cloud

[timid swift]

@wheat fern @rancid radish

[west plinth]

👀

[rancid radish]

thanks a lot, @timid swift ! hello everyone! 🙂 that's our space to talk about Cloud AI and GenerativeAI using Google Cloud!! 🥳

[wheat fern]

Hello everyone! I'm a Developer Advocate for AI/ML -- happy to discuss all things Gen AI here on Google Cloud!

[tough spade]

Cloud AI , so this is channel for vertex ai?

[rancid radish]

hey @tough spade - yes for Vertex AI too (as also BQML, GenAI, MLOps, AutoML etc)

[wheat fern]

For anyone looking for resources to learn about generative AI on Google Cloud, here's a few links:

• Generative AI YouTube playlist: https://www.youtube.com/playlist?list=PLIivdWyY5sqLRCzKJyixrIDPQKwU6XHpn
• Generative AI Learning Path: https://www.cloudskillsboost.google/journeys/118
• "Generative AI Explorer - Vertex AI" Skills Boost Quest: https://www.cloudskillsboost.google/quests/299
• Github repo with notebooks & samples: https://github.com/GoogleCloudPlatform/generative-ai/

[agile talon]

thanks @wheat fern !!! I attended your workshops in Amsterdam! they were awesome!!

great job!

[rotund cobalt]

Thanks team.. Great Job!!

[thorn gale]

can bard really send code to github

[grand vigil]

Interestingly, I completed the Learning Path few weeks ago.

[terse pagoda]

Interesting!

[chrome willow]

afternoon is new to ai is watching the vertex introduction serei

[pulsar wyvern]

Hi!
Is it me or is OpenAI/GPT more "obedient" than GenAI/Bison ?

I am trying to do some RetrievalQA (using Langchain).
I feed the model with a single webpage entry containing information about BigQuery, Spark and Dataflow (but no Kubernetes at all).
And then I give a prompt instruction to ask the model to just stitch to that webpage information:

"""Given this text extracts:

{context}

Answer the following question, using only the above information
(If you don't know the answer, just say that you don't know, don't try to make up an answer):
{question}
"""

When developing against OpenAI, and asking the question *"what are the advantages of Kubernetes?" *
I can see that OpenAI conforms with the instructions and just answers:
'The text extracts do not provide specific information about the advantages of Kubernetes.'

When migrating the same code to the whole Google/GenAI environment (so, switching the Embeddings, VectoreStore, LLM...), then GenAI still tries to be smart and gives me this answer (from its own knowledge, instead of stitching to the provided information):
'The advantages of Kubernetes are that it is a container orchestration system, it is open source, and it is scalable.'

Does anybody know how I can ensure that GenAI just gets his answers from the website I have fed him with ?

[round meteor]

Guys, I'm getting "User location is not supported for the API use" error from API, why? I'm in Turkey, it seems to be a supported region.

[glossy saddle]

Are you using paLM api? Then yes... It's only supported for US region, you can try using VPN if it's just for tweaking .

[open thicket]

Can somebody help me understand how the levels of access are assigned to database account on GCP MySQL. Is there any command that can be run on the database level to identify the permissions assigned to those accounts on the database level?

[tranquil thicket]

Just curious to figure out if any tools in AI can read out a text by creating a video with human-like expressions.

[round meteor]

Yes, PALM api. But at the time of writing the docs claimed nearly all countries as available regions.

[wraith night]

I wrote a new (pretty technical) article about custom model training and deployment on Google Cloud using Vertex AI in Go.
This article shows a different approach to solving the same problem presented in my previous article AutoML pipeline for tabular data on VertexAI in Go.
This time, instead of relying on AutoML we will define the model and the training job ourselves. This is a more advanced usage that allows the experienced machine learning practitioner to have full control on the pipeline from the model definition to the hardware to use for training and deploying.

https://pgaleone.eu/golang/vertexai/2023/08/27/vertex-ai-custom-training-go-golang/

Any feedback is appreciated.
Cheers!

[placid steeple]

Hello there.
We have built a small classification model using Google colab, and trying to deploy a prediction endpoint for a demo.
We tried to add a Cloud Function, but it seems to run out of RAM.
We'd like to stay within free tiers for this demo.
What would be the most simple way to get a prediction API usable from a web page ?

Considering the Cloud functions + Cloud run (docker) flow, but it feels like there should be an easier way.

[wintry mango]

interesting question 😄 @placid steeple model productionised is super fun. Typically the first step is to get it out of the notebook environment 😄

You'll have a couple of steps in Google colab right -> all of the cleaning, training, etc etc.

The part you need to actually implement the model is usually just the built model that is ready to do a prediction. So we might typically use something like pickling to create the model file that is now just ready to take inputs and predict straight away. Once we've got that, we can easily create a rest API

https://medium.com/mdblog/creating-a-serverless-rest-api-with-gcp-32cc62188a03

[placid steeple]

Thanks. We have the output of the saved model with saved_model.pb and keras_metadata.pb files.
We tried to load them from cloud function the same way we loaded and tested them on colab.
At the moment, I upgraded the cloud function to 4Gb RAM to try and get something working before optimizing.
I noticed that cloud function fails with errors could not load dynamic library 'libcudart.so.11.0' . I understand the google colab cpu instance has cuda installed, while cloud functions don't.

Since it's a tensorflow issue, I suppose it doesn't matter that the model is pickled or not, but I should deal with this cuda error... I guess tensorflow is trying to install cuda, which takes forever. I will try to force CPU usage from cloud functions with os.environ['CUDA_VISIBLE_DEVICES'] = '-1'

[wintry mango]

interesting. and you're just predicting ?

[placid steeple]

Didn't work, I keep getting these errors https://i.imgur.com/JlHgeWq.png
That is so strange...

Yes, I'm only getting a file data from requests.files, converting it to bytes, then an np.array, and calling loaded_model.predict(preprocessed_img), simple.

[wintry mango]

oh

[placid steeple]

I thought I could ignore these warnings, but then, why would the cloud function timeout at loaded_model.predict() . No error whatsoever returned, very difficult to debug on cloud functions.

[wintry mango]

does it load the model okay/

[placid steeple]

I think so, a print returns loaded model: <keras.engine.functional.Functional object at 0x3e9e832b22b0>

[wintry mango]

interesting

[placid steeple]

Is there a simple way to load that model on vertex AI ? It feels like it's expecting me to load the raw data and retrain a model, I couldn't figure out how to load the already-trained model.

I'll try following this article https://towardsdatascience.com/how-to-deploy-a-tensorflow-model-to-vertex-ai-87d9ae1df56

[placid steeple]

I went and created a vertex AI online prediction model using my pretrained model.
But querying it seems a lot more complex than a cloud function.
It's expecting a json object with format

  "instances": [
    { "instance_key_1": "value", ... }, ...
  ],
  "parameters": { "parameter_key_1": "value", ... }, ...
}```

No idea how to use that and how to upload my file there in the curl request (maybe it's expecting the object to describe the path to a GCS file)
The format doesn't match the document on image classification predictions either https://cloud.google.com/vertex-ai/docs/image-data/classification/get-predictions#api_1

Anyone familiar with this ?

[placid steeple]

I finally understood that I need to query vertex AI with data {"instances": [ {"image_bytes": {"b64": "base64_encoded_image" }}]} (from https://cloud.google.com/vertex-ai/docs/predictions/get-online-predictions)

I had issues Invalid argument: JSON object: does not have named input: image_bytes as reported by https://github.com/kserve/kserve/issues/942

I renamed my tensors (input was input_1 => image_bytes). Then got this error:

"error": "Failed to process element: 0 key: image_bytes of 'instances' list. Error: INVALID_ARGUMENT: JSON Value: {\n ".../2Q=="\n} Type: Object is not of expected type: float"

I modified again the input tensor to support tf.string instead of tf.float32.

Now, when querying, I get another error

$ curl -X POST
-H "Authorization: Bearer $(gcloud auth print-access-token)"
-H "Content-Type: application/json"
-d "@payload.json"
https://europe-west1-aiplatform.googleapis.com/v1/projects/${PROJECT_ID}/locations/europe-west1/endpoints/${ENDPOINT_ID}:predict

{
"error": "2 root error(s) found.\n (0) INVALID_ARGUMENT: Invalid character found in base64.\n\t [[{{function_node map_while_body_236362}}{{node map/while/DecodeBase64}}]]\n\t [[StatefulPartitionedCall/StatefulPartitionedCall/map/while/body/_1247/map/while/DecodeJpeg/_1295]]\n (1) INVALID_ARGUMENT: Invalid character found in base64.\n\t [[{{function_node map_while_body_236362}}{{node map/while/DecodeBase64}}]]\n0 successful operations.\n0 derived errors ignored."
}%

I have verified that my image content is clean base64 text. Seems like vertex AI decodes it automatically, and I shouldn't have added it in the overrided serving function
EDIT: Finally working... I can go to sleep

[grizzled dock]

Guys, I have a question regarding the JSON request with an array of instances, is this code created based on what my project has or is it something more standard. like... I don't need to use specific things from my project. im doing this project in vertex ai

[lucid coyote]

Hello everyone. I want to ask about the Document AI Workbench - Summarizer Processor. I want to create a web application using this service API but I don't see how to use the API. I get seen via Console only (https://cloud.google.com/document-ai/docs/workbench/build-summarizer-processor). My question, how to use the API for this service?

[hybrid spindle]

This is a question about google.cloud.vision.ImageAnnotatorClient()

I have difficulty finding a way to extract text from pdf files, and force the OCR to only assume horizontal text blocks, I have looked through the code, I don't see a way to define this parameter:

The following code does not work:

# Create the text detection parameters.
text_detection_params = vision.TextDetectionParams()

# Set the document orientation to horizontal.
text_detection_params.normalized_rotated_bounding_box = vision.NormalizedVertexAnnotations(
    vertices=[vision.NormalizedVertex(x=0, y=0), vision.NormalizedVertex(x=1, y=0), vision.NormalizedVertex(x=1, y=1), vision.NormalizedVertex(x=0, y=1)]
)

async_request = vision.AsyncAnnotateFileRequest(
    features=[feature], input_config=input_config, output_config=output_config, text_detection_params=text_detection_params
)

[hybrid spindle]

you can start the initial setup by asking bard.google.com

[woven narwhal]

Good Morning!

[agile talon]

morning! and welcome!

[woven narwhal]

@agile talon Thanks!!!👍🏻

[wraith breach]

Tirando los dados is a podcast of AI and data run by GDG Cloud Madrid!
In this first episode of these series, they talked to an Engineer specialized in AI who works for a Sillicon Valley Company.

Watch it again here 👇
https://www.youtube.com/watch?v=qCByQ8bNxPA

[delicate copper]

I have a error with batch predictor in my model( vertex-ai autoML). 'Error: cannot be parsed these lines' when i uploaded my jsonl correctly and combinances of it. any advice ??

[delicate copper]

i was there in GDG cloud Madrid!!

[wraith breach]

Awesomeness, welcome!!

[true orbit]

Interesting, "If you are challenged on copyright grounds, we will assume responsibility for the potential legal risks involved." https://cloud.google.com/blog/products/ai-machine-learning/protecting-customers-with-generative-ai-indemnification

[orchid flame]

hi im trying to fine tune the text bison but when i press Start Tuning, nothing happens and in console it says 500 (Internal Server Error)

[gritty cradle]

https://youtu.be/yGCggGLgROI?si=3DTTOdAy_RkXOdrq

[meager heart]

Are there really no Google img2img services, like Dalle/SD's reimagine features? I'm looking around but can't seem to find any offerings in this space, even in Imagen

not SD, i meant Midjourney

[true orbit]

Don't think so, not yet. The image editing and image generation are still in private preview (or GA, I forget which)

[livid mason]

Hi, I want to share my articles how I created a Discord Bot using NodeJS & PaLM API. Check it out! https://razanfawwaz.medium.com/create-discord-chatbot-with-palm-api-fdfefd0d6cc1

[rocky abyss]

how to create the whatapp Bot using NodeJS & PaLM API?

[carmine ingot]

hello is there any way I can host AI models on some cloud and basically do the following:
input -> AI model -> output (using api)
?

[true orbit]

That's essentially what Google's Vertex AI serving model is all about.

Intro to Vertex AI: https://cloud.google.com/vertex-ai/docs/start/introduction-unified-platform

https://cloud.google.com/vertex-ai/docs/predictions/overview

[prime shore]

Is there a way to get early access to Gemini LLM? 🙂

[west plinth]

As a developer API? Or just using it?

[prime shore]

As dev, access to API, we are in startup program

[true orbit]

The developer API will be out on the 13th

[faint olive]

Has anyone worked on a pipeline to feed 2-3 images and see it they are similar?

[half arch]

hello

[placid talon]

I'm in a middle of big project and try to migrate Google Speech to text from v1 to v2. For some reason it does not work. Anyone played with the new speech to text (v2)?
Appreciate any kind of help
https://stackoverflow.com/questions/77740107/google-cloud-speech-to-text-v2-gives-error-decoding-config-required-oneof-field

[waxen tusk]

how to gamil ip hacked helping name [e-mail] much tell me

[west plinth]

gwamo.jwanny instagram server http

[waxen tusk]

oh ty enjoy

[earnest zephyr]

hello everyone, I am building a custom chat bot using gemini pro but I found that gemini is not available for fine tuning I tried with ai studio and prompting that the chat bot working to the gemini so it can work on that way but I am not able to guess that how I gonna integrate it with my application.. if anyone of u have idea related it kindly guide ..

[nova osprey]

Thanks for drawing attention to this!

[narrow jay]

Hi there!
Hope to be in the right place for the question.
Looking for some references to seamlessly integrate google cloud functions on Shopify.
I know a few tricks but i was wondering if there was any best practice to follow.

So far webhooks and cloud functions are my primary choice.
examples:
https://www.youtube.com/watch?v=SARgBE07tis

[cobalt crest]

Hey hey all - We're just about 1-2 weeks away from releasing the full session catalogue for Google Cloud Next '24 in Vegas for April 9-11 (!!!) but as a little preview... no surprise, Gemini & Vertex AI sessions as well as related new product launch news will be one of many things that take center stage. If there's any interest here in a promo code I have just for Discord community members... let me know & happy to share w/ those who like this post directly.

[blissful rain]

Hey guys, I need help getting access to Imagen on Vertex AI. Any advice/suggestions?

[cobalt crest]

Are you on the trusted tester waitlist or have a different issue ie billing?

[blissful rain]

Thanks @cobalt crest My personal email is on the trusted tester waitlist but I would like to use (change to) my startup domain email.

[fervent nest]

Hey is there any current program running where I can get 1 month qwiklab access for free

[gilded vector]

I guess you are late for two weeks, now Statup GenAI school just started (currently week 2nd)

[cobalt crest]

Ok - gotcha. @primal nimbus do you know by chance how to help on this one?

[craggy pilot]

Im using Cloud Vision Service , and the TEXT_DETECTION feature drops my request out of the blue.
Any way to fix this ?
Like if i send 1 request in 5 min intervals , 2 out of 10 will fail.

[finite fiber]

Hi Guys... I ama novice and am trying to install gcloud on macbook but get some errors. Anyone good in this to help me out?

[formal cliff]

Hi there,
I want to implement this
reference API in PHP
https://cloud.google.com/vertex-ai/docs/generative-ai/embeddings/get-multimodal-embeddings
could you please help me with this I'm unable to get the direction.

[dull wren]

could you provide more context on what you're trying to do, I could be of help, I am also a PHP programmer

[formal cliff]

Hi Abayomi,
I want to perform some verification steps on my platform for example customer will upload their identity card and then upload the selfie.
after getting both things I want to compare the image on the identity card with the selfie image to make sure I'm dealing with the right person.
I need help with this i want to use the Google Cloud services.

[dull wren]

okay got it, and have you checked that it is possible to be done with vertex AI?

[formal cliff]

Yeah, I just read the document they are performing some image searches using text prompts.

[dull wren]

okay, please give me a minute

[formal cliff]

Sure, take your time. I'll be very thankful to you if you help me with this. I have already wasted my whole day trying to take even an initiative on this.

[west plinth]

Hey there, could you show us what you have so far?

Is the gcloud cli all set up with the right access tokens?

[dull wren]

okay, after checking the models that could be a fit for your project I think the visual-question-answering model could fit your project,

this is how you could go about it

1. merge the two images together like photo collage, so the 2 images will be beside each other, there are PHP libraries that you can use to achieve this one of them is https://github.com/tzsk/collage which uses intervention/image library
2. after collaging the two images together then send them to the visual-question-answering model API and ask the model to tell whether the image of the person in the ID card matches the selfie and then ask it to respond with a simple response like yes/no or 1 or 0

https://cloud.google.com/vertex-ai/docs/generative-ai/image/visual-question-answering

[formal cliff]

actually I'm unable to take initiative I'm looking for the step by step guide tutorial or reference to start working on this.

[west plinth]

That's fine 😄 do you have a Google Cloud account & a Google Cloud Project with billing enabled?

[formal cliff]

Yeah i have

[west plinth]

Is the Vertex AI API enabled? (sorry for asking lots of questions - just trying to see how far along it is, so that I can help accordingly)

[formal cliff]

Okay Great let me read this modal and ill feel free to ask the question if i face any issue thanks Aboyomi !

[formal cliff]

No Problem and thanks for your time.
Yes the vertex API is enabled

[dull wren]

Alright no probs

[formal cliff]

Hi Abayomi,
I've tried all methods, but they're not providing accurate answers.
even if i uploaded different objects and asking to compare the answers is yes.

[west plinth]

@formal cliff Sadly Google seems to have a policy against providing facial recognition / comparison services

Embeddings might not be what you're looking for either

As it can be prompt engineered with ease

[formal cliff]

okay now what i need to do ?

any solution will be appreciated.

[west plinth]

I'd suggest you ask a cloud GDE, those are certified experts that help people (for free !)

[west plinth]

@jolly fractal Hey Ian, sorry to disturb, can you take a look at this whenever possible?

[formal cliff]

Thanks Bedros !

[jolly fractal]

@formal cliff I don't think there is a PHP SDK for this as yet, so you would need to implement the REST API in your code.

[west plinth]

Would the embeddings suit their needs though ?

That's my biggest question tbh

Because you can submit a photo that says your name and it'll get prompt engineered / injected

[jolly fractal]

I would assume so. I am no expert on the AI stuff, I am more of the GCVE and GKE expert. 😁

[formal cliff]

can you please refer someone from your colleagues having expertise in the AI

[dull wren]

could you share your code and the response you are getting lemme take a look maybe I would be able to suggest some improvements

[formal cliff]

I'm testing the products on Google Cloud Workbench I'm not working on my own code

[dull wren]

oh okay, could you share some screenshots?

[formal cliff]

sure give me one second

[formal cliff]

here is the vertex multimodal examples

[dull wren]

okay, I mean your workbench and the response that the workbench is giving

[dull wren]

you can prompt the model API directly using PHP curl

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/imagetext:predict

[formal cliff]

Yeah, but I have uploaded images for two different persons. the person in the regular image is different from the one in the ID card image 😀
and the answer is
Response
The two images are the same person.

