fix(strategy): 探针模式下从 structured_brief 移除按细类关注词/场景子串

补充 _omit_ch8_probe_wordchart_fields：删除 consumer_feedback 内 focus_keyword_hits 与 scenarios_top；与文本挖掘主口径一致。导出脚本与 omission 说明同步；更新全量输入快照。

Made-with: Cursor

backend/pipeline/demos/dump_strategy_llm_input_md.py

@@ -0,0 +1,269 @@
+"""
+导出与「策略大模型润色」一次调用一致的完整输入，生成 Markdown（供对照）。
+
+用法（在项目 backend 目录）::
+
+  python -m pipeline.demos.dump_strategy_llm_input_md [--job-id 12] [--matrix-index 0]
+
+若未指定 --matrix-index，则默认收窄到第一个矩阵分组（与「选第一个细类」等效）。
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+# Django
+os.environ.setdefault("DJANGO_SETTINGS_MODULE", "market_assistant.settings")
+
+
+def _strategy_decisions_empty() -> dict:
+    return {
+        "product_role": "",
+        "time_horizon": "",
+        "success_criteria": "",
+        "non_goals": "",
+        "battlefield_one_line": "",
+        "positioning_choice": "",
+        "competitive_stance": "",
+        "pillar_product": "",
+        "pillar_price": "",
+        "pillar_channel": "",
+        "pillar_comm": "",
+        "audience_segment": "",
+        "competitor_reference": "",
+        "resource_notes": "",
+        "marketing_strategy": "",
+        "general_strategy": "",
+        "ack_risk_keywords": False,
+        "ack_risk_price": False,
+        "ack_risk_concentration": False,
+    }
+
+
+def main() -> int:
+    import django
+
+    django.setup()
+
+    from django.utils import timezone
+
+    from pipeline.jd.runner import build_competitor_brief_for_job
+    from pipeline.llm.generate_strategy import (
+        STRATEGY_SYSTEM,
+        STRATEGY_USER_PREFIX,
+        _omit_ch8_probe_wordchart_fields,
+        _truncate_strategy_narrative,
+    )
+    from pipeline.models import JobStatus, PipelineJob
+    from pipeline.reporting.brief_compact import compact_brief_for_llm
+    from pipeline.reporting.brief_strategy_scope import (
+        filter_brief_for_strategy_matrix_group,
+        list_matrix_groups_for_api,
+    )
+    from pipeline.reporting.report_matrix_group_evidence import (
+        load_report_matrix_group_evidence_markdown,
+    )
+    from pipeline.reporting.report_strategy_excerpt import load_report_strategy_excerpt
+    from pipeline.reporting.strategy_draft import (
+        build_strategy_draft_markdown,
+        filter_strategy_hints_for_ch8_probe,
+        report_uses_chapter8_text_mining_probe,
+    )
+
+    p = argparse.ArgumentParser()
+    p.add_argument("--job-id", type=int, default=None)
+    p.add_argument(
+        "--matrix-index",
+        type=int,
+        default=0,
+        help="矩阵分组下标；设为 -1 表示不收窄（全部分类）",
+    )
+    p.add_argument(
+        "--no-scope",
+        action="store_true",
+        help="与 --matrix-index -1 相同：不收窄 brief、不抽细类报告节选",
+    )
+    p.add_argument(
+        "-o",
+        "--output",
+        type=str,
+        default=None,
+        help="输出 .md 路径（默认：docs/planning/策略生成-LLM全量输入快照.md）",
+    )
+    args = p.parse_args()
+
+    backend_dir = Path(__file__).resolve().parents[2]
