What problem does it solve? Business users and agents need to find the right data fields in a catalog using natural language queries, but pure keyword search misses semantic matches while pure vector search loses exact business terms. This Skill implements a two-phase catalog search that keeps one response contract across both phases. ## Core Features & Use Cases - Keyword MVP: ILIKE and PostgreSQL tsvector ranking over business names and definitions behind /api/catalog/search. - Hybrid RAG phase: Combines tsvector keyword scores with pgvector cosine similarity over Gemini embeddings (VECTOR(768)) with fixed 0.4/0.6 weights. - Tribal notes with applicability: Attaches contextual notes to hits only when their applicability condition matches the query, and every response cites its source. - Evaluation gate: Enforces Hit Rate@3 >= 0.8 on a 20-query business evaluation set before closing the hybrid phase. - Use Case: A user searches "saldo vencido de la cartera" and receives catalog hits citing the creditos source, with applicable tribal notes attached. ## Quick Start Ask the agent to implement the catalog search endpoint with keyword tsvector ranking first, then add the hybrid pgvector scoring and run the Hit Rate@3 evaluation.
