Chat With Your Code

Chat with your Github repositories, built with LangChain, Supabase, Next.js and Github's API. Uses gpt-3.5-turbo-0125 by default.

How It Works

1. Log in with Github to get provider token
2. Use provider token to get all public repo metadata
3. Load repo into Supabase
   • Split code into chunks, embed file contents and store with associated metadata in Supabase (pgvector)
4. Begin chat session with repo
   • Load message history from existing chat session or create new one
5. User asks a question
   • Create vector store from existing index (repo was embedded beforehand)
6. Build runnable sequence with langchain
   • Load chat history
   • Clean up original prompt through OpenAI (ie. makes the question more clear)
   • Load context by embedding prompt and running cosine similarity search
   • Run full chain and return stream
   • Callback after chain end stores AI response in chat history DB table

NOTE: This repo is a proof of concept. Security features such as proper auth, RLS, and safe API key handling are not implemented. Don't deploy to production unless you like chaos.

Tech Stack

cvega21/chat-with-your-code is built on the following main stack:

• TypeScript – Languages
• JavaScript – Languages
• SQL – Languages
• Shell – Languages
• React – Javascript UI Libraries
• Tailwind CSS – Front-End Frameworks
• Next.js – Frameworks (Full Stack)
• Supabase – Realtime Backend / API
• LangChain – Large Language Model Tools
• OpenAI – Large Language Models
• Octokit – GitHub API SDK
• Vercel AI SDK – Large Language Model Tools