Problem
Users need grounded answers over long PDFs without the model inventing unsupported claims.
Project
ChatPDF AI
Retrieval-grounded document QA system using LangChain and FAISS with agentic query workflows.
Approach
I built a retrieval-first pipeline with chunking, vector indexing, and controlled context assembly before answer generation.
Result
The system was most useful when retrieval quality was treated as the main product surface instead of an invisible backend detail.
Core build
This project combines document chunking, embedding-based retrieval, and answer generation into a practical PDF assistant.
What I learned
- Chunking strategy affects answer quality more than people expect.
- Retrieval diagnostics are essential; the model cannot recover from weak context.
- Grounding checks are necessary if the app is meant to feel trustworthy.