Query Rewrite

RAG retrieval pre-processing layer. Detects whether a query needs rewriting, executes rewriting when needed, then searches with both original and rewritten queries. Supports 6 rewrite modes: context extraction, comparison disambiguation, coreference resolution, multi-intent decomposition, rhetorical question detection, and condition extraction. Use when: Before RAG retrieval, memory_search, wiki_search calls, or vector search. NOT for: Code generation, file operations, non-retrieval scenarios.

Query Rewrite — RAG Retrieval Pre-processing Layer

Purpose: In multi-turn conversation scenarios, rewrite fuzzy/anaphoric/multi-intent queries into structured search terms to improve RAG recall.

Core Capabilities:

• ✅ Fast detection: determine if a query needs rewriting
• ✅ 6 rewrite modes: auto-identify and execute
• ✅ Dual-path search: original + rewritten queries
• ✅ Deduplication and result merging

---

🧠 Rewrite Checklist

Before calling any retrieval tool (memory_search, wiki_search) or vector search, check the current query against this list:

Check Item	Trigger Condition	Action
🔴 Pronouns	Contains it/he/she/this/that/these/those	Rewrite → Mode 3
🔴 Implicit pronouns	"all/everything" referencing multiple entities	Rewrite → Mode 3
🔴 Comparison	"which is better/difference/compare"	Rewrite → Mode 2
🟡 Incomplete semantics	Question can't be understood without context	Rewrite → Mode 1
🟡 Multiple questions	≥2 independent questions	Rewrite → Mode 4
🟡 Rhetorical questions	"surely not/could it be/impossible"	Rewrite → Mode 5
🟢 Filter conditions	Numeric range/attribute filter/exclusions	Rewrite → Mode 6
⚪ None of above	Query is independent, complete, and clear	Search directly, no rewrite
⚪ Self-check	Rewrite result has no substantive difference from original	Don't rewrite (prevent over-rewriting)

📋 6 Rewrite Modes

Full templates and examples: references/rewrite-patterns.md

#	Mode	Trigger	Action	Example
1	Context Extraction	Query lacks subject/entity	Extract entities from last 2-3 turns	"How long is warranty?" → "iPhone 15 Pro warranty period"
2	Comparison Disambiguation	Comparison words + ≥2 candidates	Expand to A vs B format	"Which is faster?" → "PostgreSQL vs MongoDB performance"
3	Coreference Resolution	Pronouns/fuzzy references	Replace pronouns; "all"→split into independent queries (max 5)	"Do they all support it?" → ["A supports X", "B supports X", "C supports X"]
4	Multi-intent Decomposition	≥2 independent questions	Split into sub-queries with shared subject	"Color? Size? Price?" → 3 independent queries
5	Rhetorical Question Detection	Rhetorical question patterns	Extract real information needs, convert to positive query	"It won't take a month too, right?" → "Estimated delivery time"
6	Condition Extraction	Filter/limit conditions	Reorganize as subject+condition1+condition2+…	"Under $500 suitable for women" → "gift budget 500 women"

Multi-mode Priority

When a query triggers multiple modes, execute in this order:

text

Rhetorical(5) → Coreference(3) → Multi-intent(4) → Context(1) → Comparison(2) → Condition(6)

🔄 Execution Flow

text

User Query
    │
    ▼
┌─────────────────┐
│ Check checklist  │ ← compare item by item, complete within 0.5s
└───────┬─────────┘
        │
   ┌────┴────┐
   │ Rewrite? │
   └────┬────┘
        │
   ┌────┴────────────────┐
   │ YES                 │ NO
   ▼                     ▼
┌──────────┐      ┌──────────┐
│ Execute   │      │ Direct   │
│ rewrite   │      │ search   │
│ output    │      │ original │
│ 1-N       │      │ query    │
│ queries   │      └──────────┘
└────┬─────┘
     │
     ▼
┌────────────────────────────┐
│ Search: original + rewritten │
│ memory_search(original)     │
│ memory_search(rewrite_1)    │
│ memory_search(rewrite_2)... │
└────────────┬───────────────┘
             │
             ▼
┌────────────────────────────┐
│ Merge & deduplicate (↓score)│
│ Tag each result with source │
└────────────────────────────┘

⚠️ Key Rules

1. Zero-rewrite principle — Don't rewrite when unnecessary; search directly
2. Intent fidelity — Rewriting must not change user intent, only add context
3. Context scope — Default: extract entities from last 2-3 conversation turns
4. Rewrite enhances, not replaces — Original query always participates in search
5. Graceful fallback — If reference/intent can't be determined, search with original query, tag rewrite_skipped: true

📝 Usage Examples

Example 1: Normal rewrite flow

text

Context: User asked "What colors does iPhone 15 Pro come in?"
Current query: "How long is the warranty?"
↓ Checklist: incomplete semantics → Mode 1
↓ Rewrite: extract entity "iPhone 15 Pro" → "iPhone 15 Pro warranty period"
↓ Search: memory_search("How long is the warranty?") + memory_search("iPhone 15 Pro warranty period")
↓ Merge, deduplicate, return

Example 2: Multi-intent decomposition

text

Context: (none)
Current query: "Coating machine price? Warranty? Can it be customized?"
↓ Checklist: multiple questions → Mode 4
↓ Rewrite:
  - "coating machine price"
  - "coating machine warranty"
  - "coating machine customization support"
↓ Search: 4 times (1 original + 3 rewrites) → merge, deduplicate

Example 3: No rewrite needed

text

Context: (none)
Current query: "Enterprise CRM system architecture design document"
↓ Checklist: all no → direct search
↓ memory_search("Enterprise CRM system architecture design document")

📎 Resources

• Rewrite patterns reference: references/rewrite-patterns.md

📝 Changelog

Version	Date	Changes
v1.0.0	2026-07-23	Initial public release: 6 rewrite modes, rewrite checklist, progressive loading (meta+logic separation)