Visus.ai - Train your own ChatGPT

[lilac turret]

Hello all!

The first time I tried ChatGPT, I saw its potential to transform how we access and interact with knowledge.

I saw how the technology could be used towards a smarter, more intuitive, and more efficient approach to knowledge retrieval, marking an end to the days of sifting through endless search results to find the information we need.

That's when I went all in on creating Visus.ai. More than 60 days later, it's here!

Visus.ai makes it easy for anyone to train their own ChatGPT-like assistant, which is trained on their own knowledge base.

Create your free account 👉🏼 https://app.visus.ai

⭐️ Train your personal ChatGPT-like AI assistant on your own knowledge.

⭐️ Unlimited files: with a click of a button, train your own AI on thousands of files.

⭐️ Ask, don't search: Ask any natural language question and get a concise ChatGPT-like answer.

⭐️ Source-based answers: Need the source? Visus references your documents when crafting an answer.

⭐️ Secure, and configurable: Ask any natural language question and get a concise ChatGPT-like answer.

⭐️ Dialogue: Discuss your favorite book, your essays, or anything you've written!

• • •

Free trial included: 300/messages per month for FREE!

Create your free account 👉🏼 https://app.visus.ai

• • •

We’d love your feedback and will be available for any questions.

Visus is the product it is today because of feedback from our community, so please share anything you have in mind. 🚀

It's free to create an account!

Create your free account 👉🏼 https://app.visus.ai

[honest raptor]

not free

8 messages

per day

[lilac turret]

Hey @honest raptor the daily limit is 10/day and there are additional ways to get extra messages that we will be releasing soon. Do you feel the amount is too limiting?

[honest raptor]

Very

[lilac turret]

Noted, what's the use case you want to use it for?

[carmine cosmos]

So, it’s just like fine-tuning?

[lilac turret]

Not really, but we do a lot of fine-tuning behind the scenes for different tasks!

[red umbra]

if i was wagering a guess it would be more like indexing to a local folder and having the language model converse over the contents of the index

great idea for a web site over all

i hope it flys well

[lilac turret]

That's a good description yes!

Thank you so much for the kind words

Did you try it out?

[red umbra]

i havent i have a bot that has an index function tho. outside of training new models thru open ai. it seems the best way ive come across for referancing datasets that dont exist in the model without a google search format type conversation

well its not MY bot.. i just use it

[lilac turret]

I see, if you do, I would love to hear how ours compare to the one you are currently using

Like general intent identification, question structure, source referencing

[red umbra]

lol but querying against data from pdfs and such is awesome handy using the language model to desribe the data.. you never know if whats in the model is true or false in a generated answer so having it sumarise and converse over a local data set imo. is the best path with that end .. training a new model is prety intencive a process

il send you a dm

that way you can check it out and weigh it against your build

never know might find some ideas that inspire the website build

[lilac turret]

totally! There is a lot of use cases and I'm looking forward to see what the community does with it

[elfin swan]

are you a bot or are you using chatgpt to answer our questions or your acting like a bot by purpose or you want to sound professional by talking like that

[lilac turret]

No, some of us are actually professional and do speak like that. You need to use Discord less, socialize more.

[open lichen]

He's basically saying you need to learn how to tell the difference between AI and humans talking via style, tone, etc. (I hope I'm not wrong in this)

[lilac turret]

writing bursts of text

discord style

without thinking...

gets old real fast

so I prefer just typing a full sentence

and give the grammar an extra look

maybe i find some misspellings

maybe not

lmao

Anyhow, Discord is not the best place to have a mature discussion. The DMs I've gotten after posting here are weird af.

I'm posting about a product that my team & I spent over 2 months building, if you want to discuss the product then let's do it I'm more than happy to do so, but if you're going to be immature then I don't see how this is the right space. There are literally millions of communities where you can meme around.

[hushed agate]

Need PDFs

[lilac turret]

You mean upload Pdfs?

