Overview
A RAG system that chunks legal documents, stores embeddings in PostgreSQL (pgvector), retrieves relevant clauses via cosine similarity, and generates cited LLM answers grounded only in uploaded contracts.
- Role
- AI / backend engineer
- Team
- Solo developer
- Timeline
- Featured work
- My contribution
- Built the RAG pipeline from chunking and embeddings through cited LLM answers.
Problem
Long contracts make clause lookup slow, and generic LLM answers invent details that are not in the document.
- Accurate clause retrieval across long legal documents
- Preventing hallucinated answers outside source contracts
- Efficient embedding and search at document scale
Product
Upload an agreement, ask a question, and get an answer that points back to the clauses it used.
- Legal document chunking and embedding pipeline
- pgvector storage with cosine similarity retrieval
- Cited LLM answers grounded in uploaded contracts
- Contract Q&A interface for legal agreement review
Approach
Retrieval quality came first. The model only answers from retrieved context, and every response expects citations.
- Structured chunking with pgvector cosine similarity search
- Citation-required LLM responses limited to retrieved context
- FastAPI backend for scalable document ingestion and querying
What changed
Contract review becomes a grounded Q&A loop instead of scrolling for the right clause or trusting an uncited answer.
Stack
Python, FastAPI, PostgreSQL, pgvector, and LLM APIs.
More projects
Explore other work in the portfolio







