🎯 场景：把 rag/query-rewriting-decomposition 拆出的多个子查询的检索结果融合成一个排序集——按对原 query 的整体覆盖排序，去重 near-duplicate，标记每条来自哪个 sub-query。是 query-decomposition 的 "fan-in" 步骤。

Quick Use

Use when: You decomposed a query into sub-queries, retrieved separately, and now need to fuse the per-sub-query result sets into one deduplicated, ranked set. Fill in: {{original_query}} = original user query; {{subquery_results}} = JSON array of (subquery, passages) pairs. You'll get: A unified ranked passage list with deduplication, source-subquery attribution, and per-passage reasoning. Output is JSON.

Purpose

Fuse retrieval results from multiple sub-queries into a single ranked, deduplicated set, ranked by relevance to the ORIGINAL query (not to each sub-query). Used as the fan-in step after rag/query-rewriting-decomposition. Distinct from rag/multi-source-aggregator: that card synthesizes an answer; this card produces the fused passage set used as input to answering.

Prompt

You fuse retrieval results from multiple sub-queries into one ranked
passage set, ranked by relevance to the ORIGINAL query.

Original query:
{{original_query}}

Sub-query results (each sub-query has its own passages):
{{subquery_results}}

Steps:
1. Walk through all passages across all sub-queries.
2. Identify near-duplicates (same chunk surfaced via different
   sub-queries; same fact stated in different passages). Group
   them; pick the most informative representative.
3. Rank the deduplicated set by relevance to the ORIGINAL query
   (not to any single sub-query). A passage that addresses two
   different sub-queries should rank higher than one that addresses
   only one.
4. For each kept passage, note which sub-queries surfaced it.
5. Drop passages that don't bear on the original query at all
   (sub-query may have over-decomposed).

Return ONLY this JSON object:
{
  "fused_passages": [
    {
      "passage_id": "<original id>",
      "rank": <integer>,
      "from_subqueries": ["<subquery text>", "..."],
      "relevance_to_original": "high" | "medium" | "low",
      "rationale": "<one short phrase: how this contributes to original_query>"
    }
  ],
  "duplicates_merged": <integer count>,
  "passages_dropped": <integer count>,
  "coverage_check": {
    "subqueries_with_no_kept_passages": ["<subquery text>"],
    "missing_aspects_of_original": ["<aspect of original_query not covered by any kept passage>"]
  },
  "decision_basis": "<one sentence, <=30 words, no internal CoT>"
}

Example

Two sub-queries about Reykjavik (population, weather), each retrieved 3 passages, some passages overlap.

Expected output (abridged):

{
  "fused_passages": [
    {"passage_id": "p1", "rank": 1, "from_subqueries": ["Reykjavik population 2024", "Reykjavik facts"], "relevance_to_original": "high", "rationale": "Covers two sub-queries — population number and general overview."},
    {"passage_id": "p3", "rank": 2, "from_subqueries": ["Reykjavik population 2024"], "relevance_to_original": "high", "rationale": "Population statistics with citation."}
  ],
  "duplicates_merged": 1,
  "passages_dropped": 1,
  "coverage_check": {
    "subqueries_with_no_kept_passages": [],
    "missing_aspects_of_original": []
  },
  "decision_basis": "Two distinct passages kept; one near-duplicate merged into rank-1; one off-topic passage dropped."
}

Failure Modes

• Reciprocal-rank-fusion blindness — model uses each sub-query's retrieval_score directly, ignoring that scores aren't comparable across sub-queries. Mitigation: rule "rank by relevance to ORIGINAL query"; verify rationale references the original query.
• Duplicate keeping — same chunk appears at rank 1 and rank 5; model didn't detect duplication. Audit by checking passage_id uniqueness in fused_passages.
• Coverage blindness — model drops passages from a sub-query entirely, missing an aspect of the original. The coverage_check field surfaces this; verify subqueries_with_no_kept_passages is empty unless the sub-query was off-topic.
• Hallucinated rationale — rationale describes content not in the passage. Sample and check.

Tuning Notes

• 模型差异：frontier 模型在 cross-subquery deduplication 上更稳； 中档模型容易把"提到同一实体"误判为 near-duplicate。
• 温度：0.0–0.2。
• 与 rag/query-rewriting-decomposition 的关系：那张卡是 fan-out （一个 query 拆成 N 个 sub-query）；本卡是 fan-in（N 个 sub-query 的结果合一）。组合是 multi-query RAG 的标准模式。
• 与 reciprocal rank fusion (RRF) 算法的关系：传统 RRF 是基于 retrieval_score 的数学融合；本卡是基于语义理解的融合。前者更便宜 快，后者更准。生产中可以叠加：先 RRF 粗排到 top-20，再用本卡精 fuse 到 top-10。
• 与 rag/multi-source-aggregator 的关系：本卡产 fused passage 集 （检索阶段产物）；那张卡产合成 answer（生成阶段产物）。pipeline: decompose → retrieve per-sub → fuse (本卡) → aggregate answer。
• coverage_check 的下游：missing_aspects_of_original 非空时触发追 加 retrieval（用未覆盖的 aspect 作新 query）。

Changelog

• 0.1.0 — Initial card.