[dull wren]

okay 😃

[buoyant obsidian]

To be fair, this is probably really not the right type of AI to do image comparison

[dull wren]

but you could try Google Cloud Vision API

[peak sage]

🛏️ ros's coming

[west plinth]

I've already looked at the Cloud Vision API - they don't support this stuff

@formal cliff https://ai.google/responsibility/facial-recognition/
Google doesn't offer a general purpose facial recognition / comparison API - none of their products (afaik) offer this, I've looked extensively on this
You could, however, use something like OpenCV + a pretrained model and then deploy to Google Cloud (if you wish to centralize your services around Google Cloud)

Doesn't have to be OpenCV, can be anything, just that OpenCV has a huge community that can help you do exactly what you want to do here

[buoyant obsidian]

i'm using neilnet for stuff like that

[peak sage]

👀

[formal cliff]

okay thanks

[formal cliff]

any specific service?

[buoyant obsidian]

ah no mb neilnet is just a joke we have around here ahah

unfortunately doesnt exist (yet?)

[peak sage]

yet

someone on this server is named "neilnet" btw

[west plinth]

(real)

[formal cliff]

Guys, take a look at this document: https://cloud.google.com/vertex-ai/docs/generative-ai/embeddings/get-multimodal-embeddings#api-usage. It was suggested by someone from the cloud team. It seems like it could be useful for our case. However, I believe we're back to the same point we discussed

https://github.com/GoogleCloudPlatform/vertex-ai-samples/blob/main/notebooks/official/matching_engine/sdk_matching_engine_create_multimodal_embeddings.ipynb

https://cloud.google.com/vertex-ai/docs/vector-search/quickstart#install-sdk

[jolly fractal]

@formal cliff - I believe your initial question was how to implement this API in PHP. There is no SDK available for PHP currently, so you would have to implement the REST API as discussed. The SDK is currently only available for Python since that is the primary language used for this sort of stuff.

[fading bridge]

Hi guys, I'm trying to make a chatbot for a driving school website but am getting problems at this part of the tutorial
https://youtu.be/s9Np-iXOLLo?t=508
where did she get the vpc network id from to enter into the pipeline?
im so close to finishing the chatbot for teh driving school but am stuck here

[merry hornet]

Guys can anyone come on a video call and help ourteam to connect datastores to get a generative response in DFCX we will lose our jobs we will really be grateful

[dull wren]

can you explain more on what you're trying to achieve not from the tutorial, I could be of help and guide you to the right resources to implement it step by step

[dull wren]

I'm sure no one would like to do that, you can explain what you are trying to do here and someone will help out

[broken tapir]

Hi,
I'm building a chatbot for my website using a custom model I trained on Vertex AI(VertexAI -> language->My Prompts). The training went super well , and the model responded exactly as I wanted.

Here's the issue: When I integrated the API into my Flutter app (I grabbed the endpoint and parameters from "My Prompts"), the chatbot only gave generic responses, not the specific ones I got during training.

I think I might be missing something in the integration. Could you guys help me out with:
Spotting my mistakes: Any common pitfalls I might've fallen into with the integration?
Custom chatbot guide: Maybe a step-by-step on creating and integrating a custom chatbot for specific tasks in Flutter?
Troubleshooting tips: Any tricks to figure out why I'm getting generic responses instead of the trained model's output?

Thanks in advance!

[fading bridge]

I'm trying to build a ai chatbot for a driving school website

looking online a lot of peeps said to go with Dialogflow so I might give that a try

[quick pumice]

Good luck putting guardrails in. 🙂

all too often people have caused bots to just wander off talking about anything.

[merry hornet]

Okay let me describe you our entire use case.
we are trying to build a chatbot for a pharmaceutical company.Our requirement is RAG-like thing with Rich text response of the drugs within in the company and nothing beyond.This is what a sample conversation should look like BTW this is a user and pharmacist related chat bot.

User:Hi, i am feeling fatigue can you recommend me a medicine?

AI: Sure can you describe any another symptoms

User: Yes and also headache,sleeplessness

AI: Cool,i suspect you have Chronic Depression.I recommend you to contact a health professional.
User: Okay, can you recommend any drugs for time being?

AI: Sure here are my recommendations(Recommended use after consulting of professional advice only)
<Img1> with text(name,cost)
<Img2>with text(name,cost)
<Img3>with text(name,cost)

[fading bridge]

How would I do that?

[west plinth]

#cloud-conversational-ai might be of help

[dull wren]

okay give me few minutes lemme check some things out

[dull wren]

This is how you can tweak Gemini AI to make it behave like a RAG AI

this is the best I've been able to prompt Gemini AI

[dull wren]

the response

[dull wren]

this is the code

import google.generativeai as genai
from dotenv import load_dotenv
import os

import google.ai.generativelanguage as glm

load_dotenv()

API_KEY=os.env.environ["API-KEY"]

genai.configure(api_key=API_KEY)

model = genai.GenerativeModel("gemini-pro")

chat = model.start_chat()

response = chat.send_message("""
You are a chatbot for a pharmaceutical company.
Our requirement is RAG-like thing with Rich text response of the drugs within in the company and nothing beyond.

This is what a sample conversation should look like BTW this is a user and pharmacist related chat bot.

User:Hi, i am feeling fatigue can you recommend me a medicine?

AI: Sure can you describe any another symptoms

User: Yes and also headache,sleeplessness

AI: Cool,i suspect you have Chronic Depression.I recommend you to contact a health professional.

User: Okay, can you recommend any drugs for time being?

AI: Sure here are my recommendations(Recommended use after consulting of professional advice only)
recommended medicine image link wrapped in HTML img tag
text(name,cost)

Respond to the following prompt based on the description and examples above:
I am having an ulcer, can you recommend me a medicine to relief the pain?
""")

print(response.candidates)

[west plinth]

Examples perform better when the input format of the examples matches the expected input format

You can see it inserted User: I am having an ulcer, can you recommend me a medicine to relief the pain?\n\nAI: to the start of the response, which is not ideal and will cause lots of pain in developing this

Also, the ability to add image links should not be done through raw prompting alone, function calling can be implemented here to look up medicine from trusted providers and it can then output the image. However I would not add any suggestions for medicine at all. Please do not do that, @merry hornet, I know that it may be tempting for recommendations like that but it's actually really dangerous, especially if the support thing is in a pharma company website.

It's highly advised to not use Gemini or any LLM (except, say, medpalm) for medical content, instead, LLMs for support on a site should just be a friendly way to book appointments, ask questions on which professionals are available, etc
Also, I'm not sure if the person mentioned Gemini - I believe their original question was about DFCX with data stores, which supports Gemini and PaLM2 models, however it does not use the generativelanguage API as shown here and it has its own way of using it

@merry hornet pls refer to https://cloud.google.com/dialogflow/vertex/docs/quick/create-application and https://cloud.google.com/dialogflow/vertex/docs - whenever you run into any issues, describe the issue in #cloud-conversational-ai and our awesome community member @agile talon will help you with dialogflow whenever he's available 😄

[dull wren]

I've literally forgotten there's Vertex AI TBH

[wraith breach]

📢 Start 2024 strong by making a learning resolution to boost your productivity with generative AI!

Generative AI is playing a significant role in the future of technology, and Google Cloud has recently developed a range of training options to help you use generative AI in your daily job tasks through Duet AI, your AI-powered collaborator.

Check out our recommended top ten list of short trainings available on Duet AI here:
https://goo.gle/3u5XZdN

[swift olive]

What would be best to rotate an image by 45 degrees and fill the blank spots with the background color that matches most?
Trying to use GenAi Vision Edit with prompting but the results are not quite turning out as expected.

I'm not sure if that is the right solution to use for this case.

[midnight island]

@swift olive what do you mean by "background color that matches most"?

[midnight island]

if the images are as simple as your example, you don't need any sort of new-fangled AI solution:

for example (depending on how you want to fill) there are several options via imagemagick: https://www.imagemagick.org/Usage/misc/#virtual-pixel
so a one-liner would be something like:
convert my_cool_input.png -virtual-pixel Edge +distort SRT 45 my_cool_output.png

[gusty sundial]

Hi everyone,

I have recently used AutoML to train a tabular dataset for a simple use case of churn prediction. Once the model trained I have successfully deployed it to an endpoint to serve predictions. I have been able to do predictions via following ways,

1- via console under the model deployment section (for quick testing).
2- via cloud shell using a curl command with the authentication from a short lived access token.

However, I want this endpoint to be publicly accessible outside GCP domain, like via any public access using some kind of authentication (such as use of bearer token or something).

Can anyone suggest is this possible? considering both the accessbility and the relevant authentication to avoid misuse of the endpoint.

[midnight island]

having the user OAuth with a google account could work, right?

[swift olive]

@midnight island thank you so much for the examples! The image I shared is just a very simplified example. In practise I'm rotating images of portraits and due to the rotation of a rectangular image, the corners are cut off and I was hoping to fill those back. In many cases a simple color match actually works well, however in same cases a smart fill would be desired.

[midnight island]

I'd defintiely suggest outpainting the image first, and then rotating

[swift olive]

That makes quite a lot sense! Is there a paricular approach on instructing the imagen edit to expand and fill? I'm not sure if that's a good use-case for combining edits and masks or if that wouldn't be necessary

[midnight island]

seam carving is going to be the cheapest computationally (aka liquid rescale)

but that will require mostly "background" space near the edges

Otherwise, you can deploy an inpainting model like https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/stable-diffusion-inpainting

I haven't used Imagen on Vertex yet, but I assume that would be a great solution

In practise I'm rotating images of portraits

If you mean portraits of real people, please be mindful of responsible use: https://cloud.google.com/vertex-ai/docs/generative-ai/image/responsible-ai-imagen

[swift olive]

@midnight island thank you for the tips!

[swift carbon]

Is there a production API for image generation on GCP yet?

[rugged wyvern]

Hello everyone, I have a question regarding the Next'24 ticket price. What is the price for a student if he/she wants to attend the event which is held in las vegas in the month of april?

[buoyant obsidian]

I think you can register in advance here and check out the price for yourself in your own currency here:

https://cloud.withgoogle.com/next

[west plinth]

They also sent a mail to lots of people

with a code

use that code before feb 24 to get 700 USD off a full price ticket

[buoyant obsidian]

:o

that's actually decent

dammit

don't think i received that email

😭

[west plinth]

I'm not attending, i can give you mine if you want?

[buoyant obsidian]

hmmm i'll check the prices first xD

[west plinth]

Ahh nvm its not per user its a global discount

NEXT1299SESH

the email said 700 👀

[buoyant obsidian]

No but the total pricr

Because if it's 1400$

Then nope 😂

Can't afford it 😂

[zenith dagger]

it's $999 for "early bird" so that discount would bring it down to $299. It also includes a year of the Cloud training and a certification voucher. If you earn a certificate then you get $500 in GCP credits. -- Still out of my budget at the moment but I can see the value.

[rugged wyvern]

I see but the prices are expect students right. its the normal visitor price

[buoyant obsidian]

That's only for the event itself though no?

Like no hotel / flights and so on

[midnight island]

booo $1,999 for me

I'm gonna go annoy friends into giving me an alumi code

[brittle jackal]

I currently have a:
Verification Status
Pending developer action
Complete all actions below. Update your email thread with our Trust and Safety team after you have completed an action. The Trust and Safety team will continue the verification process once all actions are completed. Learn more
In which I've responded around a week ago

Is this normal & is there a way for me to accelerate this process without paying to contact support

[spice violet]

I seem to be able to see that in the email. Mine has a discount code. Have you tried checking your email yet?

[exotic stump]

Is using aistudio.google.com model tuning free?
It seems that if I want to call it programmatically, I need to use the OAuth API in conjunction with the Google Cloud SDK (which appears to require a card to be linked)?

[heady snow]

Someone please fix this terrible thing, can’t look at it..
https://ai.google.dev/tutorials/rest_quickstart

[acoustic ravine]

Hi y'all I have some questions around the high level Vertex AI Search & Conversation service. The Search App feature is pretty promising: abstracting the whole vector db / chunking / conversation state piece.

1. Data Index[ed]: Pricing for Vertex AI Search and Conversation Index Data Storage $5/GiB/mo
   I assume this is not the cost of the source files 🙂 That'd be… 100x or so greater the cost than cloud storage AFAICT. https://cloud.google.com/generative-ai-app-builder/pricing

2. How do you calculate the size of the index created for, in my case, a bucket of x PDFs (let's assume 3.5MB on average). When I've worked with vector dbs before they give you fine-grained control over chunking size, overlap, etc etc which can guide index size.

3. Does a Search App that imported data from a bucket still need the bucket content? I know the nocode widget links to them, but I'm looking at API route where I'll link to an external resource in my customised presentation instead.
   Thanks in advance!

[acoustic ravine]

4. The API as it stands is not streamed so search results can take 10+s. The lower-level manifestation of "chat with your data" might show vector responses instantly and stream the generative part to provide the best UX for people. Not sure if anyone knows if/when this is planned as an option?

[acoustic ravine]

I've found including single apostrophes even escaped are not parsed properly by Gemini

[urban elbow]

hey everyone unfortunatly my google cloud platform project was get banned due to cryptocurrency mining and i have not did that things but why this happend ?

can anyone help me

[acoustic ravine]

Wow this place is a ghost town… 🦗

[midnight island]

ha I've only just started using that API so I'm not much help, but if I come across the answer to any of that all, I'll share it :)

[buoyant obsidian]

Same

[pine charm]

And I've barely played with it because of the minimim pricing per month.

[pine charm]

I assume this is not the cost of the source files

The Pricing Example listed on that page makes it clear that it is.

[midnight island]

The $1,000 credit should go a long way, but not enough to yolo it without a plan

Which is how I have normally learned with GCP products

[acoustic ravine]

It makes no sense though -- it says 1MB per document but why are we being charged for the document being stored in the bucket AND again for the document because it happened to be indexed? I don't mind paying storage for the document indexes but… ??? Just confused here

[pine charm]

Nod. I understand. Vertex AI Search and Conversation pricing has always been ridiculous, and I've raised the issue with my contacts. See what they say.

(There used to be a $1000 / month minimum spend. But I couldn't find that in the pricing or elsewhere when I looked. So I wonder if that changed at some point.)

If I had to defend it (snort) I might say that Google doesn't do straight embedding indexing, so the index for a document may be a lot larger than the document itself. But that is... difficult... to swallow.

[scarlet palm]

Guys I someone wants to start out learning cloud tech, can someone give them a roadway?

[austere fern]

kindly check out your preferred learning path here: https://www.cloudskillsboost.google/paths

[chilly rapids]

guy's i have a problem with cloud register every time i try to create account i got this massage , i tried every thing , and nothing happend , please any help

[urban elbow]

can anyone help me in vertex ai

i am stuck.

[pine charm]

Depends on your question.

[urban elbow]

i am trying to make a model using tensorflow and deploying it to vertex ai from jupitorbook but the problem is i dont know why paramters have been changes for prediction

[turbid vapor]

anyone else having issues with duet a.i. on vscode? the extension keeps on trying to restart to update and it always fails.

[cobalt crest]

sorry im just seeing this - did you get the issue figured out? you know this im sure but that's one of the payment processing errors, so it's most often a prepaid card not being accepted, or some other incompatible FPO issue when this happens. Let me know if you need any more help if it didn't get resolved.

[cobalt crest]

just now seeing this... did you ever get an answer on student ticket pricing options?

[rugged wyvern]

some one asked me to check on the website but i couln't find any

[fluid hill]

Can someone help me in integrating chirp api into bubble.io

[chilly rapids]

hello again , im really stuck in this situation and can't create the account today i tried again and give me same erorr , and no it not prepaid card it debit card

[chilly rapids]

please if u can help tell me what deatils u need and i will share

[midnight island]

Is that Visa card you provided a prepaid/reloadable card?

Also: are you using an existing Google Workspace account/email address?

[chilly rapids]

Yes is visa card a reload able, and also yes I use already workspace account, and email address, and same card was accepted normally,

Every thing on my Google good except this cloud bill account, I have my own work space, domain, project

Google colab pro

[midnight island]

Just use a non-prepaid card

99% of the time that will fix it :)

[chilly rapids]

I do not use prepaid card 🤕🤕😔😔😔

And customer service for billing account don't even answer any thing

I'm stuck in this for almost 2 month now

[midnight island]

Hm, I might be misunderstanding. I thought "Reloadable" is the same thing as "pre-paid"

[blissful iron]

how come I only get 5 reqs/min if the the google cloud page says I should have 300 reqs/min for gemini-pro - https://cloud.google.com/vertex-ai/generative-ai/docs/quotas

[eternal gate]

Billing account?
I have similar issue same same meanwhile my Mastercard is really fine

The one I am using before got expired ever since been trying to create new one but it won’t let me

[plush bluff]

Hi. New to this. I'm trying to assemble something that can interpret free text, short sentences in an input field for booking things, say number of persons, start/end date and location. I guess "entity extraction", in short?

I just tested the Natural language API, it seems really easy to get up and running, but the entities returned were pretty...generic. F.ex numbers were returned for both number of persons and day of the month. I get the noob-ish feeling that those 2 numbers might collide if unfortunate?

So next stop Vertex AI. It seems possible to fine tune it to a much higher degree, uploading training data? Am I on the right track here, believing that Vertex ai might be the best match for my needs here? (Speaking of capabilities, the possibility to define tokens/entities to a much better degree)

Sorry if I'm vague off the charts here, my only excuse is that I'm new trying to learn about this.

[midnight island]

Sorry if I'm vague off the charts here, my only excuse is that I'm new trying to learn about this.
totally reasonable -- handling vague-off-the-charts usage is the whole point of these AI models :)

For your project, does the createEventFromDescription() function get close to your need? https://developers.google.com/apps-script/reference/calendar/calendar#createEventFromDescription(String)

or are you building a more generalized solution (ie not just making gCal events)?

[plush bluff]

Good to hear, thank you ^^! createEventFromDescription() does indeed seem interesting. Not a perfect match, what I'm working on isn't about a calendar but more like a tool for travel booking. Seems pretty similar in principle though, Enter a free text in an input box, analyse and extract entities from it to find out what the "command" is. ..which might be an answer to your question about gCal events? That is, I wouldn't say general really, it's kind of focused on an area just like the calendar stuff, but instead of calendars it's about travel booking.

[midnight island]

Good to hear, thank you ^^!

[stable swift]

Hi guys,
Lately I have being studying on AI and LLM topics. Any advanced certifications that are related to Cloud AI ?
I would really like to be a sort of AI Solution Architect 😄

[plush bluff]

I'm a bit confused here, tbh. I'm in the EU, so I use Vertex AI. I've run gcloud auth application-default login to create credentials for the API calls(right?). The JS code runs, but I get this:

[VertexAI.ClientError]: got status: 403 Forbidden.
{
  "error":{
    "code":403,
    "message":"Permission denied on resource project myproject.",
    "status":"PERMISSION_DENIED",
    "details":[
      {
        "@type":"type.googleapis.com/google.rpc.Help",
        "links":[{"description":"Google developers console",
        "url":"https://console.developers.google.com"
      }
    ]
  },
  {
    "@type":"type.googleapis.com/google.rpc.ErrorInfo",
    "reason":"CONSUMER_INVALID",
    "domain":"googleapis.com",
    "metadata": {
      "service":"aiplatform.googleapis.com",
      "consumer":"projects/myproject"
    }
  }
]}}```
Which point did I miss?

