Quantic Rag Project

What

RAG chatbot over a fictional company’s HR policy PDFs, with citation-grounded answers. FastAPI + LangChain + Pinecone, single Render service serving both the API and a built-in dark-themed chat UI.

Screenshot demonstrating the frontend view of the rag chatbot in use

Why

Quantic capstone-style build: a production-shaped RAG pipeline end to end — ingestion, retrieval, grounded generation, and measured evaluation rather than vibes.

How

Ingestion CLI chunks policy PDFs with RecursiveCharacterTextSplitter using separators tuned for policy docs (\n•, \n-), embeds with text-embedding-3-small, then wipes + upserts the Pinecone namespace so re-runs are idempotent.
LCEL chain: retriever (top_k=12) → numbered-citation context builder → gpt-5.4-mini → only citations actually referenced as [n] in the answer are returned; an insufficient-context fallback is enforced in prompt and post-checked in code.

Status / Learnings

Feature-complete; the live URL currently 503s (Render free-tier spin-down). A 20-question eval drove groundedness/citation accuracy from 80% → 90% after shrinking chunks and tightening prompts — and similarity-score filtering turned out counterproductive (min_score=0.0 won). ai-tooling.md frankly documents the AI-assisted dev process.