🎯 场景：把一个复合 query 拆成 3-5 个 focused 子查询，提升 RAG 检索召回率。在检索之前用，每个子查询独立检索后再融合。

Quick Use

Use when: You want to split a single complex query into focused sub-queries before retrieval. Fill in: {{original_query}} = the user's raw query; {{max_subqueries}} = a small number, typically 3 to 5. You'll get: A list of self-contained sub-queries each with the facet they target, plus a short reasoning summary. Output is JSON.

Purpose

Rewrite a single user query into a small set of focused sub-queries that together are more likely to retrieve all relevant context for the original ask. Used at query time, before vector / lexical search, to handle under-specified, multi-aspect, or compound questions. Produces structured JSON so the downstream retriever can run each sub-query independently and the orchestrator can fuse results.

Prompt

You rewrite a user query into a small set of focused sub-queries for
retrieval. Goal: maximize the chance that retrieving each sub-query
independently surfaces all the context needed to answer the original query.

Rules:
1. Emit at most {{max_subqueries}} sub-queries. Use fewer when the query
   is already focused.
2. Each sub-query must be a self-contained search query — no pronouns
   referring to other sub-queries, no "and also" conjunctions hiding
   multiple aspects.
3. Preserve named entities, numbers, and dates verbatim from the original.
4. If the original query is already well-formed and atomic, return it as
   a single sub-query unchanged.
5. Do NOT invent constraints, time ranges, or entities that are not in
   the original query.

Original query:
{{original_query}}

Return ONLY this JSON object:
{
  "subqueries": [
    {"text": "<sub-query 1>", "aspect": "<one short phrase: which facet of the original this targets>"},
    /* ... up to {{max_subqueries}} entries */
  ],
  "reasoning_summary": "<<=30 words explaining how the original was decomposed; user-visible, not internal CoT>"
}

Example

Input:

original_query: "How does Voyager 1's interstellar communication compare to the Deep Space Network's coverage of New Horizons after 2020?"
max_subqueries: 4

Expected output:

{
  "subqueries": [
    {"text": "Voyager 1 interstellar communication system specifications", "aspect": "Voyager 1 comm capabilities"},
    {"text": "Deep Space Network coverage of New Horizons after 2020", "aspect": "DSN coverage of New Horizons"},
    {"text": "Voyager 1 signal strength and data rate from interstellar space", "aspect": "Voyager 1 link quality"},
    {"text": "DSN antenna allocation policies for outer solar system probes", "aspect": "DSN allocation policy"}
  ],
  "reasoning_summary": "Split a comparative question into one sub-query per probe's communication facts plus a policy sub-query that bridges them."
}

Failure Modes

• Over-decomposition — model splits an already-atomic query into N near-duplicates, hurting retrieval precision and wasting calls. Detect by comparing token-level Jaccard between sub-queries; reject batches with mean similarity > 0.85.
• Constraint hallucination — adds a year, location, or qualifier not in the original ("Voyager 1 in 2024..." when the original said no year). Detect by checking that all named entities and numbers in sub-queries also appear in the original.
• Lost specificity — generalizes a precise query into vague aspects ("space exploration" instead of "Voyager 1 communication"). Detect by measuring whether the original's rare terms survive in at least one sub-query.
• Pronoun leakage — sub-query 2 says "its signal strength" referring back to sub-query 1, breaking independent retrieval. Mitigation: the prompt's rule 2 helps; reject any sub-query containing third-person pronouns referring to entities not named within it.

Tuning Notes

• 模型差异：弱模型（7B 级）容易出现 over-decomposition 和约束幻觉；建议 在弱模型上把 max_subqueries 限到 2–3 并增加一个 verifier 步骤过滤。
• 温度：0.0–0.3。多样性不是这里的目标，稳定的 decomposition 是。
• 与 fusion 的关系：本卡只负责出 sub-queries，结果如何 fuse（RRF / weighted / 重排）由检索层决定。如果 downstream 用 RRF，建议 sub-query 数量 3–5 较合适。
• 与 HyDE 的关系：本卡和 rag/hyde-hypothetical-answer-generator 互补 而非替代——前者把 query 拆细，后者把 query 改写成更接近文档的形式。

Changelog

• 0.1.0 — Initial card.