+    repo_root = backend_dir.parent
+    default_out = (
+        repo_root / "docs" / "planning" / "策略生成-LLM全量输入快照.md"
+    )
+    out_path = Path(args.output) if args.output else default_out
+
+    job_id = args.job_id
+    if job_id:
+        job = PipelineJob.objects.filter(pk=job_id).first()
+    else:
+        job = (
+            PipelineJob.objects.filter(status=JobStatus.SUCCESS)
+            .exclude(run_dir="")
+            .order_by("-id")
+            .first()
+        )
+    if not job:
+        print("无可用成功任务（需 run_dir 非空）", file=sys.stderr)
+        return 1
+
+    rc = job.report_config if isinstance(job.report_config, dict) else None
+    brief = build_competitor_brief_for_job(
+        job.run_dir,
+        job.keyword,
+        report_config=rc,
+    )
+    matrix_groups = list_matrix_groups_for_api(brief)
+    group_names = [g.get("group") for g in matrix_groups if isinstance(g, dict)]
+    scoped_label = ""
+    matrix_index: int | None = args.matrix_index
+    if args.no_scope:
+        matrix_index = -1
+    if matrix_index is not None and matrix_index >= 0:
+        mg = brief.get("matrix_by_group")
+        if isinstance(mg, list) and matrix_index < len(mg):
+            scoped_label = (mg[matrix_index].get("group") or "").strip()
+            brief = filter_brief_for_strategy_matrix_group(
+                brief, matrix_group_index=matrix_index
+            )
+        else:
+            print(f"matrix_index {matrix_index} 超出范围", file=sys.stderr)
+            return 1
+
+    gen_at = timezone.now().isoformat()
+    sd = _strategy_decisions_empty()
+    rules_md = build_strategy_draft_markdown(
+        job_id=job.id,
+        keyword=job.keyword,
+        brief=brief,
+        business_notes="",
+        generated_at_iso=gen_at,
+        strategy_decisions=sd,
+        report_config=rc,
+    )
+
+    excerpt_raw, excerpt_src = load_report_strategy_excerpt(job.run_dir)
+    excerpt_raw = (excerpt_raw or "").strip()
+
+    evidence_md = ""
+    evidence_src = "none"
+    if scoped_label:
+        evidence_md, evidence_src = load_report_matrix_group_evidence_markdown(
+            job.run_dir,
+            scoped_label,
+        )
+
+    compact_max = 80_000
+    excerpt_max = 24_000
+    compact = compact_brief_for_llm(brief, max_chars=compact_max)
+    if report_uses_chapter8_text_mining_probe(rc):
+        compact = dict(compact)
+        _omit_ch8_probe_wordchart_fields(compact)
+        if isinstance(compact.get("strategy_hints"), list):
+            compact["strategy_hints"] = filter_strategy_hints_for_ch8_probe(
+                compact["strategy_hints"]
+            )
+    ex = (
+        _truncate_strategy_narrative(excerpt_raw, excerpt_max)
+        if excerpt_raw
+        else ""
+    )
+    ev_max = min(24_000, max(3_000, excerpt_max + excerpt_max // 2))
+    gm = (
+        _truncate_strategy_narrative(evidence_md.strip(), ev_max)
+        if evidence_md
+        else ""
+    )
+
+    payload: dict = {
