Retrieval-Augmented Knowledge Assistant

Built an assistant that answers product and support questions grounded in an internal knowledge base. Chunking is semantic and aware of doc structure; retrieval combines pgvector ANN and BM25 reranking; the LLM layer is streamed with citations and a safety rail.

Highlights

• Hybrid retrieval (dense + lexical) with learned reranker
• Citations every answer, with provenance UI
• Evals in CI against a golden set; regressions block deploys