memoria-recall-flow

Hybrid retrieval improves RAG memory recall by combining HyDE query expansion with Meilisearch BM25 and vector search. It applies semanticRatio control and LLM reranking to produce accurate top-K results, ensuring efficient context synchronization across multi-tenant scoped memories.

The best way to implement multi-tenant scoping is by enforcing business_id and user_id filters during Meilisearch hybrid retrieval. This canonical pipeline ensures queries only access tenant-specific memory, preventing data leakage while maintaining accurate retrieval across scoped contexts.

Prevent unnecessary re-embedding during memory sync by applying negative caching and correct Meilisearch filtering. Avoiding full re-embedding reduces latency and cost while maintaining accurate append-only fact persistence with valid_until supersession semantics.

Handle faturamento metrics by exposing three distinct angles—bruto, liquido, and caixa—instead of a single value. This ContextSnapshot correctness prevents LLM confusion and wrong answers by providing clear, multi-faceted financial context within the retrieval pipeline.

Meilisearch hybrid search requires negative caching to prevent failed or empty queries from repeatedly hitting the index. This retrieval guardrail reduces unnecessary computational overhead and latency, ensuring the RAG memory recall pipeline remains efficient and cost-effective.

Yes, you can apply LLM reranking after BM25 and vector search to refine top-K results. This reranking step evaluates the initial hybrid retrieval output, prioritizing the most contextually relevant documents to improve precision and recall metrics for memory recall.