Cody - AI Trained on Your Business

[near sundial]

The magic of ChatGPT but trained on your business.

Use Cody to support your team or function as a highly-skilled employee that can analyze information, brainstorm ideas, and execute work.

It's super simple

1. Upload Documents or Create Articles
2. Create a Conversation & Select which folders Cody should utilize
3. Ask Cody. Cody will provide you with answers based on those documents and cite his sources for extra reliability & trust.

This is a super early free beta, trying to gauge intrest and direction on the tool, all feedback greatly appreciated: https://meetcody.ai/

[stable sundial]

love the idea

well done!

[near sundial]

Thank you! Feel free to try it out free, it's an early beta, a lot more features to come.

[wintry gate]

@near sundial Hey this is really cool? What language did you write this in?

[wintry echo]

very cool!

[sharp pier]

amazing idea , neat frontend and awesome implementation 🤘🏻

[tight smelt]

I would like to make a suggestion for integration:

If there was a way to integrate Cody's responses into Notion, it would be interesting to record and generate data during commercial calls, for example.

[near sundial]

@tight smelt Would you leave the data in Cody or want the data pulled from Notion?

[near sundial]

Thanks! Just Laravel & Vue

[near sundial]

The document uploader is still a bit shaky, but we should have a much better implementation by the end of the week. And some additional personalities soon.

Our main goal for business use was to make sure it did not Lie, since when that's passed on too employees that's a huge liability, so it's very accurate and also cite's it's work.

[tight smelt]

Extract Notion data to generate database.

In this case, Cody would be an API in Notion, am I right?

[tight smelt]

His approach was precise and shows a real difficulty in dealing with AI.

The responsibility that is transferred - partially - from the leadership to the digital interface, is not small.

What is the key indicator that AI is not "lying" in any interaction? Do you have a well defined process for this?

[silent belfry]

following

[subtle zodiac]

following

[stuck dagger]

This is a very cool idea dude. Going to test it out. Before I do, I want to upload a bunch of our copyrighted documents, and wanted to make sure they are protected. I had read through your privacy policy, and it seems so, but wanted further confirmation.

As for a suggestion, (note I haven't started my account yet), it would be incredible to track the questions that users are asking so I can see the FAQ

[rocky jolt]

Yes, I’d like to check privacy too)) otherwise the idea is awesome! Shut up and take my money!

[abstract solstice]

Hey @near sundial I'm also working on a similar project but since I'm new to AI, I'm getting stuck at things. Can you please help me answer these questions:

1. Are you using OpenAI Api's? I believe their API's stores whatever information you send to them to further improve their algorithms
2. Can you share a basic outline of how did you made your project? Only want to know the AI part. Are you doing something like comparing embeddings generated by user question and the embeddings of the data stored in the documents?

[distant tundra]

following

[near sundial]

Yes we are using OpenAI APIs but we are storing the information in an external vector database.

Documents turn into embeddings, using the API, then we run a Search and provide GPT the relevant information to answer the quesiton

[abstract solstice]

@near sundial does that mean that OpenAI APIs doesn't store information that you send to it? Does these API make any network calls or are they just a python/javascript package ?

[near sundial]

No it doesn't Open AI simply creates the Text Embeddings and we store them in a Vector database

[near sundial]

Cody - AI Trained on Your Business

[subtle zodiac]

How do you give the context back to the gpt3 models api? Also is your code open?

feel freenot to answer if you don't want

[near sundial]

We do it via text-embeddings searching a database, once the database returns the answer we tell GPT3 to answer your question using that context

[zinc bane]

This has potential really cool for training employees

[sterile oar]

Cool, this worked well in some text documents. However, with the tabular data it gives some false information

[cold stag]

This is awesome! Is this something you plan on making open source? I would really love to run my own instance of this with my own api key and make changes

[waxen skiff]

is there a way to make the files I upload, dynamic? Or will I have to upload updated versions of the docs

[near sundial]

We are actually working on API to do that!

