fix(strategy): 文本挖掘开启时策略稿不再依赖关注词/场景计数

报告配置 chapter8_text_mining_probe 为 True（默认）时，规则底稿第五节改为说明口径并去掉子串命中列表；LLM 请求从 structured_brief 剔除相关字段并附 omission_note。关闭探针时保留旧版第五节。任务 API 传入 report_config。

Made-with: Cursor

backend/pipeline/llm/generate_strategy.py

@@ -6,7 +6,10 @@ import os
 from typing import Any
 
 from ..reporting.brief_compact import compact_brief_for_llm
-from ..reporting.strategy_draft import build_strategy_draft_markdown
+from ..reporting.strategy_draft import (
+    build_strategy_draft_markdown,
+    report_uses_chapter8_text_mining_probe,
+)
 from .llm_client import call_llm, estimate_chat_input_tokens, llm_context_window_size
 
 STRATEGY_SYSTEM = """你是市场策略顾问，根据**结构化监测摘要**与业务侧填写的**决策字段**，把「规则底稿」润色为可读的策略 Markdown。
@@ -33,6 +36,7 @@ def generate_strategy_draft_markdown_llm(
     generated_at_iso: str,
     strategy_decisions: dict[str, Any],
     report_strategy_excerpt: str | None = None,
+    report_config: dict[str, Any] | None = None,
 ) -> str:
     """
     ``report_strategy_excerpt``：与同任务宿主报告第九章「策略与机会」正文对齐的节选（见
@@ -45,6 +49,7 @@ def generate_strategy_draft_markdown_llm(
         business_notes=business_notes,
         generated_at_iso=generated_at_iso,
         strategy_decisions=strategy_decisions,
+        report_config=report_config,
     )
     excerpt_raw = (report_strategy_excerpt or "").strip()
     sys_prompt = STRATEGY_SYSTEM
@@ -58,6 +63,15 @@ def generate_strategy_draft_markdown_llm(
         rules_max: int | None,
     ) -> str:
         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 (
+                "comment_focus_keywords",
+                "usage_scenarios",
+                "usage_scenarios_denominator",
+                "usage_scenarios_by_matrix_group",
+            ):
+                compact.pop(k, None)
         ex = (
             _truncate_strategy_narrative(excerpt_raw, excerpt_max)
             if excerpt_raw
@@ -77,6 +91,11 @@ def generate_strategy_draft_markdown_llm(
             "rules_draft_markdown": rd,
             "report_strategy_excerpt": ex,
         }
+        if report_uses_chapter8_text_mining_probe(report_config):
+            payload["structured_brief_omission_note"] = (
+                "已启用第八章文本挖掘：本 JSON 中的 structured_brief 已省略关注词/场景子串计数字段，"
+                "避免与当前报告正文口径冲突；请依规则底稿第五节说明与报告 §8 自拟假设，勿逐条编造子串命中列表。"
+            )
         raw = json.dumps(payload, ensure_ascii=False)
         if len(raw) > 500_000:
             payload["rules_draft_markdown"] = _truncate_rules_draft_md(rd, 200_000)

backend/pipeline/reporting/strategy_draft.py

@@ -98,6 +98,19 @@ def _risk_line(checked: bool, text: str) -> str:
     return f"- {mark} {text}"
 