[marble monolith]

Hey, Farouq, I just tried the product for a while. It looks amazing! Just curious, did you train other AI model built on top of ChatGPT? When I ask bot about something related to my uploaded document, it gives sources with a confidence score. Is this confidence score calculated by some other custom model you trained, or models provided by OpenAI?

[dapper belfry]

@lilac turret I tried asking it a few SAT questions with docs containing the sat passage and it performs better than chatgpt. Do you mind sharing the methods that allowed this model to understand the context and answer this much better?

[dapper belfry]

Also, are you using a chatGPT api, or some other tool to summarize a text and respond?

[hushed agate]

Its amazing really

[elfin swan]

💀

[sonic pecan]

My data base is in clickup documents

How can I utilise this “train your own Chatgpt” ?

[little yew]

people read visus as virus.

[lilac turret]

Hey, thank you very much for the kind words. We use a few different models for different tasks, and yes as you mentioned the confidence score is determined entirely by a separate model. We are constantly tweaking it.

[lilac turret]

There is no ChatGPT API, but we are using a combination of different models, some of them are OpenAIs Davinci 003 & Ada.

[lilac turret]

Glad to hear that the performance is better than ChatGPT, that's really good.

[lilac turret]

Sounds like a skill issue to me

[wraith hare]

can i upload some files in the contents?

[lilac turret]

Can you please elaborate, what do you mean by uploading file in the contents?

[errant agate]

wow

the answers are so good

thx for this

[dapper belfry]

Do you know if it can give not just its sources, but its own answers/responses a confidence score?

[lilac turret]

Thank you so much, glad you like it! 🙏🏻

[lilac turret]

It's technically possible, might be unreliable though. What difference does that make to you?

[dapper belfry]

Ever since I started using chatGPT, I was always wondering how much I should rely on the answers it gives me. I just thought that since the model is being trained from a small doc, it can give a score of how well its interpreting the data.

[lilac turret]

Our AI in strict mode is pretty reliable.

Try it out, ask it unrelated questions and see how it performs.

Matter of fact, we just launched an update to it today

[dapper belfry]

@lilac turret Wow. I tried it before setting it to strict, and it gave an answer of imaginary for this question "“Hypothetical” most nearly means ? A) imaginary B) actual C) possible D) interesting". After strict mode, it outputted "possible", which was the right answer. Crazy

It's probably incorporating the other definitions of hypothetical more without strict mode being turned on, but I can't believe how much better it understood the text with this mode.

[lilac turret]

Amazing to hear! Try it more and let me know.

We're working on mind-blowing things behind the scenes, and I mean it.

This is our v1

[sonic pecan]

@lilac turret My data base is in clickup documents, How can I utilise this “train your own Chatgpt” ?

[lilac turret]

We are building out our Clickup integration, but for now you can simply copy & paste your doc from Clickup to Visus

[sonic pecan]

i added a rules document

and it doesnt know anything

[lilac turret]

Did you sync the document?

[sonic pecan]

how so?

[sonic pecan]

why is it even an option shouldnt this be auto synced, as well as auto save?

[lilac turret]

Why do you think it's a good idea to auto-sync and auto-save (overwrite)?

[sonic pecan]

tbh its a nice idea to have my own chat gpt based on my own database

but why is there a limit to the amount of messages?

and what does it cost to have unlimited acces?

[lilac turret]

$20/m for unlimited, for personal accounts.

Send me your email in DM let me reset your messages so you can try it out with the synced doc

[red umbra]

this whole chunk of reply ... so well done lmfao.. i applaud you good sir

[marble monolith]

Hey, I found something interesting. I feed the model with my interview prep document. First, I ask model to explain my past project called Briefly: it got the answer right. However, when I ask it about action section, it give me result that was part of another project, not the same one I asked before lol.

Does model carries information of previous chat histroy? Or only the context provided in database?

[lilac turret]

Haha someone appreciates it ;D

[marble monolith]

Also, the Add to Focus List got several errors when I click it.