[near sundial]

Not currently but is also something we are actively developing, but it will have its limitations, more about reading a csv or excel and stating facts, versus actually writing calculations

[carmine shore]

@near sundial Very awesome project! I'm curious, how do you prepare each document for the embeddings? I can imagine that is quite difficult because each document probably has a different structure (formatting and semantics).

[near sundial]

Yeah that's the difference we have, we worked on an algorithm to strip files and format them properly so they are usable. We figured the average business owner is not going to want to be preparing data for an AI, so we tried to make it as simple as possible.

[manic cloud]

Can I share a google drive folder with all my company data to a read-only invited user and then you can use that for training and also keep an eye on it when the drive contents changes?

Also, what are the security implications of this? are the models you create stored in USA?

Could the models be stored on my google drive accessible only to your engine via encryption?

[shadow sapphire]

This is cool, does it mean you prompt the gpt api by text embedding rather than plain text. If so, this is more advance than prompt fine running i guess.

[tight smelt]

Realy!

[terse peak]

I'm fascinated by clean the platform looks, and how your tool is able to answer two questions in one response, that requires info from different files.

@near sundial would you mind briefly sharing how you ended up segmenting potentially bulk input text, that is of random format into smaller pieces before creating the embeddings to be searched? For example, are you splitting the text by sentence, paragraph or do you attempt to select larger pieces of data just under the max token limit using identifiers? This has been a bit of a struggle for me, as I'm attempting something similar to this concept of searching through bulk input text, but more geared towards legislation lawyers.

[river condor]

r

[viscid zephyr]

following

[calm plinth]

@near sundial any idea when it will be out from the Beta? I've tested it is a little bit too slow but it's quite impressive.
Another question is ... it will be possible to allow url navigation?

[tight smelt]

Hello. Something I noticed in recent times of use is that the conversation always keeps scrolling to the end of the conversation, when I want to consult some answer further up. This has greatly hindered the use of the tool.

Another problem is that the help channel is poor in content, which for example, doesn't help me understand the difference between "Factual Cody" and "Creative Cody". The beta version could have more instructional content to make it easier to use and generate more feedback from early adopters.

[ornate herald]

Following. This is awesome.

[brittle tide]

What is the largest amount of text supported by the largest document uploaded at one time?

[void viper]

I did try this out. But, it is very slow and you can only upload docs one at a time.

[chrome anvil]

funny

[dusky pendant]

What is the character limit for one bot ?

[onyx kelp]

Beautiful Concept. Hope you can scale it to full enterprise level production capacity. The potential market demand for such a solution is ginormous....

[halcyon pike]

Hello - very interested in using this for my law firm. Any place I can read more about the privacy settings and how information is kept confidential?

[spiral pond]

fantastic implementation - modular and it looks like an easy product for you to maintain. amazing work

[spiral pond]

if you read terms and conditions the gpt-turbo 3.5 will NOT use what you send for further improve their model AND will only store the inputs it for 30 days (i guess for some legal reasons)

[spiral pond]

for sure he is using embeddings. the question is how he is creating them. i think the interface is a killer.

[spiral pond]

no. the models cant be run locally or on-premise. in addition to the models, the backend that will send information to the models have storage and logs. do not share proprietary data from your company - unless you own the company and is willing to use the platform.

[near sundial]

Thanks @spiral pond for the feedback on the UX. Our next step is to release a Chatbot Widget you can implement on your website, and provide users the ability to customize their Chatbot even further.

[near sundial]

Cody - AI Trained on Your Business [FREE]

[halcyon pike]

@near sundial would you be able to explain the privacy policy a bit or show me where it is on the website? Not able to find it and I'd like to use this for my law firm but concerned about confidentiality implications

[calm plinth]

@near sundial any idea when it will be out from the Beta? I've tested it is a little bit too slow but it's quite impressive.
Another question is ... it will be possible to allow url navigation?
And I've seen there is a limit in PDF weight/length could you let me know wich is the maximum page number/weight of a pdf to be processed?