[pine charm]

That looks correct from a quick glance, tho the error "exception appending chat history" seems like an odd one. What language/library are you using, and what does your code look like?
(And what model are you going against? Quesitons about Gemini with Vertex are valid in #gemini-api)

[plush bluff]

Thanks! It's indeed example code for gemini and vertex, I'll move to the right channel right away.

[golden surge]

Azure services ❤

https://portal.azure.com/#home?wt.mc_id=studentamb_365556

[little pulsar]

Interesting, opus already shows up in my quota, but I can't use it

[{
  "error": {
    "code": 400,
    "message": "Project `removed` is not allowed to use Publisher Model `projects/removed-removed/locations/us-central1/publishers/anthropic/models/claude-3-opus@20240229`",
    "status": "FAILED_PRECONDITION"
  }
}
]

[orchid flame]

hi

[stark river]

Hi, I am trying to run a colab https://colab.research.google.com/github/GoogleCloudPlatform/vertex-ai-samples/blob/main/notebooks/community/model_garden/model_garden_pytorch_text_to_video.ipynb But I got

ResourceExhausted: 429 The following quotas are exceeded: CustomModelServingL4GPUsPerProjectPerRegion 8: The following quotas are exceeded: CustomModelServingL4GPUsPerProjectPerRegion

i change the region from us to europe but got

InvalidArgument: 400 Machine type "g2-standard-8" is not supported.

[pine charm]

That makes it look like you've spun up multiple machines without shutting them down, so you have a bunch of L4 GPUs allocated.

Take a look at https://console.cloud.google.com/iam-admin/quotas and filter by that quota name (CustomModelServiceL4GPUsPerProjectPerRegion) to see what the value is and what your usage is in the region you're working.

And then possibly https://console.cloud.google.com/vertex-ai/online-prediction/endpoints to see what endpoints you have defined and if you can shut some of them down.

[lapis fiber]

Hey guys 👋

Been trying out this cool demos from google's vertex docs to show the use cases of vertex multimodal embedding

I want to use pinecone for vector database and google vertex embeddings (multimodal embedding) for a project to build a similar search engine.

I don't get how they take into account text "timblands" (even with the typo) and "green and black" (the image part) all into one query. Anyone have an idea?

This is the demo that google provided for vertex: https://ai-demos.dev/

[pine charm]

That's part of the magic of multimodal embeddings. {:

I discuss this in a blog post (https://medium.com/@afirstenberg/using-ml-to-understand-images-3b9a06d87672 or https://code.iaflw.com/2023/09/using-ml-to-understand-images.html)

How well do you understand embeddings?

Ok, just saw your post in #gemini. This is the better place for it tho.

I think the magic in this case is that the embeddings model pulled in "shoes" and the green shoes. The "tims" has more to do with the Generative part picking the data from the search results to display.

[lapis fiber]

yup that was the part I was confused about its the timblands part 😃

[lapis fiber]

I think I have a good idea of embeddings and using vector search built some applicatiosn doing it. And I agree it is MAGIC 🪄

[lapis fiber]

ahh ok so they use GenAI to pick timblands here (even though it is a typo)

I thought they used some sort of hybrid search (dense being the multimodal embeddings) and (sparse being the keyword timblands)

Or I though they might have used unified text embedding and image embedding to take the timblands into account for search

Would love to know what they did and how it can be accomplished with pincone and google 😃

[lapis fiber]

I'll check the blogs out the papers. Looks super interestin

[pine charm]

It certainly can be done with pinecone - I think they just use Vector Matching Engine here to store the embedding, so you could swap that in.

[pine charm]

They do have a hybrid search engine. The tool used to be called Vertex AI Search and I think is now part of Agent Builder

(And the neat thing is thatr you can just store documents in Vertex AI Search, point to it as a tool in your Gemini query, and it will use it as a grounding source of truth automatically.)

https://cloud.google.com/products/agent-builder

[lapis fiber]

yh I checked that out this is a really cool outta the box RAG solution

I think it's more for search

I think my question still pretains about how they were able to get "timblands shoes that are green and black" in the screenshot above

I undersatdn for "shoes that are green and black" they use vertex multimodal embeddings but how did they get the timblands (also when it is not typed correctly)

[lapis fiber]

If anyone here knows about hyrbid search or how this was done and how I can do it please let me know 😆 been on this issue for a day tryna do what they did

[pine charm]

Did you look at the source?

Based on other stuff I've seen out of Next, I'm guessing it was multi phase:

• One that did a vector search. This pulled out the shoes and the green and black shoes.
• Another that put all of the previous results into Gemini with a prompt that included your original request. This is better at fuzzy and semantic matches, but needed some place to start.
• Based on the results from that second one - it displayed several.

[dull wren]

I hope you've been able to solve it 👍

[lapis fiber]

hmm so they take the results from vector search? embedding? and then pass the 60+ results with images to gemini to then rerank them?

Resource intensive right? There is no way and plus does gemini let you pass in 100s of images per request

I think it ight be fuzzy match but how do you combine that with multimodal embedding with pinecone?

Thats what im lost abt

[pine charm]

You're not necessarily sending the images to Gemini - just the results.

Or rather, just the metadata about the results, which includes description. And that is what might match.

[pine charm]

They're just using Vertex Matching Engine (which I thought had a different name) they say - so it isn't the hybrid match. It is just vector matching

And at that point - Vertex Matching Engine (aka Vertex Vector Search) is the same as Pinecone.

Of course, it is possible that the pictures of Timberlands could come up high enough in the vector nearby algorithm. But that does seem odd with the typo. But who knows.

Oh! This is last year's demo!

[lapis fiber]

hmm interesting

yup it seems

but I donno if they use gemini yh its last yr because 60 results each time would. I know the for image search uses pinecone and vector similar which is simple, but for that part ya it seems weird how they also then match based on the text as well?

Like I tried hybrid search with Bm25 and Vertex multimodal with sparse and dense embeddings (but the thing is Bm25 doesn't account for typo att all) so they def are doing somethin diff

[lapis fiber]

damn just realized that lol

maybe im over complicating it but its been the whole day yesteday and couldn't figure it out

[pine charm]

I think you're over complicating it. {: I'm increasingly thinking it was just the results of the embedding nearest neighbor search.

[lapis fiber]

oh so timblands wansn't even accounted for

I also though it might have been a coincidence and the red and black sheos got us the result

but then I tried just pure brand search and even with some typo it accounts for the text

or just timblands with typo

[pine charm]

I also wouldn't be surprised if the way the embeddings work tht it handles the typo ok

[lapis fiber]

yh embeddinsg is def what they use

• So they embed the image

• They also embed the title of the product it seems

• Somehow they account for both these embeddings (vertex multimodal embeds both image and text)

• So when they pass a query it looks at image embedding of the product + text embedding of the (name of product)

• Thats where im lost how do they store both those embeddings at once and account for both of them when they pass in a text query

[pine charm]

Probably as two embeddings with the same metadata.

The neat thing about the multimodal embeddings is that they share the same embedding space. So... they're just embeddings at that point.

[lapis fiber]

yh thats the cool thing you can use the same embedding for text and image

[lapis fiber]

ohh ya maybe acc

thats acc smart lowkey

but how do they search for both the embedding for the image (product of image) and embedding for the product name at the same time with one query. The way they did it was neattt

[pine charm]

Because the embedding model generates them in the same space. So they have two entries for the product - one that happens to be indexed with the image embedding and the other indexed with the text embedding of the title.

So nearest neighbor searches return them both together. It doesn't care what caused it to be indexed that way. {:

[lapis fiber]

so for the timblands green and black shoes - how did that process work from A-Z and maybe a quick overview of how that can be implemented

[pine charm]

I'm already four articles behind in what I want to write. {:

What are you doing when you get that in your console?

Using a notebook?
Writing an app that is running where?

Running on your local machine?

Did you setup the Application Default Credentials on your local machine?

You can create the credentials in the console, but in order to run the app on your local machine, you need to authenticate to the project so it knows who you are when it tries to access your project resources.

See the link that it included in the message. Specifically https://cloud.google.com/docs/authentication/provide-credentials-adc#local-dev

Linking it to a project isn't enough when you're not running on a Cloud VM.

If you're running it locally, it has to know that you are running it if you are trying to access the Metadata Server. (I don't know why you're trying to access the Metadata Server, however. Perhaps that's your question.)

To let it know that you are the one running it, you need Application credentials. I don't know what API Key you have, but STT doens't use one as far as I know.

[rough briar]

Sweet, Anthropic Claude 3 Opus is in preview on Google Cloud Vertex AI

[rough briar]

Oh, only four?

[pine charm]

Actually, you're right. It might have been 5 by last Friday night.

[lapis summit]

question. How is google ai studio being charged? i wan't to make sure i don't use an obscene about of tokens by accident. i don't see any billing in the IDE. https://aistudio.google.com/

[pine charm]

We frequently discuss this in #gemini-api , but the short answer is that AI Studio UI is and will be free, with restrictions.

[unborn tusk]

Hello,

I'm trying to configure Vertex AI using REST to use the GEMINI 1.5 model to return both text and audio in response. I've been referring to the documentation provided here: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/gemini

I'm having trouble figuring out how to specify the output formats for both text and audio in the request payload.

Can someone please help me with an example of how to structure my request payload to achieve this?

Thank you!

[pine charm]

To be clear - you want it to send audio output?

[unborn tusk]

yes @pine charm

[pine charm]

The Gemini API won't generate audio output automatically. You will need to send it through a text-to-speech system such as https://cloud.google.com/text-to-speech

[unborn tusk]

Thanks @pine charm

[valid sierra]

I'm trying to do the Vertex AI Prompt Design course from Skillsboost but can't get the first colab to run because there is never any backend available. This has been going on for two days. Am I doing something wrong? I have limited time to do these every day and it keeps getting wasted because the notebook won't run the steps (no backend)

[fading heron]

I think the login page can enjoy some ui update

[pine charm]

That's an internal web page. I'm always surprised they remembered to put the latest logo on it.

[fading heron]

Then the link is broken

[pine charm]

oh, absolutely the link is broken.

[fading heron]

They put up internal link to blog post

Lol

[pine charm]

Whats the URL for the blog post?

Hardly the first time they've done that.

[fading heron]

Llama3 on google cloud

Post

How do you ping google reps in this server?

[pine charm]

You can't, really. This is community based.
If you know the names of Googlers in a channel, pinging them is reasonable.

I've pinged some Googlers I know who are either on the team or should be adjacent to the team for them to take a look.

Here's the model card for it in Model Garden. Which I assume is what that link was supposed to go to.

https://console.cloud.google.com/vertex-ai/publishers/meta/model-garden/llama3

[fading heron]

Thanks!

[pine charm]

Ouch. Suggested deployment hardware is g2-standard-96 with 8 Nvidia L4 GPUs.

[quiet root]

I'm diving into Google Cloud and aiming to build AI solutions. To start, I'm looking to develop projects or MVPs. Do you have any suggestions? I'm also exploring the Google Cloud Skills Boost.

[pine charm]

Should be fixed now.

[polar ridge]

OK so I'm trying to get (or fix) the Gemini formerly Duet AI Code Assist VSCode Extension to work, after the udpate it's not been working. I an hihgligh texts and have it explain it, but that is about it. No full code base awareness or any of the things presented during the Google Next session. There is no singular place or source for information, and there is several conflicting statements. For exampel here: https://cloud.google.com/gemini/docs/discover/set-up-gemini#purchase-subscription it suggests optional payment , but then at the top it says it's free until July, 2024 ... SO the question is how can I get the Gemini + Google Cloud Code (Cloud Code for Visual Studio Code) VS Code Extension working and have Full Code Base context aware, code transformation, etc. https://cloud.google.com/products/gemini/code-assist?hl=en#:~:text=Full codebase awareness (Preview),comprehensive code reviews%2C and more. "Gemini Code Assist (formerly Duet AI for Developers)

---

TL:DR - why isn't Gemini Code Assist able to read the codebase or even a file, I have to highlight everything. What's the point of?

[rough briar]

Check out the cookbook as a starter

[rough briar]

You might have a look at the Insiders channel of the VSCode extension. Full code base is not yet active on the regular version. Duet AI and now Gemini Code Assist always asked to sign in into Google Cloud and associate a project with it.

[quiet root]

Thank you, @rough briar . Could you please clarify where I can find it?

[rough briar]

https://github.com/google-gemini/cookbook

[polar ridge]

THank you, I have checked that, and the full code base was active with my Duet Install (of the VSCode Insiders version) and in fact on one machine, if I don't update the plugin Duet Full Code oddly enough works. So I'm on the insiders extension, checked all the settings of the extension, I'm a part of the insiders program and trusted tester program. I'm aware the Full Code isn't in GA yet, but at least in theory or what i've been shared is that I do have it enabled, and it's active on the project. I can even (at a cost) purchase additional licenses in Vertex AI too.

So something isn't lining up.

[quiet root]

Thank you so much, Jochen 👍

[slate mango]

I want to train a model using vertex Automl. I have a folder of 1000 images. when i upload to cloud bucket get an error indicating only 500 images can be uploaded. Anyone who has encountered this behavior before?

[lapis summit]

i'm trying to call the vertex api with reqeust and api key. It doesn't seem to work and i keep getting a 404. any help would be appreciated.

i have valid api key and project Id (replaced below in code for security)

def get_gecko_embeddings(sentence):
   # Assuming you have some way to retrieve an API key and set the authorization
   key = userdata.get('GOOGLE_API_KEY')
   authorization = "Bearer " + key
   url = "https://us-central1-aiplatform.googleapis.com/v1/projects/MY_PROJECT_ID/locations/us-central1/models/text-embedding-gecko"

   headers = {
       'Authorization': authorization,
       'Content-Type': 'application/json'
   }

   data = {
       'instances': [
           {
               'content': sentence
           }
       ]
   }

   response = requests.post(url, headers=headers, json=data)
   print(response.status_code)
   if response.status_code == 200:
       r = response.json()
       print(r)
       return r['predictions'][0]['embedding']
   else:
       raise Exception(f"Failed to fetch embeddings: {response.text}")

[pine charm]

As mentioned in #gemini-api - you can't use an API Key on Vertex. You need to use a Cloud authentication method such as using a Service Account.

[frozen wadi]

Good afternoon everyone, I am following Generative AI Knowledge Base app tour and I am stuck on

gsutil cat \
    gs://courseware-coach-jason-lind-main-b4149f/dataset.jsonl \
    | wc -l

my dataset.jsonl file does not exist and I cannot figure out why... My documents are showing up in Firebase

[rough briar]

Do you have access to the cloud storage bucket?

[frozen wadi]

@rough briar I got it working, not sure what I did other than run the colab

But now I run into an issue where the model returns

Fog Computing Architecture

[Image of fog computing architecture with devices, fog nodes, and cloud servers]

Fog Computing Applications

[Image of fog computing applications in smart cities, healthcare, and industrial automation]

Fog Computing Benefits

[Image of fog computing benefits, including reduced latency, increased bandwidth, and enhanced security]

but no obvious way to link those images its extracting...

[rough briar]

Hard to say, I don't know that Colab. Sorry

[lavish agate]

Could anyone provide guidance on the ideal GPU and machine specifications for deploying Mixtral 8x22b on Google Vertex AI? I've experimented with NVidia A100 and L4, but encountered memory constraints during deployment.

[cunning patrol]

Has anyone gotten vertexAI's multimodal API working with Google cloud storage? It tells me, "caller does not have permission" but I have given the service account admin access to cloud storage and made sure that I'm using that service account.
It's this example that I'm trying to get working: https://cloud.google.com/vertex-ai/generative-ai/docs/prompt-gallery/samples/video_summarize_video_88

[burnt aurora]

I managed to get $150 free trial of google cloud for vertex ai? thanks?

I just got this, and didn't even need to put my billing details (this is actually first time using gcp), it automatically enables vertex ai api after activating the trial

since Im the billing admin and have $150 credit now, I can also enable other APIs and not just vertex with this offer?

kinda played around with gemini in vertex ai and it was cool, and some more ai services like tts, multimodal, didn't even spend a single dollar while using it. I guess that doesn't bill anything?

I think I just got this offer when I clicked "Build with Vertex AI on Google Cloud" on google ai studio

My question with this offer that is about vertex ai free trial in gcp... can I still enable other APIs like compute and stuff?

kinda feels so wrong that I was planning to try vertex ai and didn't even put my payment details, getting free $150 trial in 90days for vertex ai 😅

I'm guessing this differs from $300 offer but idk since I haven't found $150 free trial offer in google documentation

I just want to clarify things

[pine charm]

To check on credits for a project

• Go to https://console.cloud.google.com/billing
• Click on the project
• In the left navigation, go about 2/3 of the way down and select "Credits"

That shoudl show you any credits you have available. If you scroll to the right, it will tell you the "scope" (ie - services or SKUs) it applies to.

[burnt aurora]

It says any service on this billing account

so I guess that's everything

[pine charm]

that's everything. {:

[burnt aurora]

wow so I guess I can do anything with my free credit, like vms and stuff not just vertex

guess I need to set a budget so I don't spend everything

[lime oyster]

hello, which google solution is the better to create a support to call with customers, using prompt to learn how talk, and using a database to search information that the customer request?
I see gemini, agent assist and vertex, but i don't know which is better

[pine charm]

That is almost the description of Dialogflow CX, which, along with Agent Assist, is part of Contact Center AI (CCAI). Dialogflow CX is also part of Agent Builder, which may also be useful. (And they all tend to work together and have overlapping components.) We tend to discuss these in #cloud-conversational-ai

The Vertex AI Gemini API can be a component (and also has some overlap), but is somewhat lower level.

Vertex AI, in general, is a suite of tools for AI and Machine Learning. Some of the higher level tools (such as Agent Builder) are part of it, but mostly it has much much lower level tools, like multiple ML models.

[warm island]

Hey there, Devs!
Ready to level up your skills to infinity and beyond? 🚀
🚀 UPSKILL: GEN AI COURSES & CLOUD CERTIFICATES! 🚀

Demand for AI skills is exploding in the market. To help you upskill, we announced new generative AI courses and Cloud Certificates, available on-demand on both YouTube and Google Cloud Skills Boost. Once you complete the hands-on training, you can show off your new cloud credentials to employers.
➡️ https://www.cloudskillsboost.google/?utm_source=discord&utm_medium=newsletter&utm_campaign=cls-discord

➡️ https://www.youtube.com/@googlecloud/courses

[fresh geyser]

Anyone willing to share introductory level advice on using Natural Language Search on Strucutred data? I'm an AppSheet GDE looking to leverage more AI on the tabular data behind my applications for contextual summarization, reasoning, and suggestions.

[rotund phoenix]

Does anyone know of a fix for the Gemini model using Vertex AI?

Whenever I use the Gemini 1.5 pro model and send a multimodal prompt while also specifying tools / function declarations I get an API 500 internal error

[rough briar]

If I recall correctly, both resources need to be within the same project.