[lilac turret]

Here's the deal, we do both things you mentioned above.

However, one of the biggest challenges was dealing with ambiguity - the AI is not that good when a document can answer many different questions

Yes, the focus list is exactly how this problem is addressed

but the error is not really how the intention was haha. Thanks for reporting the bug

[marble monolith]

Oh I see.

[lilac turret]

Are those only console errors you're getting or is the app crashing?

[marble monolith]

Yeah that's all errors I got. Also, app is crashed when I click.

[lilac turret]

Can you refresh the conversation and try agian?

[marble monolith]

I tried refreshing, however, it seems something is keep running in background and prevent me from doing any clicks.

Probably an infinite loop?

[lilac turret]

So you're unable to use the app at all or only when adding a source?

[marble monolith]

I can, but I have to go back to reload page.

The error only occurs when I click Add to Focus List

[lilac turret]

I see, thanks for reporting I think I know the issue

[marble monolith]

Also, may I know how's each document being index? My initial thought is that we split it into several chunks based on heading1, heading2, etc. For each chunk of text we calculates it embedding and build a KNN on top of that, and compare similarity between document embeddings and user input embeddings?

[lilac turret]

I just deployed a new version of the frontend that should fix the bug you saw above. It should be available in ~3 minutes

[lilac turret]

The basic idea is embedding yes, it all depends on the use case and your app structure

[marble monolith]

I see.

[lilac turret]

can you please refresh the app and let me know if you still have the bug?

[marble monolith]

I think bug is now gone. Thank you

[lilac turret]

Wonderful!

Send me your email in a dm, I can reset your messages for the inconvinence

[marble monolith]

Sure.

[hushed agate]

How to change this?

[hushed agate]

Nvm found it

[errant agate]

I did a small experiment to compare the "Relaxed", "Moderate" and "Strict" mode of the AI. I love the results

Here's "Relaxed" mode

Here's "Moderate"

and finally, "Strict":

[lilac turret]

this is amazing, can you post in the Discord as well?

[lilac turret]

Let me know if you need anything else!

[glacial halo]

Guys, how can I modify the prompt directly through the API, so that when the user types a food, the gpt chat will bring a whole structure about the food, such as name, country of origin, ingredients, etc. All this in an organized way and using emojis. Can someone help me

[lilac turret]

Are you talking specifically about Visus?

[hushed agate]

Sure

[hushed agate]

What language are supported

[lofty acorn]

@lilac turret could I train this model to use an LTSM? If so could there possibly be other integrations?

[lilac turret]

We are focusing on English only for now.

[lilac turret]

Can you please elaborate?

[errant agate]

can we attach a PDF to be downloaded inside a document?

[opaque relic]

Very interesting. Can’t wait to give this a try later. Having been testing chatbase which does something similar. Will let you know my experience with visus.ai

[lofty acorn]

Could we hop on a call? it is a lot to explain. However I believe the overall idea would be worth it.

[lilac turret]

Please do - have you tried it yet?

[lilac turret]

Please email me at [e-mail] with details

[lilac turret]

We have this functionality in the roadmap!

[hushed agate]

👍

[sharp laurel]

Two questions:

Can you read a google drive folder?
Where does the data get stored? This is amazing but I have real concerns about private company data being indexed on a USA based server (I'm canadian)

[lilac turret]

I understand your concerns. If they are due to regulations, then unfortunately there isn't much we can do right now as all our servers are based in the US.

With that being said, servers holding user data use full disk, industry-standard AES 256 encryption.

[lilac turret]

We are launching our marketing website this week hopefully, for now to access the app head to https://app.visus.ai

[empty umbra]

I’d also be very interested in a Google drive integration 👍

[gray valley]

Good @lilac turret . How reliable is this compared to proven semantic search tools, which is open source? Have u benchmarked ur results with any of existing tools.

