LawGPT

• 🔗This project showcases Retrieval Augmented Generation on legal documents with multiple LangChain.js chains and the Vercel AI SDK in a Next.js project.
• The chain works in two steps:
  • 1️⃣First, it rephrases the input question into a "standalone" question, dereferencing pronouns based on the chat history and trimming unnecessary words.
  • 2️⃣Then, it queries the retriever (Supabase vector store) for documents similar to the dereferenced question and composes an answer.
• 🎨The main frontend logic is found in app/page.tsx.
• 🧠The api logic can be found in app/api/chat/retrieval/route.tsx.

• ⚠️ The first invocation might take up to a minute due to a cold start on the HuggingFace API. If it takes too long, please refresh the page. Subsequent prompts will work quickly!

• 👇Try asking e.g. What is the punishment for robbery? below!