[tight frost]

I have tried to fine tune a data set of 3KB on Vertex AI. It took it 3 hours 12 minutes just to fail to create an endpoint which was the last step. (most probably i lost my patience and tried to fine tune another data set on an another server account) Anybody had this problem or are there any other solutions to speed up this process?

***UPDATE: We have managed to work this out finally. As suspected you can not fine tune multiple data simultaneously. We have used Compute Engine API this time also took similar times to fine tune. But here is the spicy part. Google charges us $254 for this process. I am trying to understand how and why?😡

More on this topic can be found in this link: https://www.googlecloudcommunity.com/gc/forums/v5/forumtopicpage.inlinemessagereplyeditor.form.form.form.form
Feel free to join.

[tight frost]

We have been struggling with that since 2 weeks. Its probably that the Gemini 1.5 is not yet stable enough to work the tools.

[rotund phoenix]

Yea, I mean it's still in preview so I guess that's probably it.
For the meantime I switched to Google AI and use the Gemini API there and it works fine

[tight frost]

We have decided to use GPT 3.5 instead to work the tools. We had inconsistencies even with the Gemini API 1.0

[solar iron]

Are you going for the Vertex AI builder hackathon

DM me if you have knowledge about the field....

[pine charm]

Did you have a question about Vertex AI Agent Builder?

[solar iron]

Yeah I do have a question regarding the web scraping in the Agent data store...

Cuz it is not allowing in any domain however in some of the tutorials it shows that we can use a public domain to scrape the infromation from it like store.google.com for example

[wraith steppe]

Good evening. I'm trying to create a chatbot with knowledge base + api integration. We've tried to use Dialogflow and it seems to work pretty well. However, when we use Agent Builder with tools it sometimes switches back to english language - even using instructions to use another language. Has anyone faced this issue before? Looks like Google is still developing resources for Agent type in Agent Builder

There is no option for default language selection like Dialogflow has

[pine charm]

I don't know, but I can try to forward it to some DevRel folks to ask. Exactly what have you setup, what are you doing, and what are you seeing?

[pine charm]

Actually, let me also ping @agile talon who may have played more with the i18n aspects of it than I have.

[agile talon]

Agent builder is quite new

It is mainly english for now

[amber pulsar]

Hi, I am using the Deploy to Google Cloud option from HF to deploy a model. Then using the Vertex AI SDK for inference aiplatform.gapic.PredictionServiceClient(client_options=client_options).predict(instances=instances, parameters={'temperature': 0.0}
But this doesn't seem to have any effect and the params are not being reflected from what I see in the GCP logs. Also, the "Endpoint" container has it's own environment variables and some of them look like the models hyperparameters such as MAX_NEW_TOKENS. I don't think it is possible to modify those either unless I use a customer container, maybe?
Can anyone help please?

[glad spade]

^ I have the same question. you can change the environment vars and whatnot via the SDK but not via the UI.

I also wanted to ask about plans to update the Docker image to use the latest version of Hugging Face's TGI when serving a model-- it's using 1.4.2 now but they're on 2.x and I'd love to use an updated version

[sturdy hedge]

for the last 2 weeks i am trying to add cards details in google cloud and all my cards are just keep getting rejected please help anyone

[cedar forge]

If you're interested in learning how to use Data and AI in cloud apps, I'm helping run a series of sessions using some of our newest AI codelabs, at no cost to you (thanks to some coupon codes).
Register here, next session is June 3rd:
https://rsvp.withgoogle.com/events/cv-na

[glad spade]

Can you use us-docker.pkg.dev/deeplearning-platform-release/gcr.io/huggingface-text-generation-inference-cu121.2-0.ubuntu2204.py310 (a Deep Learning Container) for serving in Vertex AI? If so how? It seems these images are meant for running a Jupyter lab server but I'd like to use it for Vertex

[craggy palm]

How do I upload zip files to google vertex ai?

[austere fern]

thanks for sharing.

[vale obsidian]

Try IDX, has full context code assist! https://x.com/ChrisLally/status/1795151102159008253

Chris Lally | Fide ⬘ (@ChrisLally) on X

I was one of many devs excited to try full-context code assist using Gemini's 1M+ token window...no wait if you give it a try in IDX 👀

Twitter · May 27, 2024, 5:52 PM

[rain kite]

Hi there,
I am using Vertex AI Agent and successfully created a Data Store with Firestore. That Data Store should now be used as a tool, but I cannot make it work.
I have trouble understanding how to provide examples. How should Input and output be used? The user should enter a keyword/query, and the system should look for the name in the data store.
It uses the correct tool when testing it, but the result is wrong. Any ideas?

[bitter patrol]

Hi y all i am having an issue, IDK if this is the right channel or not, I tried setting up GCP with a L4 GPU i am currently trying to run a resnet network but it is insanely slow.
It appears that my mac is faster.
Maybe i missconfigured something and i am seeking for help here ...
I hope you can help 🙂

[maiden cedar]

mac m1 has an npu in it

with 20 TOPs

[exotic stump]

Can the $300 Google Cloud free trial credit be used to offset the cost of the Gemini API?

[craggy palm]

Any way to leave a request? I really wish gemini could handle .zip files.

[pine charm]

yes

[pine charm]

Your best bets to provide feedback are:

• In AI Studio, use Send Feedback from the three dot menu in the upper right hand corner
• Post a message at discuss.ai.google.dev

[brittle sable]

ISSUE IN AI STUDIO..... Your site tries to use a plain string in a DOM modification where a Trusted Type is expected. Requiring Trusted Types for DOM modifications helps to prevent cross-site scripting attacks.
To solve this, provide a Trusted Type to all the DOM modifications listed below. You can convert a string into a Trusted Type by:
defining a policy and using its corresponding createHTML, createScript or createScriptURL function.
defining a policy named default which will be automatically called.
1 directive
Source location Status
hook-exec.js:1 blocked

[brittle sable]

I run into a lot of odd bugs in there, but it's still pretty cool, I can't wait ti it's been developed a bit more

[silent vine]

Hello, kindly don't spam messages on channels 🙂 Read #rules. Thank you!

[brittle sable]

If your chat name verbiage wasn't so confusing with overlapping channel names, it'd make more sense, I'm trying to HELP YOU OUT bringing the issue forwardd, so to call me a spammer is highly insiulting when you're the ones who named the chats so confusingly

I dont have time for reading Discord rules for every chat and channel, there's an issue with the AI Studio, there's the issue, do with it what you may, peace

[silent vine]

We understand your frustration, you can just leave the message in 1 channel to avoid spamming. We are only following the #rules and I hope you understand. I will get back to you and we'll report the issue to the team.

[fleet tinsel]

Hellow everyone, I'm having issue building docke image with Google Cloud #gemini-api message

[true ocean]

Hey everyone, has anyone managed to create an AI Agent with the Agent Builder? I'm trying to create a Data Store and I'm getting this weird error about "Missing required permissions: storage.objects.get", when I'm trying to select a file to import

[pine charm]

Is the bucket in the same project as Agent Builder? Do you have the permissions on the bucket set so they can be read?

[true ocean]

Hi Allen, first thank you for the response! I'm new to Google Cloud, but AFAICT the bucket is in the same project, I only have one project.
In terms of access configuration I've set it to "Prevent public access", and "Fine-grained", with the message:
"Specify access to individual objects by using object-level permissions (ACLs) in addition to your bucket-level permissions (IAM)"
Now I've set read access to the whole project for this file.

[pine charm]

it was probably requiring object-level permissions that caused the problem, then.
Once you've changed that access, does it work?

[west plinth]

i'm very late to the party but this might just be the worst way to disclose an issue :P

[scenic locust]

Hello all, I'm trying to create a runtime for Colab Enterprise but getting this error:

[old siren]

In the below code, i was expecting to get a high cosine similarity score. But the score is small. Would you know why
import os
import numpy as np
import vertexai
from dotenv import load_dotenv
from vertexai.vision_models import Image,MultiModalEmbeddingModel
load_dotenv()
api_key = os.getenv("GOOGLE_API_KEY")
os.environ["GOOGLE_API_KEY"] = api_key

TODO(developer): Update values for project_id, image_path & contextual_text

vertexai.init(project=project_id, location="us-central1")

def get_similarity_score(emb1, emb2):
return np.dot(emb1, emb2) / (np.linalg.norm(emb1) * np.linalg.norm(emb2))

image_path = "/Users/joyeed/gcpexample/gcpllm/data/images.jpeg"
contextual_text = "Image of a kitchen"
model = MultiModalEmbeddingModel.from_pretrained("multimodalembedding")
image = Image.load_from_file(image_path)

embeddings = model.get_embeddings(
image=image,
contextual_text=contextual_text,
dimension=1408,
)
print(f"Image Embedding: {embeddings.image_embedding}")
print(f"Text Embedding: {embeddings.text_embedding}")

emb1 = embeddings.image_embedding

embeddings = model.get_embeddings(
contextual_text="Image of a KITCHEN",
dimension=1408,
)

emb2 = embeddings.text_embedding

score = get_similarity_score(emb1,emb2)

print("Score is ", score)

[old siren]

I got my answer and create the below recording
https://youtu.be/YXp6obGHy4E

[pine charm]

My experiments indicated that the actual cosine similarity will be relatively low - but will be comparitively higher than anything else.

See my testing at
https://medium.com/@afirstenberg/using-ml-to-understand-images-3b9a06d87672 or
https://code.iaflw.com/2023/09/using-ml-to-understand-images.html

[old siren]

I also noticed the same and spoke about it in the recording

[mint pivot]

Hello everyone just checkout this post regarding Google Cloud Arcade Program 2024
Join fast as the registration begins on 22nd July

https://www.linkedin.com/posts/abhiig_googlecloud-cloudcomputing-learningjourney-activity-7217383367615381504-fzmV?utm_source=share&utm_medium=member_desktop

[minor iron]

nice share

[silk phoenix]

Hi, any guide on fine tuning a model on "Sphinx Docs" site on Vertex AI

[silk phoenix]

Any help is appreciated guys

[tall flint]

Hey there. 🙂 This post is off-topic here. Please check out the channel description before you post something. I'll remove your post. Thank you

[silk phoenix]

@tall flint bro, do you have any guide

[tall flint]

Sadly nope bro 🙂 I'm just a moderator here. Hopefully someone knowledgeable will help with this.

[pliant basin]

is anyone know can we do like
Uploading image to Cloud Storage bucket using Cloud Function??

[pine charm]

Is this related to Cloud AI?
Or is it a general Cloud question (which may be better in #1079896785796997170)?
(And knowing what you've tried and what language you're using would go a long way to helping answer, in either channel.)

[split bronze]

question on document ai,
when I create a custom extractor processor via api with following code, it creates the processor, but doesn't configure the dataset for it. So it requires manual update from gcp console.

const request = {
  parent: `projects/${defaults.modelData.projectId}/locations/${defaults.modelData.location}`,
  processor: {
    displayName: slug,
    type: 'CUSTOM_EXTRACTION_PROCESSOR',
    // dataset: 'google-managed', is there a way to specify this somewhere?
  },
};
const [processor] = await client.createProcessor(request);

is there a way to implement the whole process of creating a processor, training and evaluating a processor, etc via api or document ai missing all these apis yet?

[quaint turtle]

"Cloud Guide" is what GCP needed for newbies

[wise iris]

ChatGPT Bootcamp
Start Date: Mon Jun 17 2024
https://letsupgrade.in/programs/chat-gpt-essentials?invite=eshwarsajja20714

Node.js Bootcamp
Start Date: Mon Aug 05 2024
https://letsupgrade.in/programs/nodejs-essentials?invite=eshwarsajja20714

SQL 5 Days Bootcamp
Start Date: Tue Aug 06 2024
https://letsupgrade.in/programs/sql-essential?invite=eshwarsajja20714

Building YouTube Clone using HTML, CSS and JavaScript Bootcamp
Start Date: Sat Aug 10 2024
https://letsupgrade.in/programs/youtube-clone-essentials?invite=eshwarsajja20714

C++ Bootcamp
Start Date: Mon Aug 12 2024
https://letsupgrade.in/programs/cpp-essentials?invite=eshwarsajja20714

Cyber Security and Ethical Hacking Bootcamp
Start Date: Tue Aug 13 2024
https://letsupgrade.in/programs/cyber-security-essentials?invite=eshwarsajja20714

DSA with C++ Bootcamp
Start Date: Mon Aug 19 2024
https://letsupgrade.in/programs/dsa-cpp-essentials?invite=eshwarsajja20714

[misty verge]

I hope this is the correct channel. May whoever is incharge see this request: please implement and enable auto dark mode in the Google Cloud console website for all browsers that support it! I don't know but this has now become a requirement for me! Thank you!

[wary agate]

I'm a total noob to this so I'm going to try to make this as straight as possible. Bulk PDFs of various documents -> Document OCR -> print out collected data in some form of UI.

i don't understand the videos from google on how to do this at all

[pine charm]

Does anyone use Google Cloud Speech to Text V2?
Has it stopped working for you recently?

[minor iron]

im going to scream

its not working

https://tenor.com/view/cali007-gif-9067272

"help you need help we're not the place to get it from here part from move your figure don't make that excuse to violate or there's an ordinate"

[pine charm]

Well, I'm glad I wasn't the only one, at least.

[minor iron]

Ive tried v1 and v2 and all the models (refund pls #nope) they arent even getting anything right -_-

worse than the free crap

[wary agate]

im going to die

nothing is working

[wary agate]

Can anyone tell me how I can display data from google cloud storage on the front end?

[pine charm]

The #1079896785796997170 discussion may have people who can more directly help, unless this is specifically about AI.

[minor iron]

@pine charm guess back to the ole handscribe

https://tenor.com/view/annoyed-fuck-frustrated-gif-13611303

criess

[versed light]

Can someone recommend some sources I wanna learn to use vertex ai studios bc I have those free credits but I don’t wanna use real money bc it doesn’t have any limits to stop the usage

[gloomy wyvern]

have u tried Google cloud skills boost?

they have hands on labs

[versed light]

Nope

I will check it out

[gloomy wyvern]

okay

[pine charm]

https://cloud.google.com/blog/products/application-development/run-your-ai-inference-applications-on-cloud-run-with-nvidia-gpus

Just announced - Google Cloud Run now has GPU support. The result is that you can create an on-demand LLM instances that are billed by GPU second of actual usage. You don't need to spin up a machine where you pay for idle time.

So you can, for example, train your own Gemma model, run it in ollama, and deploy it to Cloud Run... and then just call it via API.

[inland flare]

Unfortunately this isn't Google Support so no one here is able to assist with this issue. https://support.google.com/ will be able to assist with account issues. Sorry you're having trouble and best of luck!

[raw talon]

YUP

[left cosmos]

Hi everyone. Can I ask a question here about the best practices for building a RAG Q/A system on a large amount of data? I'm talking about 150 GB of literature in Armenian language. I have a few questions about using Gemini Vertex AI for this purpose. 1. How well will Google's multilingual embedding models work with the Armenian language to build a vector database?
2. How can I calculate the cost of running a RAG system created with Google Vertex AI? How much will it cost to create and host a vector database on Google's side? What will be the cost of service per month if about 100 requests per hour are expected? Perhaps there is a calculator somewhere to do the calculation?
3. How well does LLM Google Gemini work with the Armenian language?

[pine charm]

Well, for starters, I'm confused by https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/text-embeddings-api#supported_text_languages which says that Armenian is "supported" but not "evaluated". I have no idea what that means.

The pricing is pretty straightforward - you can do it with a spreadsheet. See https://cloud.google.com/vertex-ai/generative-ai/pricing#embedding-models:~:text=per 1%2C000 characters-,Embeddings for Text,-Input which is a price / 1000 characters. Assuming I did my math correctly, 150G would be about $3750 assuming no overlap in your chunks and assuming 8-bit characters. This would be your price just to get embeddings for the body of work, not including queries.

The cost of 100 requests per hour is harder to calculate, since it depends on the size of each query. For nice round numbers, let's assume each query is exactly 100 characters. So that would be $0.00025 per hour. Assuming 24 hours / day and a 30 day month that's $0.18 per month to get the embeddings for queries.

That doesn't count storage or operation costs of your vector DB.

[still lava]

Hi, so im not sure if this is the right place to ask this quesiton but, When google vision is accessing a URL of an image, is it possible to set a referer to access the image and or know what is google's referer when its trying to access the URL of that image? since i have set a bunch of rules that protect images from being directly accessed.

[still lava]

Interesting.. not a single person ever encountered such issue.

[pine charm]

Oh, I'm sure plenty of people have. They've just given up waiting for Google to do something about it.

(See https://issuetracker.google.com/issues/36758197?pli=1 for a similar, but not identical, request that is over 11 years now.)

And I'm willing to venture that not many people here have used (or even heard of) most of the Vision APIs.

[buoyant obsidian]

Maybe I didn't understand well what you're trying to do, but If the images you want to use Cloud Vision on directly come from you, wouldn't it be easier & more secure just providing them through a cloud storage bucket directly?

[still lava]

Im not storing it in cloud buckets. Its being stored in a CDN of its own.

[still lava]

this is rather unfortunate. canot use User-Agent, cant set custom referer, splendid.

[cosmic tiger]

Can you use a custom token auth parameter in the fetch URL?

[still lava]

i implemented a reverse approach kinda, basically anything outside the app would require a authenticated URL to gain access, it works

[steep elm]

Hii I'm trying to implement a recommendation engine using the vertex ai agent builder. I can't get the recommendations done using the code given in the documentation -> https://cloud.google.com/generative-ai-app-builder/docs/preview-recommendations?hl=en&_gl=1*10bo4m9*_ga*MTc3NDI2MjYzNy4xNzE5OTkxMTky*_ga_WH2QY8WWF5*MTcyNjQ3NjY4MC4xNTQuMS4xNzI2NDc3NTAyLjUzLjAuMA..#structured-unstructured

I'm initialising the client as it mentioned :

project_id = "YOUR_PROJECT_ID"
location = "YOUR_LOCATION"            # Values: "global", "us", "eu"
data_store_id = "YOUR_DATA_STORE_ID"
agent_app_id = "YOUR_AGENT_APP_ID"

client_options = (
ClientOptions(api_endpoint=f"{location}-discoveryengine.googleapis.com")
    if location != "global"
    else None
)

# Create a client
client = discoveryengine.DocumentServiceClient(client_options=client_options)

What is the working/ functional code to get recommendations?

[loud ember]

Hello all, I had created a function app in azure now I have to create a function where I have to write the c# code to encrypt EDI message do anyone have idea how to write code for that?

[silent vine]

Hi @loud ember Kindly wait for an answer, avoid spamming across channels. Thank you!

[loud ember]

Sorry for that

[silent vine]

No worries. Have a nice day!

[tall flint]

Hey there. Please make sure the question is on-topic in this channel. It would be better to ask the same in #1229901767928451082

[exotic stump]

I noticed that the Llama 3.1 API service is currently in the preview stage on Vertex AI. Is it free to use? How can I use it?
I saw the following text:
'The Llama 3.1 API service is at no cost during public preview.'

https://console.cloud.google.com/vertex-ai/publishers/meta/model-garden/llama3-405b-instruct-maas

[pine charm]

That sounds free to use, yes. {:
Here's the pricing for partner models,a nd Llama isn't listed.
https://cloud.google.com/vertex-ai/generative-ai/pricing#partner-models

[vernal phoenix]

Hey everyone! 👋 Just wanted to share my latest notebook and blog post on using #VertexAI Prompt Optimizer with custom evaluation metrics. 👨‍💻
This content shows how to enhance your prompts for generating better response for an AI cooking assistant 🥑👨‍🍳 Check it out and let me know what you think!