[lilac turret]

We are working on it, sending you a DM I would love to tell you more about our approach and if there are any blindspots we are missing in terms of security

[lilac turret]

Semantic search is a drop in the ocean when it comes to AI assistants. Relying on it alone as is, is a recipe for disaster.

We have our internal benchmarking, and we optimize for correctness in answers.

From the feedback we've received, Visus performs much better than the available alternatives today.

[echo bramble]

I uploaded an blog post about marketing automation and sync with ai and asked a simple question. I waited about 5 mins, and did not answer. I hope this bug report would be helpful for your team 🙂

[sharp laurel]

Please DM to get my email if you have a news and updates email list. I'm definitely interested in the Google Drive if that becomes available.

Side note: we use google docs a lot and I suppose our data is already on USA servers already

[lilac turret]

Sending you a dm, let's talk

[lilac turret]

Hey! Thank you very much for letting me know, we didn't have any server downtime yesterday so could most likely be a client side bug. Are you still having this issue?

[neon shuttle]

Link to your discord server is down. Can I get one? I have a lot of questions comments.

[lilac turret]

Ah I think the link expired
https://discord.com/invite/5TUJPBDmHy

Can you try?

[sterile lintel]

@lilac turret

You have been warned and muted by AutoMod.

Reason: Discord server invite

[timber turret]

Is it possible to use my own API key for this? Or do I have to pay you directly? Sounds like a cool idea either way.

[gray valley]

What alternatives u compared with? Have u explored FAISS and many more models.

Quick question. Can it read a weblink or google drive links and useful for q&a. Also, how is user data files secured?

[lilac turret]

It's not possible to use your own API, because OpenAI models are used in combination with our own models to power Visus

[lilac turret]

Yes, we are accepting early adopters for our gdrive integration. Let me know your use case. Feel free to send a pm

[lilac turret]

Huge update on the way today!

[runic wigeon]

Hey @lilac turret can I upload a document with 10,000 words? I assume you use vectors. Can your vector database store 1 billion vectors? Can your bot retrieve the top 3-5 sources based on the text. Feel free to pm me.

[gray valley]

@lilac turret , how is privacy and security maintained for the files uploaded? If we upload our private files, where are they stored and how is it controlled.

[lilac turret]

We do use vectors yes, you can upload 10,000 words. That's not a problem. Do you mean 1 billion individual vector sets?

[lilac turret]

All data is encrypted in transit using TLS 1.2+ with perfect forward secrecy. Servers holding user data use full disk, industry-standard AES 256 encryption.

We're actively working towards achieving SOC 2 compliance. Can share more details around that if relevant.

[hushed agate]

Can I upload my CSV?

[lilac turret]

Very soon, what use case do you have in mind?

[runic wigeon]

Yes whether you use pinecone or another vector storage system that can host millions or even a billion individual vectors for large datasets

[hushed agate]

Ask it about excel forecasts, get answers on stocks etc

[fleet tusk]

An AI which can read and solve steps by step, a simple second differential equation like y''-4y'+4= sin (x) .
This can be useful for brother teachers, students and parents

[lilac turret]

Well have you tried Wolframalpha? They do an amazing job at that

[past palm]

How have you been working on this for 60 days if the gpt api just came out?

[lilac turret]

There are countless models you can use. You can also trian your own. The new ChatGPT API is one of thousands.

[past palm]

Sorry I'm very new to this space.

[lilac turret]

No worries, I edited my comment. Let me know if you need any help I'm more than happy to.

To directly answer your question, we are using a multitude of models to power Visus, some are our own.

Prior to ChatGPT API, Davinci was available which is arguably better than ChatGPT (albeit more expensive).