[near sundial]

@calm plinth

We are updating to ChatGPT API by next week so it'll be very fast, and a lot better in generative applications while still leveraging your knowledgebase for context.

It's about 30 MBs

[near sundial]

We now have a Chatbot option for your website:

[charred lodge]

Yes, very nice project/product. If we can give role prompts to the bots that would be great. That allows for your HR folder example to let it act as a HR manager. etc.

[cunning bane]

Very nice, but it is probably trained on GPT~3, not ChatGPT.

As there is no ChatGPT API

[near sundial]

ChatGPT API (GPT 3.5 Turbo) was released last week, we just implemented it

[near sundial]

Hi Rene, that's our next update which will give you more control on the prompt. Currently if you goto Settings > Bots you do have a few options, which will be expanded.

[jolly oriole]

cool

[charred lodge]

That is great to hear. It looks like your filtering less than the openai interface on sensitive topics. That is great for putting docs into a broader and more critical context. A great plus! Keep that going.

[stuck dagger]

@near sundial I asked a question a while back and may have missed your response. Could you clarify on this?

[thorn venture]

From previous messages from the original poster, they use semantic search which consists of finding the most relevant looking parts of the stored text. To do that you encode parts of the text as vectors and finding similar parts of the stored text means finding vectors in the same direction (if the vectors are normalized and point in exactly the same direction it means they are the same). These are high-dimensional vectors.

After the semantic search is performed, These bits of plain text are given to ChatGPT to make a synthesis of the information to provide a likely answer to the query.

The information is fed to chatgpt in plain text but searching for the information required embedding the text in a high dimensional vector space.

There are limitations to this approach.

• The semantic search might not find bits of information that require some level of thinking or/and imagination to know that they are relevant,

• Some queries require a global understanding of the document/documents and how parts of the document are connected. A simple search for the most relevant-looking parts of the text is not enough in some cases. Maybe an example could be questions about the development of a character in a story which could require understanding the logic of events that happened and how they influence the character and in what order they happened.

Basically, some queries require a sort of knowledge graph of the document(s) that would provide information about how parts of the document are connected. This is somewhat in contrast with breaking down the document into pieces of seemingly relevant text.

A theoretical AI language model with very large memory and thorough analysis skills might be able to answer most questions on most documents but that is not the current situation and so we will have to implement partial solutions for the moment.

[open escarp]

How has the performance and response been so far with GPT-3.5-Turbo? Any issues still with speed or knowledge retrieval?

I’m keen to sign-up now but thought I’d still ask in case anyone else here is also wondering.

[shadow sapphire]

I think the lastest version is GPT-4 mode, which should have better performance than 3.5-Turbo, turbo is just faster processing than 3.5 as it means

[near sundial]

Your information is stored on AWS, one of the most secure cloud storage platforms in the world. Plus, we make sure to store all your documents anonymously for an added layer of protection. To provide you with the most accurate answers to your search queries, we share small portions of relevant data with OpenAI, but rest assured that we never share the full document. We're based in Texas, we do have a privacy policy on the homepage, but it is currently being improved by our legal team, as we just launched this beta a few weeks ago. We're actively working towards achieving SOC 2 compliance.

[near sundial]

We had just upgraded to GPT3.5 but are now in the process of implementing GPT4 shortly.

[stuck dagger]

Awesome. Thanks for letting me know.

[near sundial]

No issues, its' actually been pretty fast! It's opening a lot more opportunities to do pre & post query processing without such a big delay.

We just got accepted to GPT4, so we are looking into implementing that next after a few major features we release next week.

[brittle mirage]

@near sundial what did you use for the video?

[near sundial]

This tool! https://www.screen.studio/

[near sundial]

Yeah we are making sure, if we are retrieving information, the AI is not just offering unwanted opinions unless stated. That was a request from a user, that noticed our AI liked to chime in, on his document data, which he was using to write an article on a sensitive topic

[faint spindle]

@near sundial Do you offer this in a white label version?

