Skip to content
Leon Achata
All work
RAG

DocuChat AI

Ask questions about any PDF and get grounded answers from a hybrid search index.

The problem

Teams have answers buried in long PDFs — manuals, policies, reports — and keyword search keeps missing them.

How it works

Documents are chunked and embedded with bge-m3, stored in PostgreSQL with pgvector, and retrieved with hybrid search that combines vector similarity and full-text ranking. LangGraph orchestrates retrieval and generation, and answers come from LLaMA 3.1 on Groq for sub-second responses.

Pipeline

  1. 01Upload
  2. 02Chunk & embed
  3. 03Hybrid retrieval
  4. 04Generate
  5. 05Answer

Highlights

  • Hybrid search: vector and full-text in the same database
  • Embedding model loaded once, with GPU → CPU fallback
  • FastAPI backend with a Streamlit chat interface
  • Free-tier inference keeps running costs close to zero

Stack

LangGraphbge-m3pgvectorGroq · LLaMA 3.1FastAPIStreamlit

Want something like this for your company?

I build production versions of these systems on your data and your infrastructure.

Related service: Knowledge assistants

Book a call