Advanced RAG

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

Purpose

Add query transformation, multi-query retrieval, reranking, and contextual assembly only where baseline error analysis justifies them.

Why

Advanced stages can improve hard queries but also distort intent, multiply cost, and hide authorization or provenance failures.

How

Classify baseline failures; add one versioned stage at a time and run ablations. Preserve the original query and authenticated scope through rewriting. Generate bounded variants, retrieve each under identical ACLs, fuse/deduplicate, rerank authorized candidates, and enforce evidence/citation checks. Cap expansion, latency, tokens, and retries; fall back to baseline or abstain on stage failure.

Tradeoffs

Reranking and expansion improve selected query classes at higher latency and provider exposure. Use smaller deterministic stages where they meet the gate.

Anti-patterns

• Letting a rewrite change tenant, time range, or requested action.
• Adding components without ablation results.
• Treating reranker score as factual support.

Enterprise Considerations

Audit every transformation, review provider data handling for each stage, and monitor segmented regressions and denial-of-wallet attacks.

Checklist

• Each stage maps to measured baseline failures.
• Original intent and ACL scope are immutable.
• Ablation proves incremental quality within budgets.
• Failure fallback and citation verification pass.

References

• Precise Zero-Shot Dense Retrieval without Relevance Labels (HyDE) (research technique; evaluate locally)
• BEIR benchmark

Changelog

• 1.0.0 — 2026-07-16: Initial production standard.