OpenAI GPT 3.5 AI assistant with Langchain + Pinecone #1

I'm building an AI assistant that will actually be useful. Here's the outline of the idea for this AI assistant. The idea is to rely heavily on the Langchain library and use tools and services like OpenAI and Pinecone (OP stack).
We're currently focusing on the AI assistant's ArXiv "researcher" component. The idea here is to have a literal research assistant to help me research ideas and the latest in LLMs, AI, NLP, vector search, etc
🎙️ AI Dev Studio:
aurelio.ai
👾 Discord:
/ discord
🎉 Subscribe for Article and Video Updates!
/ subscribe
/ membership
00:00 Building an AI assistant with GPT 3.5
01:43 AI assistant overview (langchain etc)
03:44 Asking some questions on arXiv
11:03 Rest of the idea (understand YT, articles)
13:30 What will the AI assistant look like?
16:10 Next video: scraping arxiv
#openai #gpt3 #artificialintelligence #nlp #chatbot #langchain

[@AliaksandrLazerka]

Hey James! This is my first comment ever on KZbin for the past 15 years. You're making awesome content, thank you. Especially nice to see really up-to-date content. Will keep track of your videos. Also running a project development with a kind of similar stack in parallel, very promising.

[@jamesbriggs]

hey that's really awesome, really cool to hear - projects in this space are a lot of fun and I've been seeing incredible resuts, it's very cool

[@rinzler_d_vicky]

I am actually excited for this! I wanted something like this which can be given a specific domain and then the AI system is ready with all the relevant references.

[@jamesbriggs]

yeah me too! I'm looking forward to seeing how it turns out

[@zskater1234]

No way, I’ve added that to my list of ideas for this year haha. I had the same idea of having an AI assistant for research papers! Seeing you kicking off this project makes me wanna start my own version as well!

[@jamesbriggs]

haha go for it!

[@siddtem469]

@@jamesbriggs Honestly, same, and maybe I might even go for it (Slightly different idea from research papers actually)! I'm in love with your videos right now and I'd love to talk about things like this with you. I've got exams currently but I can't get my head to focus on those with what's going on in this massive world of AI.

[@filipewnunes]

I think we could make a little club than. I'm in the same path. Good luck guys!

[@Sciencehub-oq5go]

I really fancy the idea of doing backward tracking based on the references of papers. That would be super useful. I really totally like this project. It is so cool. And I especially appreciate you are sharing this. Thanks so much! Very much looking forward to this journey! And finally using it.

[@omnipedia-tech]

I've been engaged in a similar project, creating a named assistant to help me as a teacher. I don't normally do any coding in this space, and I've been contemplating how best to do more sophisticated long-term memory. This is the first I've heard of Pinecone, so it looks like I have more learning to do. This is really instructive! Thanks!

[@jamesbriggs]

sounds awesome, yeah I think you can get a lot of benefit from the memory component, maybe this will help you kzbin.info/www/bejne/qKOkdJuml9ysnac

[@lutune]

Aye! This is right up the alley of some of the products your discord channel has been working on. Great series of videos! (you broke out of the 1 a week?)

[@jamesbriggs]

with the langchain series + still covering over topics, I think 2 / week can work for a little while, there's a lot of cool stuff to cover right now

[@digidim]

Amazing work. Super excited to see how it turns out!

[@jamesbriggs]

thanks me too haha!

[@temiwale88]

James man! You're a blessing and inspiring us to share freely! God bless you man!

[@adumont]

I'm hooked! This is the only channel I have activated the notification bell 🔔!

[@jamesbriggs]

you're both epic, thanks!

[@ylazerson]

you are awesome - I'm looking forward to the next steps

[@edmundkudzayi7571]

I reckon you need no more than 30 masterfully crafted prompts to build a crude yet highly useful system. The rest is glueing things together to allow you to interact with it elegantly, with voice of course.

