Docs/06 memory knowledge/vector databases/retrieval optimization

Retrieval Optimization and Mode Selection

Version: 1.0.0 | Last updated: 2026-07-16

Purpose

Select vector, keyword, graph, and fused retrieval by evaluated query class.

Why

No retrieval mode dominates all workloads. Exact identifiers favor lexical search, paraphrases favor dense retrieval, and explicit multi-hop relationships may favor graphs.

How

Mode Select for Primary failure
Keyword/BM25 identifiers, names, quoted text, rare terms vocabulary mismatch
Dense vector semantic similarity, paraphrase exactness, freshness, opaque neighbors
Graph explicit entity relations and bounded multi-hop ontology/edge quality and cost
Hybrid fusion mixed query classes with measured complementary errors latency and tuning complexity

Route only with a deterministic classifier or validated query features. Apply ACL/time filters inside each retriever, normalize/fuse scores with a versioned method, rerank authorized candidates, deduplicate by source, and tune top-k/reranking using segmented evaluation. Fall back or abstain when confidence and evidence are insufficient.

Tradeoffs

Fusion can improve robustness but costs more. Add each retriever only when its incremental quality exceeds its latency and operational budget.

Anti-patterns

  • “Hybrid by default” without ablation evidence.
  • Comparing incomparable raw scores across retrievers.
  • Optimizing average recall while exact-ID or ACL cases fail.

Enterprise Considerations

Budget per tenant and query class, log routing/fusion versions, and monitor drift, cost, and unauthorized-candidate counts.

Checklist

  • Query taxonomy and per-mode baselines exist.
  • ACL and temporal filters precede scoring.
  • Fusion/reranking is versioned and ablated.
  • Quality, latency, cost, and abstention gates pass by segment.

References

Changelog

  • 1.0.0 — 2026-07-16: Initial retrieval-selection standard.