retrieval-judge

Recall search returns many candidates, but not all are relevant. After recall_search, apply evaluation criteria to filter results—checking title match, content relevance, and task fit—then log decisions via flight_recorder_log to keep only directly addressing results and maintain an audit trail.

Knowledge retrieval often surfaces irrelevant results. Apply a standardized evaluation checklist to each recall_search result, log judgments for traceability, and discard misaligned items. This critical filtering step ensures downstream processes work with high-confidence data.

Relevance judgments are auditable decisions. This Skill enforces standardized evaluation against a checklist and logs each judgment via flight_recorder_log, creating a reproducible trail for review and verification of which results were kept and why.

Yes. The evaluation checklist—title/type match, content relevance, task applicability—generalizes across domains. After recall_search in any knowledge-assembly workflow, apply the same standardized judgment step to filter noise consistently.

Recall search prioritizes coverage over precision, returning both relevant and tangential matches. Filtering by relevance reduces noise, ensures only directly addressing results feed downstream tasks, and prevents low-confidence data from propagating through your workflow.

The Skill applies three core checks: does the result's title or type match the query, is the content substantively relevant, and does it address the current task. These criteria filter out false positives and keep only results that directly serve the user's need.