Skip to main content
The query surface is four universal facets that map to SQL predicates, an ambient view layer, and one escape hatch. Calling agents compose flags; the CLI never parses natural language into args. Ranking modes: --mode bm25 (default — Okapi BM25 over title+text), --mode vec (local-embedding similarity), --mode hybrid (RRF fusion of both, k=60). The default only changes when the eval harness proves a mode wins. Escape hatch: --where "<raw SQL>" over the docs columns — source-specific fields stay in the VARIANT payload, reachable by SQL instead of new flags. View (ambient, set by workspace config): --instance (which brain), --class (classification ceiling), --observer (whose assertions).