Problem
Process engineers spend hours searching runbooks, tool logs, and historical excursion reports scattered across systems. Institutional knowledge lives in the heads of senior engineers and leaves when they do.
Solution
A retrieval-augmented assistant that indexes process documentation, tool manuals, and anonymized excursion history, and answers engineering questions with citations back to the source documents, so every answer is verifiable.
Architecture
- Document ingestion pipeline with chunking tuned for technical specs and tables
- Hybrid retrieval: dense embeddings + BM25, re-ranked before generation
- Claude API with tool use for structured queries against tool-state databases
- Citation-first response format: every claim links to its source paragraph
Outcome
- Turned hours of document archaeology into seconds of grounded Q&A
- New engineers onboard faster with a system that cites the runbook, not vibes
- Retrieval evaluation harness keeps answer quality measurable release over release
Lessons
- “RAG quality is a retrieval problem before it is a generation problem”
- “Citations are the difference between a demo and a tool engineers rely on”
- “Evaluation harnesses must exist before the first user, not after”
Stack
- TypeScript
- Python
- Claude API
- pgvector
- Next.js