What problem does it solve? Diagnosing why a hybrid search pipeline misses expected results is hard: rank changes can be measurement artifacts, stale databases, or genuine regressions. This Skill provides a verified methodology for analyzing AiRaccoon's FTS5+vec0 retrieval quality, tracing failures through the pipeline, and structuring improvements into measurable waves. ## Core Features & Use Cases - Retrieval Baseline Analysis: Read baseline reports, trace missed queries through normalization, chunking, BM25, vector ranking, and RRF fusion, then classify failures as structural, algorithmic, parametric, or measurement gaps. - A/B Comparison Methodology: Compare candidate retrieval improvements with a shared-scorer rule, pre-registered win criteria, and section-level hit metrics to avoid harness drift and test-set leakage. - Live-Bank Diagnostics: Audit a running server's memory bank with WAL-safe SQLite copies, promotion-queue SQL audits, and grading-coverage analysis. - Use Case: When a gate test fails after merging a retrieval change, use the integration rank-verification protocol to re-measure on the joined state, content-verify rank deltas against chunk contents, and distinguish same-knowledge trades from real regressions. ## Quick Start Analyze the retrieval baseline report and trace each missed query through the FTS5 and vector pipeline to classify the failure type.
