What
A RAG chatbot over a fictional company’s HR policy PDFs, with answers grounded in citations. It’s FastAPI + LangChain + Pinecone, deployed as a single Render service that serves both the API and a built-in dark-themed chat UI.
Why
A Quantic capstone-style build. I wanted a production-shaped RAG pipeline end to end: ingestion, retrieval, grounded generation, and a measured evaluation instead of vibes.
How
- An ingestion CLI chunks the policy PDFs with
RecursiveCharacterTextSplitter, using separators tuned for policy documents (\n•,\n-), embeds them withtext-embedding-3-small, then wipes and upserts the Pinecone namespace so re-runs are idempotent. - The LCEL chain runs a retriever (top_k=12) into a numbered-citation context builder and then
gpt-5.4-mini. Only citations actually referenced as[n]in the answer get returned, and an insufficient-context fallback is enforced in the prompt and post-checked in code.
Status / Learnings
Feature-complete, though the live URL currently 503s thanks to Render’s free-tier spin-down. A 20-question eval drove groundedness/citation accuracy from 80% to 90% after I shrank the chunks and tightened the prompts; similarity-score filtering turned out counterproductive (min_score=0.0 won). The AI-assisted dev process is documented openly in ai-tooling.md.