 
+def report_uses_chapter8_text_mining_probe(report_config: dict[str, Any] | None) -> bool:
+    """
+    与任务 ``report_config`` 中 ``chapter8_text_mining_probe`` 一致；未显式设置时默认 ``True``
+    （与 ``jd.runner.get_default_report_config`` 一致）。
+    为 ``True`` 时，策略稿第五节不再逐条列举关注词/场景子串命中，以免与当前报告正文口径冲突。
+    """
+    if not isinstance(report_config, dict):
+        return True
+    if "chapter8_text_mining_probe" in report_config:
+        return bool(report_config.get("chapter8_text_mining_probe"))
+    return True
+
+
 def build_strategy_draft_markdown(
     *,
     job_id: int,
@@ -106,8 +119,10 @@ def build_strategy_draft_markdown(
     business_notes: str = "",
     generated_at_iso: str = "",
     strategy_decisions: dict[str, Any] | None = None,
+    report_config: dict[str, Any] | None = None,
 ) -> str:
     """生成可下载的 Markdown：策略框架为主，附录为数据速览。"""
+    use_ch8_probe = report_uses_chapter8_text_mining_probe(report_config)
     d = strategy_decisions or {}
     pos = _esc(d.get("positioning_choice") or "").strip()
     kw = _esc(brief.get("keyword")) or _esc(keyword) or "—"
@@ -253,31 +268,41 @@ def build_strategy_draft_markdown(
     ckw = brief.get("comment_focus_keywords") or []
     usc = brief.get("usage_scenarios") or []
     lines.extend(["## 五、用户需求与场景 — 可写成策略的假设", ""])
-    lines.append(
-        "*下列由关注词/场景**计数**转化而来，是「待验证假设」而非结论；请结合评价原文抽样修订。*"
-    )
-    lines.append("")
-    if ckw:
-        for item in ckw[:8]:
-            if isinstance(item, dict):
-                w = _esc(item.get("word"))
-                c = _num(item.get("count"))
-                lines.append(
-                    f"- **假设**：用户决策中「{w}」被频繁提及（约 {c} 次统计命中）—— "
-                    f"*可追问：本品故事是否正面回应？传播关键词是否覆盖？*"
-                )
-    if usc:
-        for item in usc[:6]:
-            if isinstance(item, dict):
-                sc = _esc(item.get("scenario"))
-                cn = _num(item.get("count"))
-                sh = _pct(item.get("share_of_text_units"))
-                lines.append(
-                    f"- **场景命题**：「{sc}」在预设场景中约 {cn} 条、约占 {sh} 文本单元—— "
-                    f"*可追问：主图/详情/客服话术是否对齐该场景？*"
-                )
-    if not ckw and not usc:
-        lines.append("*摘要中无关注词/场景组结果，请补全评论侧分析后再写本节。*")
+    if use_ch8_probe:
+        lines.extend(
+            [
+                "*当前任务报告以**第八章评论侧文本挖掘**为主线（替代原关注词/场景条形图为主呈现）时，**请勿**再逐条照搬「子串命中次数」式假设列表——结构化摘要里仍可能含 `comment_focus_keywords` / `usage_scenarios` 等字段，**与当前报告正文展示口径可能不一致**。*",
+                "",
+                "*请结合报告中 **§8** 评论与情感归纳、文本挖掘小节，以及评价原文抽样，自拟若干条「待验证假设」（需可复核、可证伪）。*",
+                "",
+            ]
+        )
+    else:
+        lines.append(
+            "*下列由关注词/场景**计数**转化而来，是「待验证假设」而非结论；请结合评价原文抽样修订。*"
+        )
+        lines.append("")
+        if ckw:
+            for item in ckw[:8]:
+                if isinstance(item, dict):
+                    w = _esc(item.get("word"))
+                    c = _num(item.get("count"))
+                    lines.append(
+                        f"- **假设**：用户决策中「{w}」被频繁提及（约 {c} 次统计命中）—— "
+                        f"*可追问：本品故事是否正面回应？传播关键词是否覆盖？*"
+                    )
+        if usc:
+            for item in usc[:6]:
+                if isinstance(item, dict):
+                    sc = _esc(item.get("scenario"))
+                    cn = _num(item.get("count"))
+                    sh = _pct(item.get("share_of_text_units"))
+                    lines.append(
+                        f"- **场景命题**：「{sc}」在预设场景中约 {cn} 条、约占 {sh} 文本单元—— "
+                        f"*可追问：主图/详情/客服话术是否对齐该场景？*"
+                    )
+        if not ckw and not usc:
+            lines.append("*摘要中无关注词/场景组结果，请补全评论侧分析后再写本节。*")
     lines.append("")
 