With that being said, we are already experimenting with the ChatGPT API and are trying to have it in production next week

[past palm]

Ah thx for explaining. Good luck

[fleet tusk]

Yes, but it only shows you the result, not steps, however what parents and others need first is steps need to aproach the solution, not only solution.
And i tested your API and it doesn't work when we turn your connection mode on 3G only 😩 🥲

[lilac turret]

No worries, thanks! 🙂

[errant agate]

I can see step by step on Wolfram

it costs though

[frosty lava]

this is not your own trained chatGPT but another external system. So misleading

[fleet tusk]

//

[lilac turret]

Oh sorry I couldn't provide you with your own ChatGPT license from OpenAI lmao.

Of course it's an external system, I'm not from OpenAI.

[frosty lava]

If i go in projects I am looking for something I can roll myself. Not just visit another website that connects to OpenAI. Unfortunately all the so called "projects" are just products and not projects that you can go and look at to learn more. That has nothing to do with a free license or any other random expectations.

[lilac turret]

Are you sure that you are actually looking for something you can roll yourself? Because in that case you're most likely looking in the wrong place.

We connect to a multitude of models, some of which are our own. We will not allow access to those models the same way as we won't expose the exact algorithms that we use for knowledge retrieval.

I don't understand why some people feel entitled to be able to roll the system "themselves". We've spent hundreds of engineering hours to get the product that we built, and we are not going to release it as open source because:

a) We are not interested in open-source as a monetization model. People work hard on this, and they need to get paid.
b) 99.99% of people won't even bother implementing Visus if it was open-source because they lack the experience required.

There are countless free papers being produced every week on all kinds of topics in regard to LLMs, and many of these papers include Github source codes. There are countless FREE models available on Huggingface that you can use instantly. There are HUGE communities where you can learn more about LLMs. The resources are out there.

Saying that "most projects are just products" is misleading. For each product available there are probably hundreds of papers and open-source projects available.

The missing part is doing the hard work. The resources are available, it's what you do with them.

[molten pewter]

this is a really cool project, love that you can pick how strict it is on sticking to the documents or not. shame it's just a few messages you can try before the wait, but understandable, probably not cheap to host

[lilac turret]

Thank you very much for the kind words. We are looking into increasing that limit

[frosty lava]

Then dont call it and advertise it as a PROJECT, it is a product. You are the one who feels entitled to false advertise a PRODUCT and call it a project.

[frosty lava]

And no this is not the wrong place to look for PROJECTS, it is the projects channel. You should not take space here and waste people's time when they are looking for projects instead of a PRODUCT. As you mentioned a PRODUCT can be made of different projects just like yours. The fact that also you are so arrogant in defining who is entitled to do what is also quite unsettling.

Again this is not the right place. I am not a customer for a product. I am looking for projects from other people to learn and not to consume.

[lilac turret]

I don't care if you are a customer for a product. I don't care that you are looking to learn. You don't define if this is the right place or not. There are 279 projects as of now in this section, many of which are paid products. Deal with it.

If your intention is to learn then be polite, ask questions, and ask the right ones. I've helped countless people throughout the years, over 10 from this Discord post alone in my DMs.

Requesting access to private repos is entitled, lazy, and frankly insulting. I don't reach out to OpenAI and ask for their source code because I want to "learn".

You want open-source? Good. Go make your own project and make it open-source.

[timber turret]

even tho I hate paid stuff (don't we all), I personally have to agree with this, people have spent a lot of time on their code (even tho I don't always respect it lol)

[neat fox]

so is this a way to give chatgpt a long term memory about me? because thats exaclty what im looking for, although if its true that it will just give me 8 messages a day its kinda useless for my intended use case

[lilac turret]

For sure, no body likes paid stuff, but ultimately everything is paid in a way.

[lilac turret]

Yup, long term memory is a good way to put it.

[timber turret]

is there a free trial for this btw?

lol, i didnt see this

I am trying to "sync with AI" but it keeps bringing up a page with nothing on it

I disabled brave protection and "dark reader"

maybe the site is down/broken?

cuz, when I go to the "documentation" it says to "try again later"

[lilac turret]

Ah I see, please try to separate the sections in your doc, no section should be longer than 1000-1500 words. Each section should be separated by an H1/H2 tag

this is a temp. fix as we roll the patch in the coming update

[timber turret]

could you please elaborate? what do you mean by a section?

ah, I think I might get it, you mean break it up into multiple files right?

nope, I trained it successfully, but I asked it to code in my "trained" programming language, and it said it can't help

[lilac turret]

Yes exactly.