[finite venture]

@near sundial youve done a wonderful job mate

thank you for that

[faint spindle]

@near sundial I tested yesterday. Speed is initially slow. Some of my .pdfs failed on upload. Not sure why. Also noticed that Cody was referencing invalid docs

Overall, I see it becoming a very good product

[near sundial]

@faint spindle Thanks! If you have issues with the PDFs you can also try creating a document directly into the app. We are improving the document handling at this moment

[fathom stirrup]

@near sundial can this work with alpaca7b?

[near sundial]

We don't support that model yet we are currently working on a GPT4 upgrade.

[finite venture]

do you guys have any announcement channels

for new updates

[near sundial]

Yeah we have a discord you can find the link in the Get Help button on the app and you can find us on twitter (@meetcodyai)

[near sundial]

We also just deployed GPT4 momentarily if anyone wants to test it out

[golden lava]

v cool !

[golden lava]

Add GPT4 as a tag on the api-projects thread

[near sundial]

Cody - AI Trained on Your Business

[stuck dagger]

That discord link is broken

[near sundial]

Thank you we'll get that fixed!

[safe dove]

Am about to start using Cody, GPT 4 is integrated?

[near sundial]

Yes, we tried releasing it platform wide but proved to expensive, so we are releasing it case by case. If you'd like to message us on the Get Help, we can enable it, itll also be available in our premium plans soon

[near sundial]

We have a discord channel you can find on the site

[viral forge]

@near sundial interesting. I'll introduce Cody to my boss. If he's interested, would you be available to give us a live demo?

(us=Pretty large Company. We can discuss by mail)

[blazing ferry]

I need that so badly ! But from what I see it does not support French, right ?

[near sundial]

Sure feel free to email me at [e-mail]

[near sundial]

And yes it does just goto “New Bot” and on language select french

[blazing ferry]

oh nice ! thanks !

[blazing ferry]

Gosh. It's amazing ! He can answer really intricate question about local laws.

[molten wren]

Very cool interface. Very similar to our offering https://customgpt.ai/ - which supports ALL business content (like websites, helpdesks, CRM, 1400+ document formats, Youtube/Vimeo videos, etc) -- do check it out.

[near sundial]

Thanks, Wish you guys the best of luck!

[scenic estuary]

Love this idea!

[near sundial]

Thank you! We just released our website crawler as well.

[oak mulch]

amazing idea, just what my business needs actually

[prime breach]

What's the data limits on that? Couldn't one do an internal bot as well, or does it need to be publicly accessible data?

[stable delta]

what's the difference between your idea and customgpt.ai?

It's cool ! But i wonder how you guys do to fine-tune every customer's model, though they are so different and how to generate enough high quality questions and answers.

[near sundial]

I'm sure they are similar setups, you can check-out our tool, has different modes of deployment wether our team UI for desktop as well as Chatbot Widgets for websites. We also support different models such as GPT4 & GPT3.5 and have an easy UI to manage your knowledgebase.

[sly totem]

Somebody knows if exists a open-source project that can serve as your OWN knowledge base? Similar to this, only asking for not making my own project when I could improve existing one

[bitter shard]

This is a great idea. Is there a way to have the output generated collectively as either a word document or a pdf?

[old oasis]

Stopping by just to thank you. Thumbs up for the efforts!!

[near sundial]

Not at the moment, but what sort of use case did you have in mind?

[bitter shard]

I am using GPT-3.5 to apply for grants, but then it is very tedious to a;lways copy and paste the prompts and responses. I have been working on a way to integrate GPT-4 to do this. Any advancements you might have made?

[near sundial]

Not at the moment but a document editor is something we do have in our roadmap we'll keep you posted! I

[faint spindle]

@near sundial I am trying to build a CB that will follow a 10 question conversation, then reference the conversation to formulate a new structured output based on the uploaded reference doc on how to build the output. How can I do this with Cody?

[harsh fox]

I try to Image some usecase: Can I train the KI with this tool? For example Uploading a pdf of a textbook and then tell the KI to answer with the knowledge of this book I provided?

Is there a list of possible usecases?