     hints = brief.get("strategy_hints") or []
@@ -316,11 +341,16 @@ def build_strategy_draft_markdown(
     rk = bool(d.get("ack_risk_keywords"))
     rp = bool(d.get("ack_risk_price"))
     rc = bool(d.get("ack_risk_concentration"))
+    rk_kw = (
+        "评论侧量化/关键词归纳是否**以偏概全**？（需原评论抽样）"
+        if use_ch8_probe
+        else "关注词/场景是否**以偏概全**？（需原评论抽样）"
+    )
     lines.extend(
         [
             "## 七、风险与待证伪",
             "",
-            _risk_line(rk, "关注词/场景是否**以偏概全**？（需原评论抽样）"),
+            _risk_line(rk, rk_kw),
             _risk_line(rp, "价格带是否含大促/异常挂价？（需核对清洗规则）"),
             _risk_line(rc, "列表集中度与深入样本品牌是否**矛盾**？（需解释渠道差异）"),
             "",

backend/pipeline/tests/test_strategy_draft.py

@@ -45,6 +45,7 @@ class StrategyDraftTests(SimpleTestCase):
             keyword="K",
             brief=brief,
             strategy_decisions=decisions,
+            report_config={"chapter8_text_mining_probe": False},
         )
         self.assertIn("**本品角色**：追赶型", md)
         self.assertIn("- [x] **卡腰**", md)
@@ -54,3 +55,48 @@ class StrategyDraftTests(SimpleTestCase):
         self.assertIn("- [x] 关注词/场景是否**以偏概全**", md)
         self.assertIn("- [ ] 价格带是否含大促", md)
         self.assertIn("- [x] 列表集中度与深入样本品牌是否**矛盾**", md)
+
+    def test_chapter8_probe_omits_focus_scenario_count_bullets(self) -> None:
+        brief = {
+            "schema_version": 1,
+            "keyword": "低GI",
+            "comment_focus_keywords": [{"word": "口感", "count": 501}],
+            "usage_scenarios": [
+                {
+                    "scenario": "控糖/血糖相关",
+                    "count": 305,
+                    "share_of_text_units": 0.272,
+                }
+            ],
+        }
+        md = build_strategy_draft_markdown(
+            job_id=1,
+            keyword="低GI",
+            brief=brief,
+            report_config={"chapter8_text_mining_probe": True},
+        )
+        self.assertIn("文本挖掘", md)
+        self.assertNotIn("假设**：用户决策中「口感」", md)
+        self.assertNotIn("场景命题**：「控糖", md)
+
+    def test_legacy_report_shows_focus_scenario_bullets(self) -> None:
+        brief = {
+            "schema_version": 1,
+            "keyword": "低GI",
+            "comment_focus_keywords": [{"word": "口感", "count": 501}],
+            "usage_scenarios": [
+                {
+                    "scenario": "控糖/血糖相关",
+                    "count": 305,
+                    "share_of_text_units": 0.272,
+                }
+            ],
+        }
+        md = build_strategy_draft_markdown(
+            job_id=1,
+            keyword="低GI",
+            brief=brief,
+            report_config={"chapter8_text_mining_probe": False},
+        )
+        self.assertIn("假设**：用户决策中「口感」", md)
+        self.assertIn("场景命题**：「控糖/血糖相关」", md)

backend/pipeline/views/job_report_views.py

@@ -158,6 +158,7 @@ class JobStrategyDraftView(APIView):
             report_excerpt, excerpt_src = load_report_strategy_excerpt(job.run_dir)
         except OSError:
             report_excerpt, excerpt_src = "", "none"
+        rc_job = job.report_config if isinstance(job.report_config, dict) else None
         try:
             if generator == "llm":
                 md = generate_strategy_draft_markdown_llm(
@@ -168,6 +169,7 @@ class JobStrategyDraftView(APIView):
                     generated_at_iso=gen_at,
                     strategy_decisions=strategy_decisions,
                     report_strategy_excerpt=report_excerpt,
+                    report_config=rc_job,
                 )
                 src = "llm_text_ai_crawler_v1"
             else:
@@ -178,6 +180,7 @@ class JobStrategyDraftView(APIView):
                     business_notes=notes,
                     generated_at_iso=gen_at,
                     strategy_decisions=strategy_decisions,
+                    report_config=rc_job,
                 )
                 src = "structured_summary_rules_v1"
         except ValueError as e: