RAG-powered support assistant for a B2B helpdesk product

We run Deskhive, a helpdesk platform used by around 400 small SaaS companies, and we want to ship an AI assistant that drafts replies for support agents using each customer's own knowledge base. The retrieval side is the hard part: docs live in Postgres, they update constantly, and answers must cite their sources so agents can verify before sending. Scope: design the chunking and embedding pipeline, set up pgvector on our existing PostgreSQL cluster, build the retrieval + generation service in Python (FastAPI), and expose it through a clean internal API our Node backend can call. Deliverables include an evaluation harness with at least 50 golden Q&A pairs per tenant so we can measure answer quality before rollout. You would work directly with our CTO and one backend engineer. We do a 30-minute sync twice a week and everything else async in Linear and Slack. Please include a short note about a retrieval system you have actually shipped, not just prototyped.