Karpathy-GPT

Context

This app is a template for using LangChain to build a LLM Q+A assistant from any set of YouTube videos.

We use Karpathy's course on LLMs as an example.

We use LangChain to:

(1) convert YouTube urls to text

(2) feed the text into LangChain auto-evaluator to test different chain parameters

(3) with our chosen parameters, build a vectorstore retriever back-end with FastAPI (deployed to Railway)

(4) stream the generated results (answer and retrieved docs) to a front-end (deployed to Vercel)

See the notebook in /index folder:

Uses LangChain's OpenAIWhisperParser to convert urls to text in < 10 lines of code

See the text files in /eval folder:

Feed the text from step 1 and, optionally, an eval set to the auto-evaluator app
We can use this to test different parameters (see full README in the repo here)
Use the UI to run experiments
Select your best retriever, chain settings (e.g., k, split size, split overlap, etc), LLM, embeddings

See the notebook in /index folder:

Split the text from step 1 using parameters you found in step 2
Upsert the vectors to a VectorDB (e.g., in this example, Pinecone) with metadata
See this PR / notebook if you want to use locally with a different VectorDB

See the karpathy_app.py file in /api folder:

We use LangChain's load_qa_chain with a user specified LLM and prompt (see default_prompt_template)
Given a question, this will stream answer the text back to front-end and pass the retrieved documents back
We deploy this FastAPI API to Railway
See README.md in /api for local testing instructions

See /nextjs directory for nextJS app:

This will call the back-end with the query and fetch the documents / answer
Test the app locally by launching the back-end:
```
uvicorn karpathy_app:app
```
To run front-end locally with you locally running back-end, simply change the source in fetchEventSource here and here to http://localhost:8000/karpathy-docs and http://localhost:8000/karpathy-stream
To run the front-end locally, run:
```
npm run dev
```