🔗 Notebook : https://github.com/GoogleCloudPlatform/generative-ai/blob/main/gemini/prompts/prompt_optimizer/vertex_ai_prompt_optimizer_sdk_custom_metric.ipynb
🔗 Blog : https://medium.com/google-cloud/vqa-3-can-i-optimize-my-prompt-using-a-custom-metric-e3ca1e6da123

#PromptEngineering #AI #MachineLearning #GoogleCloud #LLM

[willow spear]

Hi guys, i am trying to fine tune a llm on google colab. I have colab pro sub and working with TPU v2-8. However, when I call TrainingArguments(), everytime session crashes. I tried with other hardwares, it does not crash. It seems like a special case for TPU v2-8. Has anyone ever encountered this ? or do you have any recommandation or solution ? Thank you very much.

[fluid summit]

I'm having trouble with the Client Library of Documents AI. I followed the steps from "https://cloud.google.com/document-ai/docs/libraries#use" but keep getting an error message: 3 INVALID_ARGUMENT: Request contains an invalid argument. I'm using Supabase Storage with AWS S3.

import { DocumentProcessorServiceClient } from '@google-cloud/documentai';
import { logger } from '@trigger.dev/sdk/v3';
import { env } from '@/env';
import { createAdminClient } from '@utils/supabase/server';

const filePath = translation.file_url;

logger.info('Attempting to download the file from Supabase Storage.', { filePath });

const supabaseStorage = await createAdminClient();
const { data: fileData, error: downloadError } = await supabaseStorage.storage
    .from(env.NEXT_PUBLIC_FILE_UPLOAD_BUCKET ?? 'file_upload_bucket')
    .download(filePath);

if (downloadError || !fileData) {
    throw new Error('Failed to download the file from Supabase Storage.');
}

logger.info('File successfully downloaded from Supabase Storage.', { filePath });

const fileBuffer = Buffer.from(await fileData.arrayBuffer());

if (filePath.endsWith('.pdf')) {
    logger.info('Processing PDF using Document AI.', { filePath });

    const client = new DocumentProcessorServiceClient({ apiEndpoint: 'eu-documentai.googleapis.com' });

    const name = `projects/${env.GCP_PROJECT_ID}/locations/${env.GCP_PROJECT_LOCATION}/processors/${env.GCP_DOCUMENTS_AI_DOCUMENT_OCR_PROCESSOR_ID}`;

    if (!fileBuffer || fileBuffer.length === 0) {
        throw new Error('The file buffer is empty. Cannot proceed with Document AI processing.');
    }

    const encodedFile = fileBuffer.toString('base64');

    const request = {
        name,
        rawDocument: {
            content: encodedFile,
            mimeType: 'application/pdf',
        },
    };

    logger.info('Sending request to Document AI.', { processorName: request.name });

    const [result] = await client.processDocument(request);
    const { document } = result;
}

Any assistance or clarification would be appreciated.

[little pulsar]

Wrote a short blog post on using Vertex AI’s very handy logging feature. 🙂 Didn’t know this existed until today, quite happy I found this feature!

https://www.ai.moda/en/blog/identifying-anthropic-errors-on-vertex-ai

[shadow moon]

Hey everyone 👋 I did a thing and would like to share the thing with you. It's an overview of Google AI Studio vs Vertex AI vs Gemini, what is what and when to use each: https://www.youtube.com/watch?v=MRDK8gAzDI8&t=2s

[minor iron]

You should make more videos @shadow moon

[little pulsar]

https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/batch-prediction-gemini#multimodal_models_that_support_batch_predictions

Are there any partner models, like Claude 3.5 Sonnet, that work with batch predictions?

[delicate copper]

Hello everyone,

I don't know if I'm in the right group. I need your help with something. I am a Data Scientist and I want to learn Google Cloud's data related services from start to finish. So I want to study all the learning paths from data analytic to Vertex AI. But there are many courses and they are very disorganized, I don't know in which order I should study these courses. Can anyone help me? Or which is the right group for this question? Is there a mentor group? I would be grateful if you could help me.

[pine charm]

This is probably as good a group as any. #ai-general may have another take on it. (Tho I don't have an answer to your underlying question.)

[little pulsar]

FYI Google has started charging for Llama 3.1 405B requests. Effective immediately with no notice period it seems, not fantastic. 😬

https://cloud.google.com/vertex-ai/generative-ai/docs/release-notes#October_18_2024

[little pulsar]

Instead of a course, I would really recommend just finding a problem in life, and then learning how you can use Vertex AI to solve that problem.

[buoyant obsidian]

👀

[haughty swift]

👀

[minor iron]

we turned the entire city and the state across the river into a huge nightshow. we spent $1 million this year lmao.

we call those riverboats lol and the yellow bridge, thats the big mac bridge, no one dares to even say the real name, heck, idk what it even is.

[mighty relic]

I'm trying to use the genai.file_upload functionality referenced here:
https://ai.google.dev/api/files#request-body

Whenever I try it, I get an error complaining about a missing API key that comes from https://github.com/google-gemini/generative-ai-python/blob/main/google/generativeai/client.py#L73.

Weirdly, though, I have a service account with keys that work correctly—for example, generate_content_async is fine, so the key I have works for some things.

I found no documentation specifying additional roles or permissions necessary to use that function; I looked at upload_file.md: https://github.com/google-gemini/generative-ai-python/blob/main/docs/api/google/generativeai/upload_file.md and tried following the code: https://github.com/google-gemini/generative-ai-python/blob/main/google/generativeai/files.py#L34-L74 but I feel like I am missing something obvious

Is there something special to do before one can use the upload_file API?

[pine charm]

How are you creating the genai object and calling generate_content_async and file_upload?

[mighty relic]

import google.generativeai as genai
... in my constructor for this class
            genai.configure()
            self.model = genai.GenerativeModel(
                model_name,
                generation_config=generation_types.GenerationConfig(
                    temperature=0.2,
                    response_mime_type="application/json",
                ),
                system_instruction=system_instruction,
            )
... a function in my class contains
            response = await self.model.generate_content_async(messages) # this completes successfully and the response is useful
... in a separate function
            uploaded_file = genai.upload_file(file) # this produces the error: "Invalid operation: Uploading to the File API requires an API key. Please provide a valid API key."

The credentials are provided using an environment variable, GOOGLE_APPLICATION_CREDENTIALS, which is the path to a secret file downloaded from the Google Cloud Console for my service account

[cosmic tiger]

It's been a while since I checked and can't seem to easily find the info in the docs. At one point in time, the File API only supported API Key and not oauth/service accounts. This might still be the case?

[mighty relic]

That's what my reading of the code looks like. It doesn't seem like it is documented anywhere. I assume an API key would be associated with a specific user in my org? I could probably go that route - it just seemed like the service account/key was the way to go

[cosmic tiger]

There's 3 types of auth if you look in your console; API Keys, oauth2 client IDs, and service accounts.

API Key is not associated with a specific user

[mighty relic]

Thanks - I tried using the api key and it works fine

[cosmic tiger]

Good to hear. I don't remember what the stance was from Google on other auth types for File API but at least you have it working now.

[river basin]

Any cloud engineers here in the mix?

[pine charm]

Depends what you're looking to ask. But if it doesn't have to do with AI, you might be better off asking in #1079896785796997170 instead

[river basin]

got it man

[buoyant obsidian]

[shadow moon]

For those of you working with chatbots, new video from the series "making sense of the madness": https://www.youtube.com/watch?v=BrIp7LyHrsQ

[midnight current]

Hello, can someone help with https://developers.google.com/maps/documentation/places/android-sdk/usage-and-billing ?

This looks like free unless one sees https://developers.google.com/maps/documentation/places/android-sdk/session-pricing

What i want is to create a searchbox which can take in query string from users, and now I want to send that query to google to give me 10 suggestions of autocompletions

I realized that it's an ai group, can someone move the messages to the relevant group ?

[hollow quarry]

You can go with Machine Learning Engineer path which has ai foundations, development,deployment

[eternal pollen]

Does anyone know if I can generate a Claude api key in the Google cloud platform (needs to be via Google cloud as I have credits) thanks

[cosmic tiger]

Here's some info on Anthropic's Claude models on Vertex AI
https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude

[pine charm]

You can't generate a key, but as @cosmic tiger said, you can use the Claude models through Vertex AI. (And be billed against your Google Cloud project.)

[eternal pollen]

Hmmm.... Was told my the sales folks I could...

I am using firebase to call an external API that then hosts my model that combines multiple LLMs together.

Could I generate a vertex API key then that uses the Claude model ?

[vernal ore]

ขอรหัส

[quick pumice]

We prefer english in the general channels. There are some specific language channels as well.

[pine charm]

Vertex doesn't use keys. It had other auth methods.

Where is the external API hosted? Auth from a GCP platform is fairly easy.

[eternal pollen]

The external API is hosted on Firebase, and I'm using it to call a workflow in Gumloop that manages the integration of multiple LLMs, including Claude + Gemini

I want to give gumloop the Claude api key

[pine charm]

Vertex AI doesn't use an API key. It uses GCP service accounts with credentials. If you are accessing Claude via Vertex AI - you need to use the GCP service accounts.

I asked if the call was being made from a GCP platform such as Firebase Cloud Functions because the service account is setup for you already in those cases.

[eternal pollen]

The call is being made from GCP which then calls a gumloop api, they then have their own credit system that lets you use multiple LLMs from different providers and bills you one time. If you want a discount you can provide an api key.

[pine charm]

You can't get a Claude API Key from Vertex.

[eternal pollen]

Only Gemini right

[pine charm]

You can't get a Gemini API Key from Vertex either.

You can get a Gemini API Key from AI Studio that is billed to your Google Cloud Platform account.
You can get service account credentials that work with Gemini, Anthropic, and others on Vertex that are billed to your Google Cloud Platform account.

[south coyote]

hi, to any moderator here, I have a list of text tasks to solve, I would like to use cloud service, first tier have an amount of credits, if its possible, how much time I would need to empty my credit pool?

[delicate copper]

Is there anyone who can help me on google meet bot?

[pine charm]

We can certainly try if you ask your question.

[minor iron]

That video cover is enough to win a subscribe from me

[delicate copper]

try? I need certain way

[minor iron]

well this isn't a "do it for me" channel, its a learning server, so.. you're gonna have to participate in the "learning"

[delicate copper]

how?

[pine charm]

Are you asking us to write a google meet bot for you?

[delicate copper]

Sure

I want to make a google meet bot which can join meeting and transcript it

[pine charm]

Typically this channel works best if a person has tried to do something and is having problems doing so, so they show what they've done, what isn't working, what errors they get, etc, and other people can provide suggestions.

[delicate copper]

Okay, I want to make a Meeting enhancer bot which can be used in Google meet, zoom meet and etc.
I implemented about zoom meeting already.

But according to my investigation, Google meet doesn't support such functionality

But I found other projects implemented this function already like fathom.video

[jolly elm]

Heyyyy folks I’m a founder of Intl travel advisor looking for a better coder than me to help me understand the machine learning chatbot feature for my app.

What is it? ITA Ace is a your pocket guide to making global relocation to easy and giving you confidence to fly.

My former contractor provided the code in GitHub and supposedly built it in ReAct. I can’t make heads or tails of what he did to edit or make improvements on my own - I am also a novice coder 🤷🏿‍♀️. The code is typescript primarily per GitHub and I’m hosting it in firebase. I am also trying to train the chatbot with vertex ai agent. I want to integrate components I learned in a google collab but I’m stuck 😔. Planning to put more hours in this weekend.

So that was a lot… anyone curious enough to want to help? Thanks 🥰

[minor iron]

Hey y'all- she's legit and isn't here to leech- if anyone could kindly help her out; I myself would appreciate it!!

[jolly elm]

Thank you!

Also, when I get home from work I can share the link to the Google collab I’m trying to complete and modify to better fit my use case. I’m stuck on like step 7. Hopefully this would give a better idea of how to get over the hurdle. The GitHub repository is visible in firebase at least which was a huge success.

[minor iron]

You're welcome! Hopefully someone here knows more on the matter than I do lol.

[jolly elm]

hello again, https://firebase.google.com/codelabs/ai-genkit-rag#5 this is what I was working through but I'm stuck to advance past step 6 actually

[minor iron]

Since you mentioned the underlying host (firebase) lets try https://discord.com/channels/1009525727504384150/1031906473908903989

[plucky shuttle]

You’re building a chatbot in typescript with react, hosting it on firebase, and training it with vertex ai whih is super exciting, but totally understandable if it feels overwhelming! I’d recommend finding a developer who knows TypeScript and machine learning to help untangle the code and guide you through integrating everything. Posting on gitHub, stack overflow, or hiring a freelancer could be a great next step to get the help you need.

[jolly elm]

Thank you! I will look to hire if I don’t figure it out by the end of the year. I may have to hire someone.

[tall reef]

Hi there!!, new guy here. Where to ask Questions on Vertex AI API using Imagen 3 model ? Here is the qn. anyways -Its regarding 'allow only list' in the saftey settings . Previously person/face generation was working fine through API. However, around mid last week, it asks me to request this feature. < Image generation failed with the following error: The prompt could not be submitted. Generating images containing people is currently an allowlist-only feature. Contact your Google representative to request allowlisting.> how to request people generation? also how to request 'Allow all", Is there a form? Pls advice.

[tall reef]

Hi there, any suggestion or advice on the above problem would help me.

[plucky shuttle]

you need to contact your Google coud representative directly, as stated in the error message. If you don't have a representative, submit a support case through the googlecloud console in the spport section. There isn’t a public form for this, I believe

[tall reef]

that helps. thank you

[cosmic tiger]

If you don't have a Google rep, I believe they added a public form to the documentation page recently in the People and face generation section https://cloud.google.com/vertex-ai/generative-ai/docs/image/overview#person-face-gen

[tall reef]

Thank you so much. That's exactly what i was looking for . 🤝 . Very helpful.

[languid cobalt]

The api latency for gemini models like gemini-1.5-flash-002 (non vertex) has become super slow in the past 2 weeks, whats going on? any solutions, would migrating to vertex ai solve this (any speed guarantees given for vertex)? For our production workload this is unworkable. It used to be super fast, most used to complete below 500 millisecond. Right now I'm seeing half of the requests complete in over 9 seconds. Api endpoint https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash-002:generateContent

[gloomy glade]

anyone who does cloud and is free i need some help

[pine charm]

If this is a question about Cloud AI, go ahead and ask. If this is about something else in cloud, you might want to ask in #1079896785796997170

[gloomy glade]

it's about how to get started and resources i can use to learn

[pine charm]

With Google Cloud in general, or with AI in Google Cloud?

[gloomy glade]

google cloud and AI/Ml intergration both

[pine charm]

https://cloud.google.com/docs/get-started is a reasonable place to get started, but Cloud is a big place.

[gloomy glade]

thats why i am confused

too much to learn and not many videos to cover it all

[pine charm]

You don't need to learn it all. But you do need some idea of where you're trying to go with it to know how to narrow it down. And you'll learn more along the way.

[gloomy glade]

thanks for the insight

will try to build for the project

[latent hamlet]

When I tried reopen my account, the button is disable for me.

Is anyone who can help me with this?

[buoyant obsidian]

I highly recommend you to get started on http://cloudskillsboost.google/

[gloomy glade]

thanks

[buoyant obsidian]

There are many ressources including videos & tutorials, as well as hands-on labs

so it's really great to learn & put in practice

lots of paths too: https://www.cloudskillsboost.google/paths

[gloomy glade]

will put on effort as i will have to prove myself as a intern

[buoyant obsidian]

Some of them are paired with certifications you can get

[gloomy glade]

oh thanks

really appreciate it

[buoyant obsidian]

For instance the Cloud Architect Learning Path

is also a certification you can get

Which is quite nice :)

What I like about it personally is that it's by Google, so it's up to date

[gloomy glade]

saving it all

[buoyant obsidian]

Unlike youtube videos and online articles that are sometimes years old

[gloomy glade]

btw where should i get started i know nothing about GCP

[buoyant obsidian]

So at least when you get started, you're not confused by the interface that has changed and so on :)

[buoyant obsidian]

Do you know about the cloud in general?

for instance maybe AWS / Azure, or are you getting started in the cloud as a whole?

[gloomy glade]

no

[buoyant obsidian]

oki oki

then I recommend you start with the Cloud Digital Leader Learning Path [it also has a certification that you can pass in an exam center here]

[gloomy glade]

thanks

[buoyant obsidian]

It covers the basics about what the cloud is and the challenges about it, notably for companies :)

[gloomy glade]

got it

[buoyant obsidian]

It's in my opinion a less fun one because it's just pure videos (no labs or things to try yourself), but it covers the basics so... you kind of have to start by learning all the vocabulary and so on

[gloomy glade]

then what next like i want to build a project use Image Classificaion how do i build regarding that

[gloomy glade]

good i like to go in depth from the start

[buoyant obsidian]

Then you can continue with harder paths, but more complete ones, such as the Cloud Architect Learning Path that I highly recommend to do.
It's a long one but a really covers a lot about Google Cloud, and is easy to start with, plus it also comes with a certification

Personally that's what I started with :)

Cloud Digital Leader -> Professional Cloud Architect -> Others for fun

[gloomy glade]

thanks i will keep this in mind

[buoyant obsidian]

Good luck in your journey :D

Feel free to @ me, or send a message in #1079896785796997170 if you need help with anything :)

[gloomy glade]

will do

let my journey begin

[half jolt]

Hey everyone, is there any channel to discuss about this Hackathon https://next2025challenge.devpost.com/
please let me know

[silent vine]

No specific channel at the moment. You can ask here or on #gemini 🙂

[half jolt]

Thank you @silent vine

[acoustic silo]

hey uhh

litellm.BadRequestError: VertexAIException BadRequestError - vertexai import failed please run pip install -U "google-cloud-aiplatform>=1.38"

@buoyant obsidian

any ideas

i've ran it.. but it says i havent.. like bruh

[cosmic tiger]

Are you running venv and installing it into the correct one?

[acoustic silo]

yeah

it still doesn't work

[cosmic tiger]

so you're installing in the venv and there's no errors on install
with an simple python script it fails on import, you are sure you are running the script in the same venv?

[acoustic silo]

yep.

[cosmic tiger]

you're using litellm?

[acoustic silo]

nope

[cosmic tiger]

hmm... then why litellm.BadRequestError: ?

[acoustic silo]

okay so, im just trying to use claude from my terminal

and idrk how to 💀

(using vertexai)

ive enabled vertexai, enabled claude, installed the cloud terminal thing.. but idrk what to do now lol

[cosmic tiger]

So you want to use claude models via google vertex ai. Were you following a guide? If so, which one? What's your level of experience with Google Cloud Services, Vertex, python, etc.?

[acoustic silo]

Uhh.. let's just say I'm a novice

[cosmic tiger]

Here's a guide that might be helpful to review. https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude

[acoustic silo]

uhhhh

issue

[cosmic tiger]

Have you enabled that model on your account project?

