Knowledge retrieval diagnostics
Trace missing bundles, useful positions, repeated document chunks, rewrites and packed context.
Measure retrieval of a multi-document answer bundle
A single relevant hit hides a missing independent answer component.
Locate the first usable procedure in a ranked list
Correct material present far below a noisy list is hidden by hit/no-hit reporting.
Recover result diversity when chunks repeat one document
Six high-scoring chunks can consume a list while representing only one source.
Test whether a search rewrite retains the user’s contrast
A rewrite drops which two procedures the user asked to compare.
Find evidence evicted by a context packing rule
Retrieval succeeded but a declared first-fit context pack omitted a required passage.
Separate retrieval misses from answer-generation misses
A final wrong answer receives the same remediation regardless of stage evidence.