Improve your BabyAGI with LangChain

[@coolmcdude]

I would really like a video about the vector store and embeddings. Preferably with llama 😬

[@disarmyouwitha]

ChatGPT did the same thing when they first implemented it in LangChain- they had to add chat versions for their different agents that parsed 3.5 properly like “chat-zero-shot-react-description”, so maybe we could implement something similar for llama models, if they react consistently enough?

[@gianfrancotasteri3733]

Perhaps, but also it would be good to generate a bunch of conversational examples using gpt4 and then fine tune a llama version to do that. If I had the resources I would certanly go that path.

[@KolTregaskes]

Another great video. I would be very interested to see you expand on this. A non OpenAI model would be good just for testing and fine tune our prompt before having to spend money on APIs with GOT3.5 or 4.

[@ElectroFriedBees]

Just to point out, there already exists a model fine tuned on tool use called toolpaca. I can envision a pipeline that had Koala responsible for part of the chain where you need accurate factual generation, the tool agent set to use toolpaca and perhaps even a more conversational model like Vicuna used for certain tasks like generating web searches or output.

[@waqaskhalidobeidy2673]

Thanks for the tutorial Sam. Very helpful indeed.

[@brofessorsbooks3352]

Would love to see a fine-tuning video!

[@rafaeldelrey9239]

Great video. For BabyAGI and AutoGPT, I m yet to see a really interesting use case where it works as expected. All my tries ended up with the agents totally losing focus or bringing completely wrong information. Inspired by your objective, I just tried "Find the most powerful laptop for AI Art in 2023, and give me the brand, model, and specs of such laptop". Tried multiple forms for this objective but got nowhere close to a good answer. I believe these autonomous tools might evolve into something interesting in the future, but the current ones are mostly loops with prompts.

[@samwitteveenai]

Totally agree they are loops and often loops that get off track too easily, but I think there are way that can make them useful for certain tasks.

[@rudy9546]

Great channel and videos , thanks!

[@jeanchindeko5477]

I’m not surprise all those ChatGPT (GPT3.5-turbo) variations of LLaMA, Alpaca, Vicuña, Dolly, Koalas,… are not able to perform as well as GPT-4. Everybody are trying to do ChatGPT alternative forgetting that Q&A is not the whole game. Beside Q&A, you might also want the model to make use of tool, self reflect, perform task prioritisation, etc…

[@impolitevegan3179]

Vectore store concept video would be great, even greater would be a deep tutorial

[@kostik]

So how is this different from Auto-GPT?

[@NNokia-jz6jb]

In short? What does this do?

[@johnwallis1626]

good vid sam, i had similar experience, if you want the process to find information beyond a simple google search process, it breaks down quickly, however, early stages so with changes to prompts espc. what sites to start scraping from it can improve

[@bandui4021]

Would it be a good idea to have a short video or something where you have a comparison of the different models and there adv/disadv which can be easily updated? The sheer amount of new things is overwhelming.

[@futurizerush]

Sam, you are the pro! This's truly insightful and helpful! I wanted to ask for your permission to translate your code into Chinese and share it with the Chinese community. I believe that they would greatly appreciate and benefit from it. Please let me know if it's cool to do that. Thanks!

[@samwitteveenai]

Can you contact me on Linkedin. I have been trying to get the videos up on Bilibili but been running into issues.

[@futurizerush]

@samwitteveenai Absolutely! I'm all ears if there's anything I can help with. Just sent you a message on LinkedIn. By the way, I'm from Taiwan and I'm not super familiar with BiliBili, but if you can shoot over some screenshots of the issues you're facing, I might be able to help troubleshoot. Hope your weekday's going well!

[@SaonCrispimVieira]

Thanks for the incredible tutorial! How would you adapt the vectorstore part in order to use chroma instead of faiss?

[@samwitteveenai]

If you check out the video I just released today on adding RAG to LangGraph, the first half of that is setting up a ChromaDB for RAG.

[@SaonCrispimVieira]

@@samwitteveenai Thank you! I'm having some problem in making the agents work dealing with non-english languages also! Any sugestion?

[@samwitteveenai]

@@SaonCrispimVieira This is usually down more to model choice. What models are you using and what languages are you trying to cover ?

[@SaonCrispimVieira]

@@samwitteveenai Just tried to adapt your tutorial on baby agi to portuguese, I translated the prompt's and the zero_shot agent stopped to work! I'm watching others videos from your awesome channel and now I think the problem is the output parser. I'm interested in developing an angent who reacts to a device's inner state change using the system manual in a rag pipeline to stream a action plan. I've seem so much progress in this area that I'm no sure if this framework is the better one or should I try other approaches? Like, go direct to langgraphs.

[@gustavomonteiro2087]

Sam, please consider prioritizing making this video about how to take Koala 13B and further fine tuning it on your on data. We are lacking videos showing how to further fine tune those open sourced models. Especially now with those autonomous applications, the OpenAI bill is getting way too expensive for everybody. 😅

[@nesdi6653]

if you can do the vectorstore embeddings with FAISS why even use pinecone

[@samwitteveenai]

Pinecone would be for if someone wants to persist it in the cloud.

[@amitpanchall81]

How to use this with SqlDatabase chain

[@samwitteveenai]

you could plugin that in as a separate tool.

[@moistweener]

Do any of these work with mobile? Anyone know

[@abhijitkadalli6435]

Won't this be even more cost effective if we used something like vicuna or alpaca? And what if used the dolly version would this be commercially viable?

[@abhijitkadalli6435]

ohhhhhhhhhhhh LESGOO YOU Actually used a model other than openAI you are the best...

[@meishadow74]

This is what im thinking when i know about langchain and pinecone, there must be a way to use other thn openai api key, for example hugging face api or mybe just use local model like vicuna/koala/gpt4all etc etc.. edited: i just finish the video and yeap, its possible to use Koala etc, but need to finetune the model to be able to do the task.. imma test it, maybe i find something or someway to get it work..

[@samwitteveenai]

I showed using Koala at the then end of the video which is very similar to Vicuna (probably a bit better) and it doesn't work without finetuning it for this task.

[@impolitevegan3179]

I've played with these a bit and it looks like raven would be much better for this task than koala.

[@samwitteveenai]

interesting I didn't try Raven on this, does it handle the tool stuff as well?

[@impolitevegan3179]

@@samwitteveenai sorry I wasn't clear, I haven't implemented a auto-gpt with raven yet. I hope the authors of raven will implement until the weekend, so I don't have to. They added chat functionality a few days ago, so it seems they are actively working on it. I meant it would be better because 1. in my experience it's better than all others including koala, dolly 2.0, alpaca... 2. it does not have the context length problem as all others, including gpt4 api. This opens new possibilities.

[@ricardocosta-w6l]

Sam, do you have a discord or other medium were we can talk?

[@samwitteveenai]

Currently not. perhaps in the future.

[@ricardocosta-w6l]

@@samwitteveenai I will be over the moon when I can discuss with you some ideas and projects. :D