+        "job_id": job.id,
+        "keyword": job.keyword,
+        "generated_at_iso": gen_at,
+        "strategy_decisions": sd,
+        "business_notes": "",
+        "structured_brief": compact,
+        "rules_draft_markdown": rules_md,
+        "report_strategy_excerpt": ex,
+        "report_matrix_group_evidence_md": gm,
+        "chapter8_text_mining_probe": bool(
+            report_uses_chapter8_text_mining_probe(rc)
+        ),
+    }
+    if report_uses_chapter8_text_mining_probe(rc):
+        payload["structured_brief_omission_note"] = (
+            "已启用第八章文本挖掘（探针为主）：structured_brief 已省略顶层「关注词/场景子串计数」、按细类 feedback 中的 focus_keyword_hits/scenarios_top，"
+            "以及「与条形图同源的 strategy_hints 句子」，避免与报告 §8 主口径冲突。**不得**再以这类子串计数或预设场景占比作为论据。"
+            "用户与评论侧须依报告 §8 文本挖掘归纳及 `report_matrix_group_evidence_md`；**促销、满减、券价差**须与报告第六章、`price_promotion_signals` 及下方 `report_strategy_excerpt`（第九章）对齐，不得省略报告已写明的活动建议。"
+        )
+
+    user_body = STRATEGY_USER_PREFIX + json.dumps(payload, ensure_ascii=False)
+    full_chars = len(STRATEGY_SYSTEM) + len(user_body)
+
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+
+    lines: list[str] = [
+        "# 策略生成 · 大模型一次调用的「全量输入」快照",
+        "",
+        "> **生成方式**：本机 `pipeline.demos.dump_strategy_llm_input_md` 按与 "
+        "`generate_strategy_draft_markdown_llm` 相同的 payload 组装逻辑导出。",
+        "> **与线上一致性**：与真实接口相比，表单字段此处均为空默认；"
+        "你只要把当时提交的 `strategy_decisions` / `business_notes` 代入即与线上等价。",
+        "",
+        "## 快照元数据",
+        "",
+        f"- **任务 ID**：{job.id}",
+        f"- **关键词**：{job.keyword}",
+        f"- **run_dir**：`{job.run_dir}`",
+        f"- **矩阵分组**：{matrix_index if matrix_index is not None and matrix_index >= 0 else '未收窄（全部分类）'}{f' → 「{scoped_label}」' if scoped_label else ''}",
+        f"- **本任务可选细类（节选）**：{group_names[:20]}{'…' if len(group_names) > 20 else ''}",
+        f"- **第九章节选来源**：{excerpt_src}，约 {len(ex)} 字符",
+        f"- **细类报告节选来源**：{evidence_src}，约 {len(gm)} 字符",
+        f"- **System 字符数**：{len(STRATEGY_SYSTEM)}",
+        f"- **User 消息字符数**：{len(user_body)}",
+        f"- **合计约**：{full_chars} 字符",
+        "",
+        "---",
+        "",
+        "## 1. System 提示词（完整 `STRATEGY_SYSTEM`）",
+        "",
+        "```text",
+        STRATEGY_SYSTEM,
+        "```",
+        "",
+        "---",
+        "",
+        "## 2. User 消息（完整：`STRATEGY_USER_PREFIX` + JSON）",
+        "",
+        "以下为网关 **user** 角色一次发送的完整字符串（前缀 + 单行 JSON）。",
+        "",
+        "```text",
+        user_body,
+        "```",
+        "",
+        "---",
+        "",
+        "## 3. 同上 JSON 的排版版（便于人眼查看 `structured_brief` 结构）",
+        "",
+        "说明：若与第 2 节有任何不一致，以第 2 节（真实入参）为准。",
+        "",
+        "```json",
+        json.dumps(payload, ensure_ascii=False, indent=2),
+        "```",
+        "",
+    ]
+    out_path.write_text("\n".join(lines), encoding="utf-8")
+    print(f"Wrote {out_path} ({out_path.stat().st_size // 1024} KB)")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