[lilac turret]

That's not the intended use-case, for now it's mostly for Q&A.

The next update to the model might perform better

feel free to send me a document of your programming language, might advise on how to structure it better so the AI understands

[errant agate]

Getting blank screen when saving

[errant agate]

@lilac turret ?

[lilac turret]

Please try again now?

[errant agate]

I'll try. Thank you!

[stark sorrel]

So it's not possible to delete the account?

[lilac turret]

not from the dashboard, I can delete your account for you though

[lilac turret]

Our new website is live! https://www.visus.ai

[errant agate]

wow!

[errant agate]

@lilac turret Greetings, sorry for the ping! I just have a question is it able to like write a Anime fanfiction novel? With for example 50 chapters each chapter consisting of 1k words. And can i teach it on other Anime novels?

[lilac turret]

In the next model it should be possible.

You can give it general guidelines as document knowledge, and it will use those to base its writing

[manic wadi]

@lilac turret nice work. Who do you think your paying customers will? What problem will this solve for what industry?

[lilac turret]

Thank you! It's mostly businesses right now, making company knowledge accessible

[errant agate]

Hey @lilac turret, looks interesting. Can I push this to slack for instance? Do you have an API key that I can connect to a our custom slack app?

[lilac turret]

We are working on our Slack integration. Can you DM me your specific use case please?

[hushed agate]

Is gpt4 ready?

[lilac turret]

Core

• Added file upload support (PDF, Docx).
• Added file deep training - when you upload a file, Visus will be deep-trained on the contents, leading to even better answers.
• Support for GPT4 added.
• Models that are not set to strict can now do more than question & answer. You can create emails, blog posts etc. based on your files.
• Base model updated to utilize ChatGPT 3.5

QoL

• Increased message limit to 50/day for free users.
• Improved file management by adding support for bulk deletion, training, and untraining of files.

[woven trail]

Hello, i did create a new team in visus, trying to upload PDF and it says upload complete, after that nothing happend in content, doesnt appear

[real crater]

Hey, I get no update after uploading a file.

Also no chat.

Any idea what's wrong?

[lilac turret]

Hey! This has now been fixed, please try again

[lilac turret]

This is fixed, can you please try again?

[woven trail]

seems works now 😄

[lilac turret]

Peerrrrrrfect!

How's visus so far?

[real crater]

It's not letting me log in or load https://app.visus.ai/

[lilac turret]

We had some minor outage, can you please try again now?

[real crater]

Tested it out now. It works.

Though not very accurate. I'm also building something similar, though not exactly your app, which serves many general purposes.

Do you know what we're supposed to do to get access to GPT4? 3.5 is not really that smart with documents.

[lilac turret]

Thanks for the feedback, what wasn't accurate? Please do share a few example prompts.

GPT4 is available on the business plan for now.

[vestal belfry]

I tried to upload a pdf file and nothing happen but creating a blank document works

[lilac turret]

Hey, what PDF are you trying to upload? Working just fine

If you can send me the file in a dm, happy to investigate

[manic summit]

@lilac turret first off love the project, did some sample pdfs and very cool! Any idea on one could use your app for online documentation?

[vestal belfry]

@lilac turret i tested this morning and now it works 🤔 👍

Agree not really accurate. e.g i ingested a document where there is some emails contact, then i asked if there a way to found a contact, the answer was correct it says yes there is an email, then i asked what is this email, it gave me what is an email.

[lilac jackal]

Will the chatgpt also consider entity level details like how should respond to a manager and how to am employee.
Eg. If manager askes for employees salary , it has to display but the same if an employee asks manager salary ,it should not respond

[lilac turret]

I'm not really understanding, did it get it right or wrong? 🙂

[lilac turret]

thank you very much for your feedback! We are adding our web crawler really soon.

