Problem
Users need grounded answers over long PDFs without the model inventing unsupported claims.
Applied AI & Agents · 202480% on manual test set
ChatPDF AI
Retrieval-first document QA with LangChain + FAISS and OpenAI Agents SDK workflows, reaching 80% accuracy on a manual test set (N ≤ 50).
Approach
I built a retrieval-first pipeline with LangChain chunking and FAISS vector search, then added OpenAI Agents SDK workflows for document querying and response generation.
Result
The Streamlit app reached 80% accuracy on a manual test set (N ≤ 50) with answers grounded in source text.
Step 1 of 7: Upload PDF
Highlights
- Built a retrieval-first document QA system using LangChain and FAISS to ground answers from source text.
- Integrated Agents SDK for agentic workflows in document querying and response generation, resulting in 80% accuracy on a manual test set (N ≤ 50). Deployed on Streamlit.
Core build
This project combines LangChain document chunking, embedding-based FAISS vector search, and OpenAI Agents SDK workflows for querying and answer generation in a Streamlit app.
What I learned
- Chunking strategy affects answer quality more than people expect.
- Retrieval quality limits answer quality; the model cannot recover from weak context.
- Retrieved passages should remain inspectable when assessing an answer.