[@jamesbriggs]

yeah a few good prompt templates + voice would be cool

[@jacobsmall271]

Thanks so much for these videos!

[@klammer75]

I’m here for the duration! Lead the way James!🥳🤩🤓

[@jamesbriggs]

thanks man, looking forward to showing more!

[@kevon217]

You practically are my AI assistant… Good luck with this project!

[@henryciao]

Hey james ive been loving your videos with a focus on Langchain in particular and thry have helped me work on my first AI projects. However one thing that i seem to not be getting is where all this "long term memory"/"context" enters the equation. Is all of it part of the prompt at every query? Doesn't that make the model incredibly expensive since it increases the amount of Tokens passed through? I am sure i am missing something here, but just to add some context could you let us know what your target "cost" for this project is? How much should the AI assistant be costing you?

[@vaibhavdeokar5522]

Love your content❤️

[@CristiVladZ]

Look forward to follow along

[@jamesbriggs]

thanks!

[@alivecoding4995]

A very cool project!

[@RustuYucel]

Wow! Looking for a similar project but too noob to make from scratch. A multiple documents analyser and Q&A tool would be a perfect tutorial

[@jamesbriggs]

yeah it's coming soon :)

[@inchane4933]

this looks super cool James, any update of the status of this project? Cheers!

[@kobekapoor]

I’m so keen to see if you figured out how to get the bot to not respond to things it does not know “for a fact”

[@scarlett_j]

What I do, basically, is just feed it back its enhanced prompt and ask it to interpolate on that data, with a variety of reserved questions. E.g.: Prompt 1: niches in business. -- Return (1): "blah" Prompt 2: what are a variety of examples of the following: -- Return (2): "blah blah" Prompt 3: please summarize the main goals and then explain about benefits of the following: -- Return (3) "blah blah blah" Prompt 4: please write an introduction into the subject of getting into the following, then give possible varieties: : in relation to: -- Return (4) Prompt 5: please summarize and write paper about key topics relating to: + " and " + : and please write a lengthy research paper, including anything important. # *** *** *** Then all work that needs to be done is writing the variables, catching returns, feeding the old returns or prompts into the new prompt, asking it to summarize data, then tell it to write an introduction, a script, or any types of research. As well as you can do both. Don't forget to make it also summarize the key things of data so you have a new prompt. Then you could e.g.: ask it to write any monetary, any beneficial properties, detailed description, or paper for instance. Also

[@chrismaley6676]

Hi James, an exciting project. I'm assuming the project will support multiple tasks. How are you planning on mapping the user's query to the langchain chain or agent? Will you use pinecone metadata?

[@jamesbriggs]

I think a "task identifier" prompt to begin with, then use the output of that to decide what we're doing

[@chrismaley6676]

@@jamesbriggs, lol I'm always learning something new from you thanks.. I would love to participate in the project in any way possible. I could help with the UI. Disclaimer, I'm not a frontend developer but I can get a UI working for testing and demo purposes.

[@mustafadut8430]

Hi James, In my opinion, different languages have different grammatical structures, vocabularies, grammatical rules and differences in meaning. Therefore, situations may arise where the exact equivalent of a word or expression in one language does not carry the same meaning in another language. Everything is made of weights, not atoms :)

[@daharius2]

Very interesting project! Now that a ChatGPT endpoint has been released, I hope it will be a lot easier to get the chat bot functionality implemented.

[@jamesbriggs]

yeah definitely will be interesting

[@deanshalem]

Hey, so is this video 1? How can we se the notebook in full?

[@li-pingho1441]

Your content is awesome 🎉🎉🎉💪💪💪💪

[@jamesbriggs]

thanks I appreciate it!

[@Unknown-ki8yk]

I think this kind of research assistant can help me to finish my PhD earlier :)

[@chrisalmighty]

I'm an chartered accountant and I'm building exactly a similar thing for accounting research and ability for an assistant that can help me Research accounting standards and guidance or reason through problems and write memos.