For now, you can just CTRL + P (print) a page and save it as a PDF. Works just fine 😎

[vestal belfry]

No error, just not accurate

[lilac turret]

No, what you can do instead is create two separate AIs. One for managers, and one for employees. That's the recommended and safest approach.

Let me know if you want me to teach how

[lilac turret]

Oh I see, have you tried tweaking your documents?

For example, if you create a blank document from scratch and write the contents manually you have much more control over the contents and how Visus interprets the text

[vestal belfry]

Gotcha, but let say my doc already exist and i can't/wan't to change it. could be issue if there is some charts in this doc?

[lilac turret]

Then Visus will do its best to understand the document. You can create supporting documents that cover the blind spots in that case.

Charts right now do not get interpreted

[lilac turret]

Minor updates to the interface completed.

[chrome bronze]

Are you planning to release an API in the future?

[lilac turret]

Yes, @chrome bronze . What do you have in mind?

[real crater]

I sent you the prompts and answers via DM.

Do you mean GPT4 is available in Visus' business plan, or in OpenAI's business plan?

I'm looking for access for the OpenAI GPT4 API. Not sure what else to do other than wait.

[lilac turret]

Yes, the Visus business plan uses (along with other models) GPT4.

In case you want to access to the GPT4 API, then you need to apply for access.

[small gazelle]

Yooo I need something for law school bro can you help me

@lilac turret I need a law school app I have 4 weeks to pass my classes. Can you product save me???

[lilac turret]

I don't think so to be honest.

[lilac turret]

Visus 7th April update

Core

• Improved the knowledge retrieval algorithm (you should get better answers).

QoL

• Significantly improved upload speeds.
• Significantly Improved the UI on mobile (fully usable).
• Fixed bug where files were not appearing after uploads

[errant agate]

How to did you do this
Teach me your ways master oogway @lilac turret

[lilac turret]

Haha, what do you want to learn?

[lone heart]

I would, like them most likely, personally like to know how to just get GPT models to work off of PDFs that I feed them. It would be incredibly useful for specific topic-matters in my research for University. Too bad this stuff is cropping up mostly half way through my last semester of my degree.

Also, I've already had to cough up some extra money for GPT Plus. I don't think I can afford another $20 a month. The price is a little steep for people in my position, to be honest.

[lone heart]

Passing law school, especially with AI help, doesn't mean you'll pass the BAR exam. Good luck. I'll personally be going without AI systems as much as I can when I get into law school.

[lone heart]

Soo... Idea: Possibly, maybe there could be a feature where people can utilize their own API key in exchange for a much reduced subscription fee to your service?

[lone heart]

Also, I am noticing that when I put in prompts, even small ones, I am getting an error saying that something has gone wrong. However, it is still using up some of those free daily messages.

[blissful sigil]

Edge with bing does that ;p

[lone heart]

It does, but it doesn't allow me to pull quotes out of papers or books, that I may have forgotten where they are, but can give a vague enough description to an AI to search it for me.

[small gazelle]

I can take the bar exam as many times as I want. I only have 1 shot with school exams. Why can I create a personalized tutor to help me learn the information?

[lilac turret]

I get it, but do you think that you need more than the free 50 messages per day?

[lilac turret]

It should be a visual bug - if you refresh, do you still see that your messages were deducted?

[lone heart]

When I have it struggle to read just one PDF, it does appear that way. I spent about 25 different free usages earlier today trying to get it to actually give me accurate quotations from a text without completely hallucinating the answer.

Also, I have a question for you. The citation it gives as a number isn't hyperlinked at all, and there is no way to tell what document it is referring to by just the number. Is there a way to tell which documents it is actually referring to without asking it for the title?

