← Lab
BUILDINGDraftStarted 2025-03
Grounded retrieval for runbooks
Can retrieval over internal runbooks answer on-call questions without inventing steps that do not exist?
- Domains
- AI
- Reliability
- Technologies
- pgvector
- PostgreSQL
- TypeScript
- Embeddings
Hypothesis
Retrieval quality, not model size, is the limiting factor for useful operational answers.
Objective
Answer on-call questions strictly from indexed runbooks, with citations and an explicit no-answer state.
Architecture
Documents are chunked by heading, embedded and stored with their source anchor. Answers must cite at least one retrieved chunk or refuse.
Implementation
Placeholder — implementation notes will be published as the experiment progresses.
Findings
Placeholder — findings pending.
Next steps
Build an evaluation set that specifically tests refusal behaviour.