What problem does it solve? A RAG query returns the wrong file at top-1, uniformly low scores, or empty results, and you need a systematic way to find the root cause instead of guessing at weight tweaks or re-indexing. ## Core Features & Use Cases - Hypothesis-ordered checklist: Walks chunk presence, score gaps, weight policy, multilingual expansion, and server asymmetry in a fixed order, stopping at the first concrete finding. - Concrete probe commands: Provides ready-to-run Qdrant scroll queries, probe_weights.py invocations, and compare_queries.py parity audits for both Python and .NET retrieval servers. - Guardrails against dead ends: Documents weight bounds (0.80-1.30), when --force-full re-indexing is actually required, and which fixes (embedder swaps, blind re-ingestion) to avoid. - Use Case: A user asks why the query "payment refund policy" returns an ADR instead of the payments lifecycle spec; the skill guides checking corpus presence, comparing raw versus post-weight scores, and confirming whether an over-boosted weight caused the mismatch. ## Quick Start Debug why the RAG query "how are inventory reservations released" returns the wrong document instead of the inventory spec.