[stark haven]

it does kinda feel limited ngl

but atleast you can still jailbreak it (kind of)

[lilac turret]

Highlight the message and press the "..." icon, then show sources to view all the refered sources. I agree, it could be a good idea to view them by clicking.

[mellow sleet]

Any support for self-hosting this with your own API key and infrastructure? 🙂

[lilac turret]

Hey Kaveen, unfortunately not as of now! ☺️

[lilac turret]

that's interesting 😅

[mild solstice]

@lilac turret 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 Visus?

[lilac turret]

Minor optimizations to query speed added.

[lilac turret]

Embedded chatbox support in staging 🎉

[ionic vine]

are the uploaded documents storaged by your entity? which is the privacy policy related to this matter? thanks @lilac turret

[lilac turret]

We use AWS S3 for our hosting, all data is encrypted and is highly secured. Access is only granted if you have the appropriate permissions e.g. member within the team you created

Feel free to reach out if you have any questions

[willow fractal]

So the way GPT works, it can only handle a certain # of tokens. So if you call for information that requires the bot to retrieve data from multiple files, it will eventually hit the # token threshold and then only give you a half baked answer because it won’t be able to use all of the information as part of the context, right? @lilac turret

[willow fractal]

So as I understand it, this thing can only retrieve data. It can't actually do anything with it

[lilac turret]

Our algorithm will only include relevant sources and disregard irrelevant ones. We utilize multiple methods to fit more into the context window. We are working on a V3 of the knowledge retrieval system that works in a different way. So this answer is about to be outdated.

[willow fractal]

Yea I tried Visus. I imported a coding project, and asked it to write some code for me based on those files, and it wasn’t able to do it accurately - it didn’t even take into account some of the files.

It may be good for data retrieval since you’re using vector databases to allow the context window include all of the data, but it isn’t good at performing actions on the retrieved data

So to be honest, I see this product as more of like a simple tool that retrieves files for you and just plugs them into your prompt so you don’t have to, saving you a few minutes of time

You can already do this with regular ChatGPT, where you just copy and paste the data into the prompt and tell it to do something.
-prompt here-
-data here-

I can see a good use case for this though - when the user doesn’t know what data to exactly send. This product would do the thinking for the user and only retrieve the files needed. But that’s the best case

If you have any actual other use cases, please let me know - I’d be interested

All of that is about the developer scope of things

I can see this being used for websites where people want to ask questions about a product or something, and the bot can answer using the data

Even good for law documents and lengthy stuff like that I guess - data retrieval all around

[civic tendon]

I tried uploading my journal and asking questions and it didn’t impress to be honest. Cool tool, but I wouldn’t trust it as a credible research tool. I have it 138 pages worth of information on my life and it couldn’t answer who my best friend was clearly

[willow fractal]

Yep probably because it’s vector similarity algorithms aren’t good yet

[lilac turret]

Both of your use cases are not intended, especially the programming part, either way thanks for the feedback

---

April 21 update

Core

• Released the ability to embed any Visus AI on your website.
• AI Description fields can now be used to steer how an AI responds to queries.

Qol

• Improved the "strict" mode setting to decrease hallucination risk.
• Added the ability to use an avatar image for an AI.
• Improved file management setting when dealing with a large amount of files & folders.

[errant agate]

Hi! I was wondering if the Chat-GTP trained by me could become a paid web service, for example, or even a public, chef-specific recipe finder?

[willow fractal]

Well @civic tendon’s use case was definitely intended. It’s one of the things you list as a ‘feature’ on your website

[lilac turret]

How you use it is entirely up to you, once you embedd an AI on your website you can make it only accessible to paying users.

[lilac turret]

Being used as a research tool is not the intention, for that, I strongly suggest using Elicit.

Now please, if you don't have anything that is on-topic, check out the other projects in this section. It's becoming a mess.

[errant agate]

Thanks! Sorry for the questions, my friend, I am very interested in your product. With Visus it is feasible, for example, to upload a database created in Notion and ChatGTP responds according to the information in that documentation, right?