Hybrid RAG
Version: 1.0.0 | Last updated: 2026-07-16
Purpose
Fuse complementary retrieval modes when workload evidence shows that one mode is insufficient.
Why
Lexical and dense retrieval often fail on different queries; graph retrieval may add value for explicit relationship questions. None is universally required.
How
Establish per-mode baselines, then run complementarity and ablation analysis by query class. Execute only justified retrievers under identical pre-retrieval ACL/time filters. Fuse with a documented method such as reciprocal rank fusion, deduplicate source versions, rerank, enforce context diversity and token budget, and verify citations. Route simple query classes to one mode to control latency.
Tradeoffs
Hybrid retrieval may improve coverage but increases infrastructure, tuning, latency, and incident surface. Graph retrieval is an emerging optional branch, not part of the minimum definition.
Anti-patterns
- Claiming universal recall gains or fixed percentages.
- Sending every query to every index.
- Fusing candidates before each retriever enforces authorization.
Enterprise Considerations
Apply consistent deletion and correction across all indexes, record per-retriever provenance, and reconcile tenant isolation and residency controls.
Checklist
- Complementary errors justify each retriever.
- ACL and lifecycle controls are equivalent across indexes.
- Fusion, routing, and reranking are versioned.
- Ablation, latency, cost, and citation gates pass.
References
Changelog
- 1.0.0 — 2026-07-16: Replaced universal hybrid mandate with evidence-based fusion.