[acoustic silo]

yeah

can i share my terminal with you? lmao

i literally have no clue what im doing 💀

[cosmic tiger]

hmmm... maybe we should walk back a bit and ask, what are you trying to do? GCS, vertex, etc. may be a bit too advanced to start with?

[acoustic silo]

i just wanna have a chat with claude lmfao

but not with the claude.ai

i wanna use the google cloud vertex thingy...

https://tmate.io/t/nWDPRQfjwZXYZpK6yRH3PJUsz

if you wanna take a look at my terminal, theres a tmate link there

[cosmic tiger]

No need to look. So you just want to use the model. You don't need to write code or build anything yourself?

[acoustic silo]

wdym?

i just wanna use the model

but i hate the claude.ai website and it has barely any free messages, id rather talk to it through the google cloud vertex thing and pay through that..

and with the $300 free credits...

[cosmic tiger]

ok, so for that you can

1. log into https://console.cloud.google.com/ -- Google Cloud Console which seems like you already have
2. select your project -- seems like you should already have a project setup from the sound of things
3. navigate to https://console.cloud.google.com/vertex-ai/studio -- Vertex AI Studio
4. on the left menu you can select the mode such as Chat
5. on the right you can select the model you want to use
6. assuming you have the model enabled already otherwise you will need to enable it

[acoustic silo]

it just keeps saying

ERROR Quota exceeded for aiplatform.googleapis.com/online_prediction_requests_per_base_model with base model: anthropic-claude-3-5-sonnet-v2. Please submit a quota increase request. https://cloud.google.com/vertex-ai/docs/generative-ai/quotas-genai.

[cosmic tiger]

in the UI or via code?

[acoustic silo]

UI

[cosmic tiger]

if you go to https://console.cloud.google.com/vertex-ai/publishers/anthropic/model-garden/claude-3-5-sonnet-v2 does it indicate that it's enabled?

[acoustic silo]

😐

like i have said before

it is enabled.

[cosmic tiger]

Just walking through the process since I don't know what has or hasn't been done.

Do you know how to look up quotas? I don't have any of those models enabled on my account so I can't remember if it defaults to 0 (zero) and requires a request to be submitted.

[acoustic silo]

idfk

i dont even understand the current usage part, because i've NEVER had it work lmao

[cosmic tiger]

in the filter area search for claude and it should show some models to filter on

[cosmic tiger]

So it looks like you have 0 on your quota so it supports the errors you are seeing.

also, this additional page may help with some partner models setup https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-partner-models

[acoustic silo]

no but even when i use something like opus

it still doesn't work..

[cosmic tiger]

can't really tell you what or why you have usage against your account that you say you haven't used but that's a different problem 🤷‍♂️

[silent vine]

@acoustic silo The use of profanity is not allowed here. Kindly refrain from doing it again. Thanks

this is your last and final warning for breaking server rule #1

[obtuse jungle]

hello how is everyone here??

need a good roadmap to getting into cloud this year can you all please suggest and guidee

[scenic flume]

Check out the learning paths on skillsboost https://www.cloudskillsboost.google/paths/

Also another option https://cloudresumechallenge.dev/docs/the-challenge/

[obtuse jungle]

Thankyou so very much

[vernal phoenix]

Hey everyone! 👋 Just wanted to share my latest notebook and blog post on serving multiple LoRA adapters on Vertex AI!

This content provides a practical example of how to deploy a Gemma 2 model with multiple LoRA adapters on Vertex AI using Hugging Face Deep Learning containers to enable SQL and code generation tasks.

Check out the article and notebook here:
Notebook : https://github.com/GoogleCloudPlatform/generative-ai/blob/main/open-models/serving/vertex_ai_tgi_gemma_multi_lora_adapters_deployment.ipynb
Blog : https://medium.com/google-cloud/open-models-on-vertex-ai-with-hugging-face-serving-multiple-lora-adapters-on-vertex-ai-e3ceae7b717c

Let me know what you think and...Happy coding 🤗

[obtuse jungle]

AMAZING!!

[gritty cradle]

Whatsapp + Gemini = 🔥

[vernal phoenix]

Hey everyone! 👋 I want to share the new Google Cloud tutorial on how to evaluate open models with Vertex AI!

The tutorial is part of the Building with Hugging Face on Google Cloud series 🔥 and shows how to Gen AI Evaluation service on Vertex AI to evaluate Gemma 2 for a summarization task.

Here's the video: https://www.youtube.com/watch?v=AUSunZXC2rg&list=PLIivdWyY5sqIwEOfjCSVl87ND7Rn3m1Fd&index=2

Also you can find the notebook in its description to reproduce the content!

[vernal phoenix]

Hey Cloud AI crew! 👋 I had some time to experiment with Ollama and Vertex AI!

Below you have a step-by-step tutorial walks you through deploying a Gemma 2 SQL adapter using Ollama on Vertex AI.

https://medium.com/google-cloud/hey-ollama-how-about-running-on-vertex-ai-03cded7bfd0b

As always, let me know what you think and...Happy coding 🤗

[pine charm]

I don't think there is a GCP free tier for the Generative AI stuff, but there is a $300 cloud credit available.

[pine charm]

As far as I know, yes

[buoyant obsidian]

New Vertex AI Agent Builder customers also get 1000$ of GCP credits, if that's something you're looking for :>

(A list of the services that have free tiers is available here: cloud.google.com/free)

[willow shale]

Thank you Google Team for the cloud credit.
We made this AI Solution for the Indian Judiciary and the state government also encouraging such projects.
most of the legal professionals make use of it.
'AI Advocate' Trained on all Indian 11,000+ acts
https://aiadvocate.live

[pine hearth]

I've been trying to use the vertex ai retail recommendation but I'm having a hard time filtering the recommendations. Is that valid filter v2: (availability: ANY("IN_STOCK")) AND (NOT brands: ANY("foo")) AND (NOT brands: ANY("bar"))?

because this filter on a predict call returns some products in either foo or bar, and there's no way of evaluating filters in the google console

[autumn sapphire]

[ Cloud Computing Series– Powering the Future with Cloud Innovation ]

Hello, GDGoC dwellers! 👋🏻

Get ready for the Cloud Computing event series!

𝗧𝗲𝗰𝗵 𝗧𝗮𝗹𝗸 #𝟭 – 𝗖𝗹𝗼𝘂𝗱 𝗖𝗼𝗺𝗽𝘂𝘁𝗶𝗻𝗴 𝗨𝗻𝘃𝗲𝗶𝗹𝗲𝗱: 𝗘𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻, 𝗜𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻𝘀, 𝗮𝗻𝗱 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀
📆 Friday, February 7, 2025
🕒 19:30 – 21:10 WIB
🔗 Zoom Meeting

Explore cloud computing evolution, cutting-edge technologies, and real-world case studies from companies like Netflix and Spotify.

𝗣𝗿𝗮-𝗦𝘁𝘂𝗱𝘆 𝗝𝗮𝗺 #𝟮 – 𝗚𝗲𝘁𝘁𝗶𝗻𝗴 𝗦𝘁𝗮𝗿𝘁𝗲𝗱 𝘄𝗶𝘁𝗵 𝗖𝗹𝗼𝘂𝗱 𝗖𝗼𝗺𝗽𝘂𝘁𝗶𝗻𝗴: 𝗦𝗲𝗿𝘃𝗶𝗰𝗲𝘀, 𝗧𝗼𝗼𝗹𝘀, 𝗮𝗻𝗱 𝗕𝗲𝗻𝗲𝗳𝗶𝘁𝘀
📆 Saturday, February 8, 2025
🕒 19:30 – 21:10 WIB
🔗 Zoom Meeting

Learn cloud services (IaaS, PaaS, SaaS), tools like Compute Engine, and insights on the Juara GCP program.

𝗦𝘁𝘂𝗱𝘆 𝗝𝗮𝗺 #𝟯 – 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗖𝗹𝗼𝘂𝗱-𝗕𝗮𝘀𝗲𝗱 𝗣𝗿𝗼𝗷𝗲𝗰𝘁𝘀: 𝗛𝗮𝗻𝗱𝘀-𝗼𝗻 𝘄𝗶𝘁𝗵 𝗚𝗼𝗼𝗴𝗹𝗲 𝗔𝗽𝗽 𝗘𝗻𝗴𝗶𝗻𝗲, 𝗖𝗼𝗺𝗽𝘂𝘁𝗲 𝗘𝗻𝗴𝗶𝗻𝗲, 𝗮𝗻𝗱 𝗖𝗹𝗼𝘂𝗱 𝗦𝘁𝗼𝗿𝗮𝗴𝗲
📆 Sunday, February 9, 2025
🕒 08:30 – 11:45 WIB
📍 Room B2-205, Campus B STT Terpadu Nurul Fikri

Build a cloud-based portfolio website and enhance your cloud development skills with Google App Engine and Compute Engine.

Regist on bevy now: https://s.id/GDGOC_CLOUDCOMPUTING

See you there, folks! 🙌🏻

[silent vine]

Would be great to add it on #events 🙂

[gritty cradle]

Gemini + Whatsapp = 🔥

[swift olive]

Has anyone played around with exporting models from Vertex Ai to Tensorflow.js? I'm seeing a significant decline in prediction performance comparing the trained model on an endpoint vs running in browser from the exported model.

[untold maple]

Can you help me?
I am a AI developer.
I'm going to build a site integrate Gemini with Vertex AI with the primary objective of enhancing our machine learning models, specifically classification models.
I would like to ask for help from a developer with expertise in this matter.

[silent vine]

For gemini try to ask on #gemini

[sonic bridge]

We are trying to use Google Evaluation Engine. However when we are invoking it from a langraph agent it is giving the following error

2025-02-18 22:40:56,683 - ERROR - An error occurred: 403 Permission 'aiplatform.endpoints.predict' denied on resource '//aiplatform.googleapis.com/projects/llama3-1-integration/locations/us-central1/publishers/google/models/gemini-2.0-flash-thinking-exp-01-21' (or it may not exist). [reason: "IAM_PERMISSION_DENIED"
domain: "aiplatform.googleapis.com"
metadata {
key: "resource"
value: "projects/<project-id>/locations/us-central1/publishers/google/models/gemini-2.0-flash-thinking-exp-01-21"
}
metadata {
key: "permission"
value: "aiplatform.endpoints.predict"
}
]

We have tried with all kinds of permission and does not seem to like it. Is it something that #gemini developers in Google have to enable? Any help much appreciated.

[cosmic tiger]

What library or SDK are you using and what authentication method?

[subtle panther]

Hi guys!

I have a question: does the LM Notebook API exist or are there plans to develop it? If anyone has information, documentation or any news on this subject, could you share it here?

I really appreciate your help!

[pine charm]

It does not exist.
Google has not made any announcement that I'm aware of that they're working on one.

[subtle panther]

Got it, thank you very much! I hope that in the near future they develop some way of using it via API

[little pulsar]

Is there a way to restrict a parameter in a Vertex AI API call to an allowlist? It seems like Vertex AI cannot be used with any IAM Conditions, which seems... really wrong? This seems like a basic thing to ask for.

[tacit charm]

I completed Gen AI Beginner path on google cloud skills boost , but what is the purpose of these badges and ranks ? Do they provide any benefits?

[sand wigeon]

Hey folks, my YouTube Data API v3 uploads are stuck at 1-2 views (just me!) even after verifying my app. Manual uploads used to get 400-500 views. Is it a privacy setting glitch or something else? Any quick fixes?

[tacit charm]

Hey I have 35 credits in cloud skills boost so what should I do ? For that credits?

[terse spindle]

Guys I really need help

[silent vine]

what kind of help?

[terse spindle]

I am unable to sign up for google cloud, with errors that are not my fault

I cannot contact google support

and it is halting my thesis

I am genuinely stuck and might have a freak out cause the deadline is creeping and I cannot work

support keeps pointing me to different areas, accounts keep saying im not the billing admin when I have not done anything

I am really at my wits end here

[silent vine]

https://cloud.google.com/contact

[terse spindle]

: ^)

I am the billing admin

Getting juggled by chatbots is also getting more and more frustrating

[buoyant obsidian]

Why so? Do you have any error that shows up?

[vale igloo]

I was testing vertex RAG Engine, it seems to be hallucinating. I just used Vertex AI Docs to create corpus. and it is giving weird answers. I am using Flash 2.0

[halcyon sedge]

Hello everyone! What’s the best way to implement transformer models for analyzing long texts with N-Grams and NER in an automation flow (e.g., Make.com)? Has anyone worked on this before?

[vernal phoenix]

Hey everyone! 👋

If you are struggling with APIs for Vertex AI model deployment, this is for you 😉

Vertex AI Model Garden now simplifies open model deployment using the new Deploy API. 🔥

TL;DR:
💻 Deploy models like Gemma 3 with just ~3 lines of code.
🤖 Automate deployments with the interactive CLI.
🛠️ Configure with ease: set machine types, GPUs, replicas, use Spot VMs, etc.
⚠️ Catch quota/policy errors early.

Check out Blog post with code examples
👉 https://www.googlecloudcommunity.com/gc/Community-Blogs/Introducing-the-new-Vertex-AI-Model-Garden-CLI-and-SDK/ba-p/888386

Give the new SDK a try and let me know what you think! 👇

[mint pivot]

Exciting news! Enrollment for Google Cloud Arcade 2025 has officially started, and this is your chance to dive into cloud computing, earn Google Cloud skill badges, and unlock amazing opportunities!

Why Join Google Cloud Arcade 2025?
Hands-on experience with Google Cloud Labs
Earn industry-recognized skill badges & showcase your expertise
Exclusive mentorship & learning resources
Unlock exciting rewards & swags
Boost your career with real-world cloud skills

Enrollment is open now—don’t miss out!

Join the TechOps Community to Register & Stay Updated!
To help you maximize your learning experience, I have also created a WhatsApp community where I’ll be sharing opportunities, insights, and guidance throughout the program.
Join the WhatsApp Group Here: https://chat.whatsapp.com/IanqvEyYQmWGkpKTKezkJy

Let’s make the most of Google Cloud Arcade 2025 and take a step toward becoming cloud-ready professionals! Tag your friends, spread the word, and let’s build an amazing learning community together!

[mighty harbor]

Does anyone know what the rate limit for "text-embedding-005" (Vertex AI) is? I can't find it anywhere on Vertex AI docs, as well as the quotas and limits dashboard on Google Cloud Console. Thanks!

[pine charm]

I believe this is reported under "Regional online prediction requests per base model per minute per region per base_model" for a base_model value of "textembedding-gecko". It appears to be 1500

(well, 1500 in us-central1 anyway)

[mighty harbor]

thanks!

[vale igloo]

Hi there, I was trying to implement Veo 2.0 APIs into my SDK https://github.com/gunpal5/Google_GenerativeAI

I am getting this error:
{
"error": {
"code": 400,
"message": "Async process failed with the following error: Feature text to video is not allowlisted for project 103876794532.",
"status": "FAILED_PRECONDITION"
}
}

[pine charm]

Ok.
Is it supposed to be allowlisted for your project?

[vale igloo]

I am not sure where to add my project to allowed list.

[pine charm]

Veo 2 is not in public preview yet.
If you are in the private preview group, typically they permit a project. If you want another project, you'd have to talk to your contact at Google.

[vale igloo]

I'll try that, Thank you

[vivid notch]

Hello, I am interested if gemini 2.0 flash fine tuning model has function calls and if yes could you show me how to make it? I tried for non tuning model and its worked but for tuned model it isnot

[primal wolf]

Hello guys,

I have a python script that requires a strong machine (currently I am running a 32 GB and 8 proc) and I might still run out of memory, is there a better option than using a CloudRun Job?

[buoyant obsidian]

If you want to use more memory you could switch to Compute Engine

especially if it's a one-time task

You create a VM, SSH into it from the console, execute your python script and when it's done delete the VM!

[primal wolf]

@buoyant obsidian , its supposed to work as a cronjob,

pickup a a task that it has a status ready_to_build, process it and go on to the next one

and in the future it should be able to do more tasks concurrently

[thin birch]

Vertax looks promising, can we still apply for the 1000 credits?

[polar belfry]

Hi

[trail belfry]

Does anyone know if the Agent Development Kit supports structured output?

[gusty magnet]

Hi, is this the correct discord channel to enquire about doubts regarding GCP Vertex AI Media Studio doubts?
Specifically about using Chirp 3 HD Voices Custom Voice Settings. I wanted to know whats the pros and cons of Custom Voice vs Built in Voices. What does "I consent to Google using this voice to create a synthetic voice model" involve?

[gusty magnet]

In ADK github pages doc, under 'Agents --> LLM Agents --> Sturcturing Data', details for using pydantic is mentioned. Maybe that will help?

[pine charm]

This is a good channel for that.

[pine charm]

What does "I consent to Google using this voice to create a synthetic voice model" involve?
Pretty much what it says. If you want a model based on your voice, you have to give google permission to use your voice to create the model.
What other questions do you have about it?

[gusty magnet]

Hi, thank you for your response.

Basically wanted to know if we're consenting to google useing for improving google's services or just for providing us the voice clone service and not chirp model for training.

For example, when we use Google AI Studio

• "Free Tier" means Google use our input for improving services related to RLHF.
• "Pay as you Go" means Google will not use the data for AI training. It only keeps logs for legal purposes.

So is this consent similar or different, I did not see any similar distinction in that Custom Voice Generation workflow.

Also, would like to know whats the optimum audio length for a good custom voice training when asked to submit a .wav file

[pine charm]

IANAL, but the phrase "to create a synthetic voice model", which you have to record, certainly sounds like it limits the use.

If you have questions, you can contact your account rep when you get added to the allow-list.

[pine charm]

The documentation says "up to 10 seconds".

[sacred forum]

Hello @mint pivot I am sorry but your post is against the #rules , rule #4

[kind vigil]

Is there a channel or perhaps even separate discord for Google ADK (Agent Development Kit) ?

[native salmon]

I haven't found one, but this channel seems like the best so far.

[last fox]

Heyy guyss, I'm dealing with Google's new ADK, I was running the sample repo provided my Google, but I'm getting errors in credentials, but I provided all right credentials, can any one help me out from this if you have a good knowledge on GCP and AI

I'm using free tier

@pine charm can you help me

[cosmic tiger]

• Which sample repo or tutorial are you using?
• What kind of auth are you using?
• What errors are you encountering?

[last fox]

https://github.com/google/adk-samples/tree/main/agents/llm-auditor this is the repo

for info : gcloud auth application-default login
gcloud auth application-default set-quota-project $GOOGLE_CLOUD_PROJECT i also ran this in Cloud CLI

@cosmic tiger

[cosmic tiger]

Have you worked with Application Default Credentials, google cloud cli, and setup a project before?

[last fox]

yeah i setup the project, i have the project and region

i'm able to print the access token as well

@cosmic tiger

[cosmic tiger]

Are your environment variables set correctly?

[last fox]

yeah

[last fox]

plz come in my dm, i'll share you the problem properly

[cosmic tiger]

There shouldn't anything confidential so can just communicate in the channel.

Have you verified the credential file is created and accessible after you authenticate?

