Fix Poor Retrieval Accuracy in Our RAG-Based AI Agent (FastAPI + OpenAI)

We have an AI agent in production that answers questions over our company documents using RAG, but the retrieval quality has gotten unreliable and we need it fixed. The main problems: 1. It misses answers that are clearly in the docs the right information exists, but the agent says it doesn't know or returns something unrelated. 2. It sometimes hallucinates answers confidently with information that isn't in our knowledge base. 3. Citations are off when it does cite a source, it often points to the wrong document or chunk. The stack is FastAPI + OpenAI + a vector database (pgvector / Pinecone). The pipeline works end-to-end this is a quality/accuracy problem, not a "build from scratch" job. Scope: - Diagnose why relevant chunks aren't being retrieved (embeddings, chunking, or search config) - Improve retrieval accuracy review chunking strategy, add hybrid search / reranking if needed - Tighten the prompt + grounding so the agent stops answering outside the retrieved context (reduce hallucinations) - Fix citation/source mapping so answers point to the correct document - Show before/after on a few real test questions Deliverables: - Retrieval accuracy noticeably improved on our test questions - Hallucinations reduced / properly grounded answers - A short explanation of what was wrong and what you changed You're a good fit if you have: - Proven production RAG experience (not just tutorials) - Strong grasp of embeddings, chunking, vector search, reranking, and hybrid search - Worked with pgvector / Pinecone, OpenAI, and FastAPI - Experience reducing hallucinations and grounding LLM answers