What problem does it solve? Answering questions from ingested documents requires more than a single vector search: naive top-k retrieval loses recall, dilutes generation with raw child chunks, and has no way to signal insufficient evidence. This Skill implements the full hybrid retrieval pipeline so queries against a document corpus return ranked, citation-backed parent chunks or an explicit insufficiency signal. ## Core Features & Use Cases - Hybrid first-stage recall: Runs BM25 full-text search and pgvector dense vector search in parallel, then fuses results with Reciprocal Rank Fusion (k=60). - Cross-encoder reranking and parent resolution: Reranks the top 50-100 fused candidates down to 5-10, then resolves child chunks to their 1,000-2,500 token parent chunks before generation. - Permission-safe filtering and sufficiency gating: Compiles tenant_id and permission_scope into SQL WHERE clauses, and applies a bounded corrective-retry loop (max 1) before escalating aggregation queries to the RLM path. - Use Case: A user asks a multi-part question over a 1,000-page ingested document set; the pipeline classifies the query, decomposes it, retrieves and reranks evidence, and returns cited parent chunks or an "insufficient evidence" answer instead of a guess. ## Quick Start Use the rag-retrieval skill to answer this question from the ingested documents and return ranked parent chunks with citations.