[last fox]

no, whats that

how to verify the credential file

[cosmic tiger]

Maybe start here and read about ADC and get a better understanding of it. https://cloud.google.com/docs/authentication/application-default-credentials

There's also a section on troubleshooting the ADC setup

[last fox]

yeah i got that, its running now

[cosmic tiger]

Good to hear

[jolly pike]

hi online ones

[remote roost]

Hello all, I just tried to participate in a Google Cloud Skills Boost challenge and I saw this notification.

I am wondering what does it mean? Do I need to set up a separate account with my student email?

[hoary wasp]

The labs, generally you should run in incognito mode, but If you run in your web browser not problem because you use the credentials provided by lab, but is better to follow the instructions

[sacred forum]

Hello @wheat cave please provide context with your post

[remote roost]

Yeah, I eventually figured that out, thanks.

[neon mulch]

spam links take care guyss

someone who is so good in cloud deployment can tell me how to write a nodejs script that can take a zip file and deploy it to gcp or cloud shell whatever and mention me please as messages is getting too big

[craggy palm]

can I buy credits for 2.5 pro? or is that still not ready?

[hidden pivot]

oh yes

I guess billing has been delayed the past couple days... I thought I was vibe coding on the credit I had still....

boy was I wrong 😂 😂

[neon mulch]

any cloud or web

[elder siren]

Hey folks 👋 What are the IAM roles required to access App Design Center? I'm getting an error that I don't have the necessary permissions but I've got all 3 of the ADC permissions in IAM plus a bunch of other ones.

Found it -- Project IAM Admin, Service Usage Admin, App Design Center Admin, and App Hub Admin.

[gleaming latch]

hi guys, I need some help, we are using vertex ai with gemini (2.0 flash, 2.5 flash & pro). We are calling these from a cloud function hosted in same location as the model, however, when using flash is taking around 10-20 secs on average to generate the response (the total tokens processed are like 2k) I've seen plenty of comments out thhere about the latency issues, anybody faced this and knows how to solve it? this is making gemini imposible to use in prod ready services

[dreamy flower]

Greetings, Running into an issue trying to use the AgendDevKit ADK.

Have everything installed , and working great locally.
Just trying to run tests and getting :

E   ModuleNotFoundError: No module named 'google.adk'```

**Does anyone have any ideas?**

Running `adk web` works perfectly. 
I have logged into the glcoud with `gcloud auth application-default login`

[shrewd orbit]

Hi everyone, I'm completely new to Google Cloud Platform and eager to start learning. My focus is in AI and ML, so I'd really appreciate it if anyone could recommend a well-recognized course/certification tailored to those areas.

[pine charm]

https://cloud.google.com/learn/training/machinelearning-ai
Has a number of courses and certifications from Google, some of which may be what you're looking for.

[shrewd orbit]

Yeah thank you very much, in addition do I need to learn cloud services the GCP if so where can I get started with? Any Coursera certification that is offered by Google is available?

[pallid stag]

I've been getting this error over vertex for past a hour :(

I didn't use much and yet rate limited?

I'm new to GCP, on free trial and don't know whom to contact or resolve this issue through; any help?

[pine charm]

That looks like AI Studio, correct?

Gemini 2.5 Pro Preview has a very low quota at the moment.

[pallid stag]

I mean, I didn't use more than 4 times in this whole day; is it that low? It's on vertex ai so paid tier should have higher quota right?

and I see nothing have exceeded the quota either...

[pine charm]

Ok, so you are using Vertex. Sorry, wasn't clear from your initial image.

On Vertex, Gemini 2.5 uses Dynamic Shared Quota. So there is no specific quota. And there will be no quota reported.

So 429 means that there is no available resources at that moment.
See the URL that is in the error message you posted.

(imho - DSQ is the abolute stupidest "feature" in Vertex)

[pallid stag]

No problem! thanks for helping allen

I thought quota will be shown on quotas & system limit as name suggests

URL page suggests I should check the quota limit which I didn't exceed

Honestly I'm at end of wits 😕

I've been trying to develop PoC for my startup with GCP then suddenly this happened, any more suggestions on how to resolve this issue?

[pine charm]

Try using 2.5 Flash instead.

[pallid stag]

Okay

[pine charm]

It isn't you. It is that 2.5 Pro is resource intensive, highly popular, still in preview, and Google has insufficient resources allocated / available.

[pallid stag]

ah, Resource exhausted. Please try again later. Please refer to https://cloud.google.com/vertex-ai/generative-ai/docs/error-code-429 for more details.

Same error I guess?

[pine charm]

See https://cloud.google.com/vertex-ai/generative-ai/docs/dynamic-shared-quota

[pallid stag]

Understandable, as it is preview, I shouldn't expect reliability all the time

Sure!

Okay I understand that overall demand on the shared pool at the moment is very high; I wish error message had been more clear/understanding

Nonetheless, thanks a lot for the help, Allen!

I'll wait for entire day and hope the demand on this specific shared pool to reduce

[distant shale]

hey guys anyone here know what the pricing per million tokens would be if i deploy deepseek r1 in vertex ai model garden on SGLang 160K context speculative decoding 1 host (8 NVIDIA_H200_141GB; a3-ultragpu-8g)?

[static stirrup]

Go down the rabbit hole in cloud skills boost.

[eager galleon]

🗿

[gray rune]

Has anyone worked withe Google ADK with env_vars directly being pulled from the Cloud Secret Manager? I'm having a really hard time making it work 🙂 The ADK engine won't start on the Agent Engine platform on Google Cloud as soon as I introduce a secret variable.

Following this part of the documentation btw:
https://cloud.google.com/vertex-ai/generative-ai/docs/agent-engine/deploy#environment-variables

[stuck garden]

Hello all 👋 ,

For those interested in the Google Agent Development Kit (ADK)🤖, I've written a blog post detailing my experience building a basic Conversational News Assistant Proof of Concept .

The article covers the initial setup, tool integration, and overall impressions of using ADK for a simple agent. You'll also find a link to the GitHub repo within the post, making the POC available for you to test, fork, or contribute to.

Link: https://medium.com/google-cloud/from-zero-to-news-hero-building-a-conversational-ai-agent-with-google-adk-a-hands-on-poc-0737f4c0c6c3 📄

This POC was primarily a test of the framework's fundamentals. I'll be sharing more as I explore advanced ADK features in subsequent articles.

Feedback and discussions are welcome! 👇

[tribal pike]

Hey, Everyone Atharv here

I got with a issue on Cloud Skills boost

My university had partnered with a third-party platform to provide students with free credits for completing cloud-related courses. However, due to a mistake, my registered account was deleted from Google Cloud Skills Boost.

Is there a way to recover the deleted account or reassign the credits to a new account under the same university email?

[mild stream]

Data Analytics Roadmap is nice ❤️ lol we are going to need it haha

[sleek ravine]

Is there a way to increase api quota for Gemini 2.5 TTS models?

[pine charm]

These are still preview models, so probably unlikely.

[viscid fulcrum]

What is Agent 2 Agent. I understand "Agent Builder" on GCP that is now " Ai applications " but what is A2A?

[pine charm]

a2a is a new protocol (and library) that is designed for agents to communicate directly with other agents, share controled information, etc.

[livid verge]

You're giving them a voice, man. What if they demand robot rights?

Lkl

I mean I guess they can only complain to each other so they can get stuff seriously that's pretty awesome because I mean that's a big problem at the moment them not being actually able to properly communicate with each other so you can't have multiple agents working on the same problem without overlap

I wonder if like humans they'll just choose not to communicate critical information just for shits and giggles.

[lapis jacinth]

hi all,,, when running multiagent program using goodge ADK specially with adk web it is impossible to run as it is always complaining for sub process etc

has anyone encounted this issue?

now i am forced to run using CLI which is not that interactive like adk web

i think the moment we put MCPTools for the tools used by agent it just break

pls share you comments if you have already seen and fixx this issue

i tried both on windows and linux - nothing facncy simple agent application with lets say a MCP server for MySQL and MCP server for files system

[lapis jacinth]

@stuck garden any pointers

[stuck garden]

In the last couple of week I didn't had time to work on it. But just now I wanted to test sth with the POC I publish and I have some errors too. Also the service I had running in Cloud run was complaining, and it used to work. Maybe some change in the libraries? Did you try updating the libs and the SDK?

[lapis jacinth]

Yeah I think adk mostly shine in gcp ecosystem

Even the messages like Parts are gemini specific

This is has another Avidence.. If u use MCPToolBox for database then it's working fine even on local windows machine

[stuck garden]

I resolve my problem by Updating to the new version of google-adk in Python, and moving from dataclass typyng to BaseModel

[tardy ravine]

A case against Vibe Coding... from a Vibe Coder. It’s a long read, but maybe someone here will find it interesting.

[grave forum]

https://youtu.be/0S3Kute8gEQ

[kind eagle]

guys

I have some queries about google cloud arcade

[humble pilot]

Can anyone please tell me how to create a basic ai agents with vertex ai, I tried but getting invalid grant access

[buoyant obsidian]

https://www.youtube.com/watch?v=8rlNdKywldQ

[neat cloud]

I am doing the google cloud flower image classification model and the tutorial seems to be out of date. Please help with the error messages.

https://cloud.google.com/vertex-ai/docs/tutorials/image-classification-custom/serving

gcloud functions deploy classify_flower
--region=us-central1
--source=function
--runtime=python37
--memory=2048MB
--trigger-http
--allow-unauthenticated
--set-env-vars=ENDPOINT_ID=${ENDPOINT_ID}

I have to update to python 3.9 because python37 is not even depreciated it's gone.

gcloud functions deploy classify_flower
--region=us-central1
--source=function
--runtime=python39
--memory=2048MB
--trigger-http
--allow-unauthenticated
--set-env-vars=ENDPOINT_ID=${ENDPOINT_ID}

Traceback (most recent call last):
File "/layers/google.python.pip/pip/bin/functions-framework", line 5, in <module>
from functions_framework._cli import _cli
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/functions_framework/init.py", line 31, in <module>
import flask
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/flask/init.py", line 7, in <module>
from .app import Flask as Flask
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/flask/app.py", line 27, in <module>
from . import cli
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/flask/cli.py", line 17, in <module>
from .helpers import get_debug_flag
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/flask/helpers.py", line 14, in <module>
from werkzeug.urls import url_quote
ImportError: cannot import name 'url_quote' from 'werkzeug.urls' (/layers/google.python.pip/pip/lib/python3.9/site-packages/werkzeug/urls.py)

ChatGPT says
✅ Ultimate Solution

You cannot patch third-party packages like Flask/helpers.py in a clean deployment. So, if the error comes from Flask's own code, then:

❗ You must downgrade Flask to 2.2.x and Werkzeug to 2.2.x.

That doesn't work because requirements.txt already follows that advice:

Flask==2.1.0
requests==2.24.0
google-cloud-aiplatform~=0.5.1
tensorflow-cpu~=2.8.0

Where us Werkzeug in the code anyway?

I also had to change this:
gcloud projects add-iam-policy-binding flowerai-463021
--member="[e-mail]"
--role="roles/cloudbuild.viewer"

To this:
gcloud projects add-iam-policy-binding flowerai-463021
--member="user:[e-mail]"
--role="roles/cloudbuild.builds.editor"

[neat cloud]

Attached is the code:

[burnt chasm]

I also have a few versions that step through basic deployment from the cloud shell, a Linux VM and the VM on Chrome OS

[burnt chasm]

Another good ADK example is the Instavibe code lab. This article goes step by step on how to deploy it from Firebase Studio:

[copper nexus]

@burnt chasm@stuck gardenHey, #new-publications might be a better place for these, that's where all new articles go!

[exotic hill]

Rate limits are applied per project, not per API key.

Does that mean You can get around the rate limits by adding more projects for your project?
What are the limitations to creating a new project (amount, pricing) ?

[primal flame]

Anyone can help me? why i can't use Claude model on Vertex AI

[tulip steeple]

Hello dear sir I was doing my lab in gen Ai exchange program but got this error and lab was quitted can u help Organization Policy constraint constraints/vertexai.allowedModels violated for projects/614738821393 attempting to use a disallowed Gen AI model veo-3.0-generate-preview. Please contact your organization administrator to fix this violation by adding publishers/google/models/veo-3.0-generate-preview:predict to the allowed values. For more info, see https://cloud.google.com/vertex-ai/generative-ai/docs/control-model-access.

Request ID: 1984854035337227032 and this error

Been getting this error can anyone help

[delicate shadow]

Short answer: https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/claude/use-claude#regions

To maintain overall service performance and acceptable use, the maximum quotas might vary by account

in plain English, this means you haven't spent enough money for Google to consider you a user who won't abuse their PAYG quota by using just registered accounts
if you want to speedrun it, then look here
the exact amount you need to pay varies: could be $20, could be $50
better than waiting for the end of the billing period tho

[delicate shadow]

yes, you can, GCP quota works in mysterious ways

What are the limitations to creating a new project (amount, pricing)?
https://support.google.com/cloud/answer/6330231
differs from the account
default value of projects per Organisation - 10

[quick glacier]

@silver oasis @silent vine can you guys help me?

I am doing Prompt Design in Vertex AI: Challenge Lab

[silent vine]

Unfortunately we can't help you with that.

[proud hazel]

Hello community

There is no proper documentation on Imagen: EDIT_MODE_PRODUCT_IMAGE

I got this error trying with Imagen 3-capability-001:

{"error":{"code":400,"message":"Image editing failed with the following error: Product image editing is not supported on Imagen 3 Capability Preview.","status":"INVALID_ARGUMENT"}}

Do you have a use-case on how to use it?

[burnt chasm]

For anyone with access to qwiklabs there are 2 new ADK courses live with both labs and quizzes. https://www.cloudskillsboost.google/public_profiles/4cfcca49-bc82-4d90-bb00-eb244a0243ec/badges/16893058

[native salmon]

Hello,
I'm trying to use the DiscoveryEngine API to get a list of provisioned products

$cred = gcloud auth print-access-token $headers = @{ "Authorization" = "Bearer $cred" } Invoke-WebRequest
-Method GET -Headers $headers
-Uri "https://iam.googleapis.com/v1/projects/ai-agent-builder-test-project/serviceAccounts" | Select-Object -Expand Content

Invoke-WebRequest -Method GET
-Headers $headers -Uri "https://discoveryengine.googleapis.com/v1beta/{name=projects/*}" | Select-Object -Expand Content

The first call works, but the second call fails with a 404.

404. That’s an error.

The requested URL /v1beta/%7Bname=projects//operations/%7D was not found on this server. That’s all we know.

I'm trying to call this API
https://cloud.google.com/generative-ai-app-builder/docs/reference/rest/v1beta/projects.operations/get

[native salmon]

Well, silly me, it's a gRpc endpoint.

[placid peak]

I’m using Vertex AI with Flask in Cloud Run. It works once after deployment, but fails on later requests with "could not resolve project_id" errors. I'm using the default service account with Vertex AI User role. It works fine in Cloud Shell, but not consistently in Cloud Run. Any idea why this happens or how to fix it?

[pine charm]

Where are you getting that error?
Have you checked the logs to see what they say?

[placid peak]

@pine charm
Yes, I’ve checked the logs — the error appears directly in the Cloud Run logs when the app tries to execute crew.kickoff().

Here’s the traceback:

Error generating report: litellm.APIConnectionError: Could not resolve project_id
...
File "/usr/local/lib/python3.12/site-packages/litellm/llms/vertex_ai/vertex_llm_base.py", line 442, in get_access_token
raise ValueError("Could not resolve project_id")

This happens inside litellm, which is used under the hood by langchain_google_vertexai.VertexAI.
Even though I explicitly pass the project and location, it still fails after the first request in Cloud Run.

[pine charm]

scratches head Why is LangChain using LiteLLM?
(Tho I'll admit, I'm more familiar with the JavaScript side than the Python side of LangChain. And that could be a red herring.)

I would be expecting all of them to use ADC and not need you to specify the project id at all. It's clearly an auth-related issue if it is happening in get_access_token.

How long in between your requests?

[placid peak]

Yeah, I was surprised too — I wasn’t explicitly using LiteLLM myself. I’m just importing VertexAI from langchain_google_vertexai, and it looks like under the hood it’s relying on litellm to handle the Gemini model calls. So it might be an internal dependency or default behavior.

You're right — ADC should ideally take care of the credentials automatically. But even though I’m explicitly passing project="demo-ai-app" and location="us-central1", it still throws:

ValueError: Could not resolve project_id

This only happens in Cloud Run. The first request always works right after deployment, but any follow-up requests fail with that error. If I wait 5–10 minutes and try again, it works once more — then breaks again on the next one.

I think what's happening is that when Cloud Run starts a fresh container (right after deployment or idle timeout), the credentials are properly loaded or refreshed. But on subsequent requests, something in the container (maybe litellm or LangChain) tries to reuse a cached token or improperly resolves the project context, which causes it to lose access to Vertex AI.

So yeah — it feels like an access token caching or refresh issue, possibly due to how litellm or the VertexAI wrapper is managing auth state in a stateless container environment like Cloud Run.

[vernal phoenix]

Hello there 👋 ,

Ivan here from the Vertex AI Advocacy team. I just want you to know that Vertex AI Agent Engine now support custom installation scripts.

With the new custom installation scripts for Vertex AI Agent Engine, you can now use a shell script to bake system-level libraries and non-Python dependencies directly into your agent's container image at build time.

TLDR:
⚙️ Simplified Deployments: No more runtime workarounds for complex dependencies.
🤝 New MCP Pattern: Easily deploy agents with their own tools as self-contained servers using the Model Context Protocol (MCP).
🚀 Faster & More Reliable: Shifting setup to the build step means faster agent startup and fewer runtime errors.

Check out the blog, code, and docs to get started 👇

Blog: https://discuss.google.dev/t/deploying-adk-agents-with-mcp-on-vertex-ai-agent-engine-using-custom-installation-scripts/250649
Documentation: https://cloud.google.com/vertex-ai/generative-ai/docs/agent-engine/deploy#build-options
Tutorial: https://github.com/GoogleCloudPlatform/generative-ai/tree/main/agents/agent_engine

[native salmon]

Agentspace - How to share an agent I created with Agent Designer with someone else in my Organization? Is it possible? Or are they only local to me?

[west plinth]

@buoyant obsidian Mr agentspace help him out 🙏🙏

[dusky tide]

Hi, I’m coming from #ai-general message

Here’s my issue:

I want to set up a RAG engine with metadata support.

• i used this documentation as base: https://cloud.google.com/vertex-ai/generative-ai/docs/rag-engine/use-vertexai-vector-search#retrieve_relevant_contexts_using_the_rag_api
• I’ve scraped ~100 websites into Markdown.
• Each has metadata (URL, title, etc.). It’s also in the Markdown, but that can be ignored.
• I want to query it with top_k and metadata filters.
• Data should be embedded into a Vertex AI Vector Search instance.
• Retrieval should (1) return some metadata, like the original URL (not just the gs:// path), and (2) allow filtering by tags, customer IDs, etc.

I can’t figure out how to configure the RAG engine so that crawled data in a GCS bucket retains usable metadata.
I tried the JSONL format but had no luck.

