memoria-recall-flow

What problem does it solve?

This Skill solves weak or inconsistent Copiloto memory retrieval by enforcing a canonical recall pipeline (negative-cache → HyDE expansion → Meilisearch hybrid → LLM reranking → append-only persistence) and by applying business-logic correctness for faturamento metrics with multi-tenant scoping.

Core Features & Use Cases

• Canonical hybrid recall pipeline: Runs NegativeCache, HyDE query expansion, Meilisearch hybrid (BM25 + vector), and Llm reranking to produce top-K results efficiently and accurately.
• ContextSnapshot correctness for faturamento: Exposes three legitimate faturamento angles (bruto, liquido, caixa) instead of a single value to prevent LLM confusion and wrong answers.
• Retrieval safety via cataloged gotchas and metrics: Avoids known failure modes (e.g., missing negative-cache, wrong semanticRatio usage, incorrect filtering, unnecessary re-embedding) and ties changes to required daily retrieval metrics (including recall@5, precision@5, latency, and cost).

Quick Start

Ask the AI: "Use hybrid recall to find what we learned about faturamento90d and return the correct bruto, liquido, and caixa angles for business_id 4, top 5, with reranking and metrics-ready context."