[@WereOutOfBuziness]

James huge fan and thank you for the content! Question, how can we incorporate feedback to our models that leverage semantic search + prompt engineering (GQA)? This way end users to apps can help assist the semantic search for questions in the future

[@jamesbriggs]

you can do a few things, (1) add some explicit feedback from the user (like a thumbs up/down button) and using this you can create a dataset to fine-tune the LLM and retriever model, or (2) infer implicitly whether the answers are good or not based on whether the user is clicking on links (if there are any), trying a similar prompt again (so trying to reformalize the query), or other ways you can think of

[@WereOutOfBuziness]

@@jamesbriggs interesting! I like solution 1 (the explicit feedback button)! Let’s say a Q&A app leverages baseline embedding and completion models (e.g. text-embedding-ada-002, text-davinci-003) - you’re recommending to fine-tune the embedding model and completion model? Is there a simpler approach to incorporating the feedback and still leveraging the base models? I’m concerned that fine tuning a model will result in less performant models than the baselines offered by OpenAI. If fine tuning is the only optimal approach does that mean fine tuning the completion model only? Thanks again for your input :)

[@CarstenWinsnes]

Love where this is going... if you need help with that Web UI, feel free to reach out. Thanks for all the great content here!

[@jamesbriggs]

much appreciated!

[@MrAusdrifter]

Obsidian life Had a bit of learning curve but I find it a nice note system, atleast where you have direct access to the files

[@jamesbriggs]

Yeah I do like it, great software!

[@nathancanbereached]

I can't tell by reading the description, does your course go over indexes, pinecone, and langchain? I feel like that by itself is worth $15

[@jamesbriggs]

no the course is from a couple of years ago, focuses more on huggingface, haystack, and transformer models in general All the pinecone, llms, and langchain content will be shared for free here :)

[@GrigoriyMa]

hello, please tell me what is the name of the application for notes that you use? Thanks

[@jamesbriggs]

Hi, it’s Obsidian

[@glennbrien6677]

Looking forward to this project playing out. I have a similar but slightly different need. The data I’d like a system to read is a set of interview transcripts. We do interviews with a company’s employees about the organisational culture. It would be fantastic to be able to reason about the culture. Eg: Q: what organisational values have been commonly expressed amongst the respondents? Any thoughts you have for such an architecture most welcome?

[@TheAnubhav27]

Why not use Chroma DB instead of Pinecone?

[@jamesbriggs]

From what I’ve heard it doesn’t seem to have any pros over other (open source) options and being built in python cannot compete on performance with equivalent rust based DBs, but if you have any experience with it please share - as for why I use Pinecone, it’s the easiest to setup, has the most useful features (single stage filtering, native Splade support, etc), is managed, includes a 5M vector free tier, and I work there 😁 - nonetheless I may cover more options in the future, I’ve already talked about faiss and elasticsearch a lot

[@AngusLou]

Is the part#2 of AI assistant released?

[@jamesbriggs]

coming later today

[@nikhilgjog]

Thanks! could you share your notebook?

[@jamesbriggs]

just added everything, for now can find it in here: github.com/aurelio-labs/arxiv-bot These are just demo notebooks and a scraper script, don't expect anything pretty 😅

[@RatafakRatafak]

Hi, the repo you shared does not contain the jupyter notebook. Can you please upload it there? Thanks!

[@li-pingho1441]

🙏🙏🙏

[@saschakorth]

Maybe the Zebras are stripey because of a missing temperature setting? I'm just guessing.

[@jamesbriggs]

I wish that was the problem haha! Temp is 0.0, I set it early in the notebook (I don't think it was in the video), will share the notebook soon

[@RobertHoffmann]

what editor is that in the video ?

[@jamesbriggs]

VS code with chandrian owl theme and jupyter extension

[@RobertHoffmann]

@@jamesbriggs Thanks 🙏

[@UserName________]

This video shows level 2 - but level 1 is easier, just take GPT Playground chat template and you are set in an hour.