Deep dives

Context recall

How one call turns a question into prompt-ready memory: hybrid retrieval, entity expansion, and the weakEvidence abstention signal.

Context recall is the read primitive everything else exists for: one call, one prompt-ready block. See the API reference for parameters; this page explains the behavior.

What happens on a recall

  1. Hybrid retrieval (BM25 + vector) ranks documents for q - each hit reports which sources matched.
  2. The top entities are expanded with their current facts and recent timeline.
  3. Everything is fused into contextMd, ordered for prompt injection.

Abstention beats hallucination

If the best hit scores under 0.5, the response sets weakEvidence: true. Your agent should treat that as "the brain does not know" - caveat the answer or proceed without memory, never present thin matches as established fact. Memory layers that always return something teach agents to invent provenance.

Recall discipline for agents

  • Recall at task start, not mid-stream - one good contextMd injection beats five micro-queries.
  • Scope with pathPrefix when the task has a lane (/workspace/teams/platform/).
  • Use includeBodies only when the caller won't make follow-up doc reads (single-shot pipelines).

On this page