Retrieval Evals
Test Retrieval With The Queries Users Really Asked
A Query Wind Tunnel — real user queries launch left through four replaceable modules toward a target on the right. Misses mark which stage drifted. Compare Production vs Candidate A vs Candidate B trajectories on the same batch.
Launch
Query RewriteChanged the original meaning
RetrieverDidn't find the correct documents
RerankerLowered key results
Answer GeneratorCited content that doesn't support the conclusion
TargetCorrect doc + answer
Tunnel idle — Build Eval Set from real failures, then Run Retrieval Test.
- Build Eval Set — from real user queries
- Run Retrieval Test — score current stack
- Add A Candidate — model, index, or config
- Inspect A Miss — module that drifted
- Promote Winning Version — mark recommended
Failed Queries Are Product Feedback.
The Intelligence Hidden In Accumulated Queries.
QuerySilt is the Search Intent Diagnostic Layer for B2B SaaS, AI support, developer tools, and enterprise search — not a new search engine. Mine zero-result, rewrite, and retrieval failure silt into decisions.