Search prompts: diagnose candidate coverage and zero results

Last updated: 4 Oct 202611 min read

tutorial

Candidate coverage asks whether retrieval produced a potentially useful, permitted item before ranking rearranged the list. An empty result can follow a missing document, stale index, analyzer mismatch, overstrict filter, or permission rule. A prompt should request a bounded diagnostic ledger: query ID, index snapshot, normalized tokens, filter chain, eligible count at each gate, and known answer availability. It must not recommend bypassing access control simply because a relevant private item exists.

Operational case

Beacon has five empty queries among 47. Two have catalog records not yet indexed, one uses an alternate part spelling, one is filtered by a category condition, and one would match only a document the contractor role cannot access. These are fictional diagnostic findings for the project packet, not five identical ranking defects. The permission case remains empty for that role; the proposed fix is a safe public answer or clearer no-result message, not exposure of the private record.

47 queries: 42 with candidates + 5 empty
empty causes: index lag 2; spelling 1; filter 1; permission 1
permission empty: do not relax access rule
owner: index | analyzer | filter | content/permission

Performance and review cost

Tracing Q queries through G gates costs O(QG) logged gate checks when counts are available. Replaying full retrieval can be more expensive because it touches the index and filter services. Gate counts narrow the investigation, while missing diagnostics should stay unknown rather than be assigned a convenient cause.

Common Mistakes

Do not call every zero-result query a ranking bug.
Do not bypass a permission filter for better coverage.
Do not guess a cause without gate-level evidence.

Connected lessons

prompt engineering

search relevance