You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Search results flag name-only graph matches — each returned triple now carries its query-overlap decomposition (name_overlap / term_overlap). Triples that matched on entity names alone are tagged name_only, listed after term-matched triples, and counted in the result's notes (when none matched on terms, the note says the graph arm returned mentions, not answers). Ranking is unchanged; the MCP search tool surfaces the tag and places name-only triples in a trailing section.
Relevance floor on by default in sentence search — units whose cosine similarity to the query falls below the floor (0.25) are withheld from sentence-form results and counted in the result's notes, so what the merge surfaces is what cleared relevance, not merely what fit the budget. Callers can set a different floor or disable it (floor_cos=None).
Added
MCP server surface — ontosmith mcp serves the built graph to AI callers over the Model Context Protocol (stdio; new mcp extra). Five read-only tools — search, entity, neighborhood, connect, triple — over the same store binding as serve. Every tool returns one envelope: source documents hoisted and referenced by short ids, claim–evidence items with citation keys, a hard token budget with signalled truncation and a resume cursor, conflicts flagged on both sides rather than resolved. No ask tool and no query language: the caller's model composes answers from the returned evidence.
Fixed
Query-time sentence index no longer peaks in memory — a build that predates the shipped sentence index gets its sentence universe built at query time; that matrix was materialized in memory (the float32 embeddings plus their float64 widening at once — about 4.7 GB for a 127k-sentence corpus), which could kill serving on a modest machine at first search. The matrix is now file-backed beside the store (identical values, evictable pages) and reused by later processes instead of being embedded again; the prepare note says which backing served.
Sentence index rebuilds or refuses on embedder dimension mismatch — the query-time cache keyed reuse on the sentence texts alone, so a matrix cached under one embedder was served under another; with different dimensions the merge died on a raw matrix-product error, unnamed and without a remedy. A cache hit is now validated against the bound embedder (a one-text dimension probe) and rebuilt on mismatch; a shipped index served under a mismatched embedder refuses by name, stating both dimensions and the remedy.