If there’s an up-to-date example or docs you can point me to, I’m happy to dig in.
Alternatively, if I’ve misunderstood RAG capabilities with Vertex AI Vector Search or Pinecone, please point me toward the correct approach.

Thanks

[buoyant obsidian]

It's currently not possible to directly share Agents you made with Agent Designer to specific colleagues.

What you can do though is make them available to all your colleagues (company-wide), or subsets of them that may have been created by the Agentspace admin at your organization.

I recommend that you contact your organization's Agentspace administrator so they add your agent to the "From your company" section of the Agent Gallery, or contact your Sales Rep / Partner Engineer / Customer Engineer at Google so they help you with that :)

[native salmon]

Okay, so no official way yet, but I was able to add this by grabbing the Agent ID from my Agent Designer created Agent and then use that ID in my Agent URL field, with the normal preamble of project/locations/global/agents/agentid and it now shows up in my organization.

[native salmon]

Of course now I get an error when running it under a different user, so that didn't work.
"error": {
"code": 429,
"message": "Too many planner turns for one request. Limit: 10",
"status": "RESOURCE_EXHAUSTED",
"details": [
{
"@type": "type.googleapis.com/google.rpc.RequestInfo",

[native salmon]

Thanks. I am the Agentspace Administrator and have been able to Publish those to different Application instances.

We have a user that desires to share a really helpful agent she created via Agent Designer with her team.

So sounds like I would have to replicate that functionality using the ADK and publish to make available to others.

At least for the time being. Probably not worth the effort if it is on the roadmap.

[buoyant obsidian]

If you have different instances per team that could work but if not there isn't a way to do this yet

[dusky tide]

bump in hope for input 😨

[sullen pilot]

@dusky tide Use restricts for the metadata you need in the query filters (tags, IDs).
Use embedding_metadata for the metadata you only need to retrieve along with the snippets (URL, title).

code
JSON
{
"id": "chunk_id_123",
"embedding": [0.1, 0.2, ...],
"restricts": [
{"namespace": "tags", "allow": ["tag1", "tag2"]},
{"namespace": "customer_id", "allow": ["cust123"]}
],
"embedding_metadata": {
"source_url": "https://original-site.com/page",
"title": "Page Title"
}
}

[dusky tide]

But how do I get the data+metadata into an index/rag engine?

ATM the spurce_url shows the GS:// url

[sullen pilot]

@dusky tide Do NOT use ImportRagFiles.
Create a script that generates a data.json file with your snippets, embeddings, filter restrictions (tags), and embedding_metadata for return data (URL, title). Create a MatchingEngineIndex pointing to the GCS where you saved the data.json.
When querying, use the filter parameter to filter by tags and return_full_datapoint=True so that the response includes your embedding_metadata with the original URL.

[dusky tide]

so i do not use the RagEngine for this use calse at all? "just" an index

[sullen pilot]

To achieve what you want, you should stop using ImportRagFiles and build the index directly in Vertex AI Vector Search.

[dusky tide]

oh or wait, this here

Create a MatchingEngineIndex pointing to the GCS where you saved the data.json.
is the input into the rag, yes?

[dusky tide]

ah ok, got it

[dusky tide]

so i need to make the emmbedings "myself" instead of relying on the ragengine to do this.

[sullen pilot]

@dusky tideThis is just an example, adapt the instructions to your needs.

[pallid stag]

I was usin gemini 2.5 flash lite for past few months and never been charged so high. Today, I got billed for 10k usd!

I immediately disabled the billing account and when I tried to contact support in chat, it said I'm not on paid account? naturally my bank doesn't have that huge amount and still large amount of cash on pending

what can I do on here

most of times it would be 30$ or around charge and this time it's damn fucking high that I don't think I could pay; saw some reddit claiming google using debt collect agency to pursue

[copper nexus]

Hey, probably you already looked at this place since you tried contacting support through chat but gonna link it anyway. https://cloud.google.com/support/docs/get-billing-support
I'd suggest to try through chat again and say that what was responded isn't applicable to your case.
If not that, on the left on the panel there is cases, you can see if theres way for you to open a case for your project through "get help" button in cases page.

[pallid stag]

I've tried and read all of those 😭

[pallid stag]

got charged for something I didn't use [2.5 flash native image gen???]

[pallid stag]

@copper nexus thanks for the support, it seems officially this has been labelled as bug

[copper nexus]

Lets gooo! Happy for ya!

Hope you get refund soon

[pallid stag]

Yep!

[sacred sigil]

Hello everyone,
I've run into a blocking issue with Vertex AI and I'm hoping someone in the community might have some insight.
The Goal:
I'm trying to use Gemini models (e.g., gemini-1.5-pro) from a Node.js application running on a GCE VM. The application uses a service account for authentication.
The Problem:
All API calls from my application fail with a NOT_FOUND: Model 'gemini-1.5-pro' not found error.
Debugging Steps & The Core Issue:
To diagnose this, I have bypassed my application and am using the gcloud CLI directly on the VM. I've confirmed the following setup:
Billing: The Google Cloud project is linked to an active and valid billing account.
API Enabled: The "Vertex AI API" is enabled for the project.
Permissions: I am using a service account that has the Owner role on the project.
Authentication: Running gcloud auth activate-service-account --key-file=... succeeds without any errors.
However, when I run the definitive test command to list available models, the API returns an empty list:
code
Bash

This command succeeds without permission errors

gcloud ai models list --region=us-central1 --project=<MY_PROJECT_ID>

But the result is always:

Using endpoint [https://us-central1-aiplatform.googleapis.com/]
Listed 0 items.
I have tested this against multiple regions (including us-east1) with the same result.
My Question:
Why would the gcloud ai models list command successfully connect to the Vertex AI service but return an empty list for a project that has both billing and the API enabled?
It feels like the models are simply not being provisioned for my project. Is there another service, permission, or a common configuration step I might be missing?
Any help or pointers would be greatly appreciated. Thank you

[pine charm]

Gemini 1.5 has been deprecated for several months now. New projects haven't been able to use it, and even existing projects will lose access on September 24th.

https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/1-5-pro

[dense agate]

Any chance Google's own example code doesn't work correctly?

I'm trying to test the simplest voice conversation with Gemini Live API, using the exact code from this example [https://cloud.google.com/vertex-ai/generative-ai/docs/live-api/streamed-conversations#start_a_conversation] (Python Gen AI SDK)

And it works well for one turn - I say "Hi", I can hear the model responding, but any subsequent voice audio is either not being sent or the response audio is not being received because the script seems to "hang" or rather the model stops responding after the first response.

I tried spinning up my own scripts using the examples but keeping running into the same issue. I tried with different models, native audio and the "legacy" ones. Issue keeps persisting.

Be on this for two days straight, would love some insight! Thanks!

[young loom]

Hey I'm getting this error any solutions to this?

[narrow furnace]

Hi i've deployed the agent on Vertex AI agent engine is it possible to interatc with the deployed agent with ADK web? i want to use the web ui

[solid oriole]

ADK web is good for local dev

but for production use, I think u shud consider alternatives

[narrow furnace]

Thank you

[narrow furnace]

One more problem i want to test Google Memory bank in the local agent but i got really confused with the documentation, do i need to deploy agent in the agent engine first? in order to use memoery bank? and if so can i run said agent with web ui?

[solid oriole]

nope, no need to deploy, u can use adk web ui

[narrow furnace]

Thank you!

still i'm trying to set it up right now

[dawn sand]

有没有想合作开发一款输入法app的
有别于现在大部分输入法app
这款输入法app 会把大量常用词 没意义词  和用户输入的字词混在一块
还可以实时更新 可以减少用户输入法被监控的可能
有没有一块开发的 联系联系
Is there anyone interested in cooperating to develop an input method app?

Different from most existing input method apps, this one will mix a large number of common words, meaningless words, and the words typed by users together. It can also be updated in real time, which may reduce the possibility of the user's input method being monitored.

Is there anyone willing to develop it together? Please feel free to contact me.

[pine charm]

It is possible, but as @solid oriole said, it is not a good idea. Not least of which is that it exposes a lot of your internals, not to mention other sessions! A security nightmare for production use.

[pine charm]

What issues are you running into?

[narrow furnace]

I'ts just that english is not my main language and reading document is rather challenging, i'm trying to set up a memory for my simple agent, i want to know how to set it up since i only know how to run ADK WEB right now.

i have trouble navigating documents i only see how to set up memoery bank for the agent engine instance, but i don't quite understand how to set it up in my root agent python script

or do i need to run my agent programmatically?

[sudden stirrup]

Hi! Hope u all good! Do you know where I can find public posted Studio Google AI Builded app? Want to test some or discover new idea

[narrow furnace]

Hi excuse me, i'm using a command adk web path/to/your/agents_dir --memory_service_uri="agentengine://myagentengineid"

i've create my agent engine instance

does this mean when i tell something to the agent like "i like train" i should save the memory in the memory bank yes?

but it seem it does not work, how do i verify if i succesfully set up and use the momory bank
i've make sure i have export the Environment Variables
and i've read somewhere that memoery bank now only support us-central, but i'm using singapore location is this the cause?
I'm really confused right now thanks !

[fierce glade]

hi guys, can someone help me? i making feels cloudFunctions v2 but i receave a stranger logg, i search in all internet but dont find anything, i using Python 3.11 (Ubuntu 22) to buil my cloudRun CloudFunction v2, follows the log a said:

WARNING: All log messages before absl::InitializeLog() is called are written to STDERR

E0000 00:00:1756832732.556250 8 alts_credentials.cc:93] ALTS creds ignored. Not running on GCP and untrusted ALTS is not enabled.

[narrow furnace]

Hi is there a way to verify wether the agent in agent engine got the memory bank set up correctly?

[alpine heron]

idk if this is the right spot but is the Cloud Run AI event open to students? it asks for company info, etc. just wondering if students can still register

[solid oriole]

yep they shud be, assuming the company info and such fields are optional

[pine charm]

Definitely welcome. Tho stuff run by the GCP team tends to have more of a business spin.

[sacred sigil]

hey does someone know more about the known bug ?
i need to patch it up or it will eat all my general purpose credit instead of the trial credit for gemini code editor and cli

[pine charm]

I've never heard of this bug before.
But I would point out that LLMs are not sources of truth. If it hasn't given you a link to a resource saying this - it may be hallucinating.

That said, if you look at the "scope" column on those Credits pages, you should see exactly which products the credit applies to. Verify that the trial credits are limited to those products, or it is treated the same as other credits you may have gotten.

[cosmic tiger]

As far as I understand it, it's not a bug. I'm assuming you are referring to the Trial credit for GenAI App Builder credit. Due to all of the product name changes it is unclear what that actually refers to. The terms and condidtions of that promo also doesn't seem to be readily available. No one seems to have a definititive answer as to what falls into that promo. I've personally never reached out to Google and don't have any record of it either. What is universally clear is that AI Studio API and Vertex AI API for do not fall into that promo.

[cosmic tiger]

Unfortunately, Google Cloud console is limited in this regard. There is no scopes/skus listed in the credits screen and only states you should see the terms. However, those terms and conditions are nowhere to be found. 🤷‍♂️

[pine charm]

Ah, it's one of those.

[buoyant obsidian]

What's the bug?

[pine charm]

The image suggests that for Cloud Credits, it is charging against one bank of credits instead of another.

[sacred sigil]

sorry for late reply.

The "known bug" appears to be related to the behaviour of the Gemini CLI and Gemini Code Assist when it comes to token usage and billing.

Based on public community discussions on platforms like GitHub and Google forums, the bug is often described as follows:

When the CLI's tools (e.g., file-editing commands) encounter an error or a bug in their own logic, they can fall back to inefficient methods, like reading and rewriting entire files.

This fallback method consumes a massive amount of input tokens because the full file content is repeatedly fed into the context window.

The issue is that this high token usage is a direct result of a tool's internal bug, but it still gets deducted from the user's rate limit or trial credits, even though the process was unproductive.

In essence, you're paying or using up your tokens for the AI's own mistakes. The user in your Discord screenshot was likely highlighting this exact issue—the frustration of seeing their general purpose credit get depleted due to a bug in the Gemini CLI, rather than for successful, productive work.

This is a recognized problem, and discussions suggest that the ultimate fix would be for the service to differentiate between usage caused by bugs and normal API usage.

[sacred sigil]

not the genAI credit. i know that one is for the vertex and for me to overhaul the wiki i have on my page. at the moment i use vertex secure key and embedding for reach in my handbook and gemini flash lite for the wiki to answer

[pine charm]

If you have security concerns about a package, raising them on the library is a better approach than messaging two random people on discord. {:

That said... just tried this:

$ npm install langchain

added 38 packages in 7s

8 packages are looking for funding
  run `npm fund` for details
$ npm audit
found 0 vulnerabilities

So I can't duplicate what you're seeing.

[sacred sigil]

should we carry on in dm ? you know better 😉

i think matter on version of the dependency.

[sacred sigil]

maybe point me the right direction please

[pine charm]

Possibly. I did a fresh install.
But that just means that they fixed the dependency on a vulnerable version already.

[pine charm]

You can go to npmjs.com and get to the page for each package, which will link to the github repository where you can report it.

But, to be honest, I take just about everything from an LLM with more than one grain of salt. In this case, I'd ask it to explain what makes it think that and verify it yourself.

[sacred sigil]

gotcha, i will remove that post to not cause any panic and i will dig around a bit. thax

[cosmic tiger]

Ok, got it. Your original message and screenshots made it seem like you were questioning the credits rather than your restated bug.

[vernal phoenix]

Hey @everyone! 👋

A quick and exciting follow-up on all the great feedback some of you shared for Vertex AI Memory Bank, the managed memory service on Vertex AI. We've been busy turning some of your suggestions into features, and the new release is now live!

TL;DR of the new features:
👉Memory Lifespan (TTL): You can now set memories to automatically expire after a set time. Great for keeping your agent's knowledge from getting stale.
👉Custom Knowledge: You can define custom memory_topics the agent should care about and provide few-shot examples to teach it how to extract that specific info.
👉Model Selection: You can now choose the specific embedding and generation models you want to use, so you can optimize for cost, performance, or multilingual support.

I have published a full announcement and a get started tutorial over on the Agent channel of the official Google Cloud Community forum.

📚 Blog: https://discuss.google.dev/t/announcing-customization-features-for-vertex-ai-memory-bank/261941
📓 Tutorial: https://github.com/GoogleCloudPlatform/generative-ai/blob/main/agents/agent_engine/memory_bank/get_started_with_memory_bank.ipynb

Hope you find these updates useful! Happy to answer any quick questions here, on Linkedin or X.

[hoary fossil]

Hi, I’m exploring backend development with Python and automation. Curious how others are using cloud tools in their workflows.

[vernal phoenix]

Hey there! 👋

Vertex AI just dropped a great simplification for building multi-agent systems.

They introduced the new native integration of the Agent2Agent (A2A) protocol on Vertex AI Agent Engine.

TL;DR:
🔗Universal API: Use the A2A open standard so your agents can communicate easily with any other A2A-compliant agent.
📦Simplified Deployment: No more managing separate runtimes or writing tons of glue code! Deploy your entire A2A agent as a single service directly on Agent Engine.
⚙️Fully Managed: Go from code to a secure, scalable, serverless endpoint with a single SDK call.
🤝Interoperable: Build agents with a clean, reusable API that makes it easy for other applications to connect and interact.

To get started, check out the following resources:

📓 Notebook
✍️ Blog Post

I'd love to hear your feedback. And if you have questions, you can also connect with me here, on LinkedIn or X/Twitter.

[vernal phoenix]

Hey #cloud ! 👋

I am excited to share about the new Code Execution on Vertex AI Agent Engine in preview!

It's a managed service that gives you a sandbox for your agents to run generated code.

TL;DR:
⚙️Fully Managed: It's a simple API call, so you can forget about maintaining infrastructure.
↔️Flexible: Works with your LLM or agent framework you're already using, like the ADK.
⚡Stateful: Sandboxes are persistent for multi-turn conversations.
🛡️Isolated: Runs LLM-generated code in a hardened and isolated sandbox.

If you want to know more, here you have notebook and tutorial to get started.

As always feel free to share your feedback and if you have futher question, reach out !

[sacred forum]

Hello @runic jetty please don't spam the same message in multiple channels

[runic jetty]

ok sir sorry i just want to get knowledge. i apologize

[sacred forum]

Don't worry 🤗

[runic jetty]

ok

[vernal phoenix]

Hey everyone!

Google just released the Agent Payments Protocol (AP2), an open protocol to create a secure payment layer for the agentic transactions.

Today's payment infrastructure was built for humans clicking "buy". When an autonomous agent needs to make a purchase (especially a "human-not-present" one, like "buy these shoes if the price drops below $100"), how does the ecosystem handle trust, authorization, and accountability? How do we protect against agent "hallucinations" leading to bad purchases?

AP2 aims to address these questions by engineering trust into the system using cryptographically signed Verifiable Credentials (VCs):

• Intent Mandate: Captures the user's conditions for an agent to act in their absence.
• Cart Mandate: The user's explicit, signed approval for a specific cart in a "human-present" scenario.
• Payment Mandate: A signal to the payment network (issuer/network) that an AI agent is involved.

This is designed as an open extension for protocols like A2A and MCP and will work with any agent framework (like CrewAI, LangGraph, etc.). It even supports crypto/x402 alongside traditional cards and bank transfers.

Check out the samples and the full spec on the GitHub repo: https://github.com/google-agentic-commerce/AP2.

Also we are going to record a The Agent Factory Podcast (https://www.youtube.com/playlist?list=PLIivdWyY5sqLXR1eSkiM5bE6pFlXC-OSs) today together with the PM of AP2. If you have questions, please drop in this thread.

Happy building!

[teal falcon]

Is it possible to get a gpu vm for training ml models via gcp free trail?

[west snow]

yes it is

has anyone run into their instance nuking itself when hitting high VRAM spikes? I’m seeing this issue on the Deep Learning Deploy image with an L4.

[teal falcon]

How?

[west snow]

Yeah, GCP’s free trial gives you $300 credits, you can spin up a GPU VM with that. Just be careful since GPU hours burn through credits fast.

[teal falcon]

I tried, it doesn't

Allow me

To do it

[west snow]

whats the error youre getting

[teal falcon]

You can't utilise a gpu in

A free trial

Please activate

To use full gcp

Features

[west snow]

just activate

your credits will prevail

[lone sequoia]

Hey guys I hope you're doing great!
I'm having a project to create chatbot with LLms langchain and RLHF from a Google docs file or PDF and I don't know where to start to do it from scratch! Please do you have any resources? Also I wanna deploy it with azure ! I would be very happy if you help me 😔🤍🤍

[solid oriole]

please dont cross post 👍

[west snow]

have you tried langsmith?

its great when youre not that technical

or opal?

[lone sequoia]

No I didn't but is it required to use openai key? It's not free to use even deployment

[west snow]

you can use the permanent free gemini api key

60 request per minute limit, but should be enough for a chatbot:)

[lone sequoia]

Is it unlimited? It's okey for 60 request per minute

And is it free ?

Without card details

[pine charm]

Yes, if you go to ai.studio you can get a free API Key.
LangChain and LangChainJS have full Gemini support.

[lone sequoia]

Okey I will try it thank you brother