Find candidates
SQLite FTS5 pulls up to 30 keyword matches. Dictionary notes sort first.
Keyword search finds the candidates. A small neural model on your machine reads each one against your query and puts the best first. No API call, no key, nothing leaves the vault.
SQLite FTS5 pulls up to 30 keyword matches. Dictionary notes sort first.
A cross-encoder scores every query and passage pair together, on your CPU.
Dictionary notes get a small boost and journals a small penalty. A tie-breaker, not an override.
Bundles prune matches under the relevance floor and list them in meta.skipped with the reason.
With include_bodies, bodies are added in rank order until the token budget is full.
Query: database migration rollback. Live output from three test notes written for this page.
The guide moves above the calendar. The note that repeats the query still wins here, so every match carries its rerank_score and you can see when to distrust the order. Scores bunch near 1 when every candidate is plausible; the order is what carries the information.
Defaults shown. Pass rerank: false to files_search to skip it for a single call.
{
"search": {
"rerank": {
"enabled": true,
"candidate_limit": 30,
"dictionary_boost": 0.08,
"timeout_ms": 300
},
"verdicts": { "relevance_floor": 0.15 }
}
}The rules block injected into CLAUDE.md, GEMINI.md and the other agent files is now a short stub. The full reference lives in KONTEXTA.md and is read only when needed.
83% fewer tokens in every session's standing context. Counted with the cl100k_base tokenizer, a neutral proxy rather than Claude's own. KONTEXTA.md itself is 4,107 tokens, loaded on demand.