A retrieval assistant over a manufacturer's technical documentation for blind rivet nut systems. Delivered as client work, now complete.

Production runs against a managed corpus: source documents are imported from object storage into a hosted RAG engine, and at query time the app retrieves excerpts, injects them into a Gemini call, and passes the result through a guardrail layer before it reaches the user. Nothing is indexed on the host.

An earlier local path still exists behind an explicit opt-in flag, and it is the more interesting piece of engineering: parse the manuals to markdown, split them with a table-aware splitter because the answers live in the tables, then retrieve with vector and BM25 in parallel and fuse the two rankings. It was replaced by the managed corpus for operational reasons, not because it worked badly.

Client intellectual property: architecture only. No corpus, no prompts, no configuration, no production data.