Knowledge Engine ranks your records by meaning — not keywords — using local embeddings, an in-memory cache, and a threshold-aware top-K cut, then hands only the closest matches to the AI step that follows.
Every screen that needs to shortlist its own records — competitor patents, compliance rules, prior art — hands its query text and candidates to one shared function. It never re-implements embedding or scoring logic.
An in-memory cache, keyed by each record's own content, means editing one record invalidates only its own entry — every other record's embedding stays reusable across every domain that shares the engine.
minScore + topK — keeps the downstream AI prompt focused, never overloadedIf Ollama isn't configured, unreachable, or every candidate falls below the threshold, Knowledge Engine drops to a keyword match instead of returning nothing — the downstream AI step still gets a relevant shortlist.
{ matches, degraded }Knowledge Engine talks directly to your own local Ollama instance — it has no server-side counterpart and never can, because Render's servers have no network path to your machine. The AI step that follows is separate and can run wherever you've configured it.