[near sundial]

Yeah you can do that

[merry rose]

Testing, thank you!

[calm plinth]

@near sundial I wrote an enquiry for a business call some time ago, Angelo from Italy I had no answer...
anyway I would like to Know if it will ba possible to:

1. use API building a proper user interface with my domain
2. any example of customi bot window (i have no example on site)

[small arrow]

cool

[near sundial]

We currently don't offer an API but do plan to offer it by the end of the month.
You can embed the bot currently on a website, and remove the powered by cody if you're a premium subscriber

[calm plinth]

Ok thank, so , at the moment, you can only remove powered by Cody or you can cuatomize also the look n feel?

[open briar]

+1 for this project from a happy new customer!

[near sundial]

Thank you @open briar !

[molten wren]

No data limits -- customers have built bots with many gigs worth of data. It does not have to be public data (there is now an upload option where you can upload PDFs or documents in 1400+ document formats). see this example https://adorosario.github.io/LevinBot/index.html -- built with 30 years worth of research documents.

[vital flame]

@near sundial What vector database are you using?

[near sundial]

@vital flame We use Pinecone!

[vital flame]

Thanks! The free plan is not released. I have to get on the whitelist.

[near sundial]

We have the look and feel options releasing this week along side a few other features

[open escarp]

Is it possible to cc-in shared collaborators for conversations?
I.e. I want to have an LLM-assisted group chat with a team member, accountant, etc., amplified with provided docs/context and conversational memory.

[vital flame]

Hi, I need help understanding the registration flow and usage of OpenAI + vector database pinecone.
I tried using Pinecone, but I'm having difficulty understanding when I generate the embedding and register it in Pinecone's index. And then displaying the conversation history.

I plan to develop a Node.js application where the user can register texts. And via chat, they will be able to ask questions about the registered texts. I will also need to display the user's conversation history.

[near sundial]

Not at the moment, but that's interesting, so you'd want to have, you and your colleague messaging the LLM in the same conversation?

[tawdry thorn]

@near sundial love this idea. I have a connection with a chief security architect. If you pair this platform with the ease of mind the businesses want when uploading sensitive information I think this can be a home run to land larger ticket customers. DM me if you like an intro.

[open escarp]

Yeah pretty much this - like a multiplayer, LLM-amplified conversational interface in which I can add collaborators. Ideally the chatbot would have basic system prompt instructions, file uploads, parameters, memory, external db integrations, etc.

Data recall would be nice to have as a standard, but personally I’d be more interested in using these sort of shared interactions (humans + LLM) for ops/strategy/product development.

[exotic spindle]

Its nice - good job! 👍
Is there some API or something too? So you could implement the "new" model in a discordbot?

[near sundial]

We have an API coming out in like 2 weeks!

We just release our widget customization options as well

[exotic spindle]

nice, thanks for the answer! 🙂

[gaunt swan]

API is coming out in a couple of days, DM me if you're looking to beta test it.

[near sundial]

Our API is now available: https://developers.meetcody.ai/

[exotic spindle]

Nice thanks for it! 🤝

[twin thistle]

Do you have an API for content upload in real-time?

[umbral ridge]

create a form

[near sundial]

We have that coming in July, currently is just conversational API and Bot Selector

[near sundial]

form for what?

[tight smelt]

I keep running into issues when trying to paste codeand i get a server error

[exotic spindle]

API?

[old oasis]

Nice update @near sundial thx again for the effort

If I might suggest something...when sharing the public link for the bot, it will be great to have a link preview, "rich preview" feature...your logo at least 🙂

[near sundial]

We just added the logo feature but yeah thats a great idea

[near sundial]

As requested a bit late but now we have the Slack integration & Discord rolling out next week!

https://meetcody.ai/blog/slack-ai-integration/

[exotic spindle]

very cool! : D

[exotic spindle]

@nocturne oasis could you delete this please? :x

[mystic cairn]

JFYI: when I try to use the help to start creating content in google chrome is not working

[near sundial]

Thanks @mystic cairn We'll have the team look into it!