backend/pipeline/llm/generate_strategy.py

@@ -13,6 +13,39 @@ from ..reporting.strategy_draft import (
 )
 from .llm_client import call_llm, estimate_chat_input_tokens, llm_context_window_size
 
+
+def _omit_ch8_probe_wordchart_fields(compact: dict[str, Any]) -> None:
+    """
+    第八章文本挖掘（探针）为主时，去掉与**预设关注词/场景条形图**同源的统计字段，
+    避免与报告 §8 文本挖掘主口径「两用数据」。
+
+    仅影响传入大模型的 ``structured_brief``；``brief`` 全量仍可由规则稿使用。
+    """
+    for k in (
+        "comment_focus_keywords",
+        "usage_scenarios",
+        "usage_scenarios_denominator",
+        "usage_scenarios_by_matrix_group",
+    ):
+        compact.pop(k, None)
+    cfb = compact.get("consumer_feedback_by_matrix_group")
+    if not isinstance(cfb, list):
+        return
+    slim: list[Any] = []
+    for g in cfb:
+        if not isinstance(g, dict):
+            slim.append(g)
+            continue
+        slim.append(
+            {
+                k: v
+                for k, v in g.items()
+                if k not in ("focus_keyword_hits", "scenarios_top")
+            }
+        )
+    compact["consumer_feedback_by_matrix_group"] = slim
+
+
 STRATEGY_DATA_RULES = """**全局禁止编造（适用于输出全文各节、各表、各段；独立策略稿与报告第九章策略归纳**共用**本段，硬性）**：
 - **事实与数字**：销量、GMV、占比、价带、条数、份额、券面额、满减/满折门槛、到手价、店铺/品牌计数与排名、SKU 数、接口返回量等，**仅可**来自**本次调用输入 JSON** 中已给出的字段（策略稿为 `structured_brief`、`rules_draft_markdown` 内摘录、`report_strategy_excerpt`、可选 **`report_matrix_group_evidence_md`**（与同任务报告第五～第八章细类大模型小节同源）、`strategy_decisions`、`business_notes`；第九章嵌入为 `competitor_brief`、可选 `prior_chapter_llm_narratives`）；**禁止**凭空新增、改口径或写成「已监测证实」而无字段支撑。
 - **主体与名称**：**禁止**引入上述输入中**未出现**的**具体**品牌名、店铺名、SKU 名、商品标题作为**事实陈述**；若 `strategy_decisions`/备注/brief/节选已含则可写；否则用「头部/同类竞品」等泛称或「待业务指定对标」。
@@ -153,13 +186,7 @@ def generate_strategy_draft_markdown_llm(
         compact = compact_brief_for_llm(brief, max_chars=compact_max)
         if report_uses_chapter8_text_mining_probe(report_config):
             compact = dict(compact)
-            for k in (
+            _omit_ch8_probe_wordchart_fields(compact)
-                "comment_focus_keywords",
-                "usage_scenarios",
-                "usage_scenarios_denominator",
-                "usage_scenarios_by_matrix_group",
-            ):
-                compact.pop(k, None)
             if isinstance(compact.get("strategy_hints"), list):
                 compact["strategy_hints"] = filter_strategy_hints_for_ch8_probe(
                     compact["strategy_hints"]
@@ -195,9 +222,9 @@ def generate_strategy_draft_markdown_llm(
         }
         if report_uses_chapter8_text_mining_probe(report_config):
             payload["structured_brief_omission_note"] = (
-                "已启用第八章文本挖掘（探针为主）：structured_brief 已省略「关注词/场景子串计数」及「与条形图同源的 strategy_hints 句子」，"
+                "已启用第八章文本挖掘（探针为主）：structured_brief 已省略顶层「关注词/场景子串计数」、按细类 feedback 中的 focus_keyword_hits/scenarios_top，"
-                "避免与报告 §8 主口径冲突。**不得**再以词频或预设场景占比作为论据。"
+                "以及「与条形图同源的 strategy_hints 句子」，避免与报告 §8 主口径冲突。**不得**再以这类子串计数或预设场景占比作为论据。"
-                "用户与评论侧须依报告 §8 文本挖掘归纳；**促销、满减、券价差**须与报告第六章、`price_promotion_signals` 及下方 `report_strategy_excerpt`（第九章）对齐，不得省略报告已写明的活动建议。"
+                "用户与评论侧须依报告 §8 文本挖掘归纳及 `report_matrix_group_evidence_md`；**促销、满减、券价差**须与报告第六章、`price_promotion_signals` 及下方 `report_strategy_excerpt`（第九章）对齐，不得省略报告已写明的活动建议。"
             )
         raw = json.dumps(payload, ensure_ascii=False)
         if len(raw) > 500_000:

backend/pipeline/tests/test_strategy_draft.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 
 from django.test import SimpleTestCase
 
+from pipeline.llm.generate_strategy import _omit_ch8_probe_wordchart_fields
 from pipeline.reporting.strategy_draft import build_strategy_draft_markdown
 
 
@@ -163,3 +164,24 @@ class StrategyDraftTests(SimpleTestCase):
         self.assertIn("样本内品牌较分散", md)
         self.assertIn("price_promotion_signals", md)
         self.assertIn("price_promotion_signals", md)
+
+    def test_ch8_probe_omit_wordchart_nested_in_consumer_feedback(self) -> None:
+        compact = {
+            "comment_focus_keywords": [{"word": "x", "count": 1}],
+            "usage_scenarios": [],
+            "consumer_feedback_by_matrix_group": [
+                {
+                    "group": "饼干",
+                    "comment_rows": 10,
+                    "focus_keyword_hits": [{"word": "口感", "count": 5}],
+                    "scenarios_top": [{"scenario": "早餐", "count": 2}],
+                }
+            ],
+        }
+        _omit_ch8_probe_wordchart_fields(compact)
+        self.assertNotIn("comment_focus_keywords", compact)
+        self.assertNotIn("focus_keyword_hits", compact["consumer_feedback_by_matrix_group"][0])
+        self.assertNotIn("scenarios_top", compact["consumer_feedback_by_matrix_group"][0])
+        self.assertEqual(
+            compact["consumer_feedback_by_matrix_group"][0].get("comment_rows"), 10
+        )