From 50a9867d0218feff21389e6a433735ae2456275a Mon Sep 17 00:00:00 2001
From: hub-gif <2487812171@qq.com>
Date: Tue, 21 Apr 2026 11:43:50 +0800
Subject: [PATCH] =?UTF-8?q?fix(pipeline):=208.3=E6=83=85=E6=84=9F=E5=AD=90?=
 =?UTF-8?q?=E6=A0=87=E9=A2=98=E9=99=8D=E4=B8=BA=E4=BA=94=E7=BA=A7=E4=BB=A5?=
 =?UTF-8?q?=E4=BE=BF=E7=BB=86=E7=B1=BB=E6=8A=A5=E5=91=8A=E6=91=98=E5=BD=95?=
 =?UTF-8?q?=E5=91=BD=E4=B8=AD?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

strategy 侧 load_report_matrix_group_evidence 按 #### 细类 抽取时在下一行 #### 即截断，导致正/负向主题正文丢失；在 matrix_group_focus 成稿后对四个固定小节做 ####→##### 后处理，并补充单测。

Made-with: Cursor
---
 backend/pipeline/llm/generate_sections.py     | 30 ++++++++++++++++++-
 .../reporting/report_matrix_group_evidence.py |  2 ++
 .../test_report_matrix_group_evidence.py      | 22 ++++++++++++++
 3 files changed, 53 insertions(+), 1 deletion(-)

diff --git a/backend/pipeline/llm/generate_sections.py b/backend/pipeline/llm/generate_sections.py
index c51f857..3d002ee 100644
--- a/backend/pipeline/llm/generate_sections.py
+++ b/backend/pipeline/llm/generate_sections.py
@@ -34,6 +34,31 @@ SENTIMENT_LLM_SYSTEM = """你是电商/食品类用户研究助手。输入 JSON
 
 **篇幅**：若 JSON 含 ``matrix_group_focus``（单细类范围），本节总字数约 **500～1200 字**，勿再按全关键词池写「全行业泛化」；若**不含**该字段（全量池），总字数约 **700～1600 字**。简体中文，语气客观。"""
 
+# 嵌入报告 8.3 时外层为 ``#### {细类名}``；若内文仍用同级 ``#### 正向体验主题``，
+# ``extract_level4_sections_by_group_title`` 会在第一个子 ``####`` 处截断，导致策略摘录/心得侧「同细类报告摘录」拿不到正文。
+_SENTIMENT_INNER_H4_TITLES: frozenset[str] = frozenset(
+    {
+        "正向体验主题",
+        "负向评价主题归因",
+        "混合评价中的典型张力",
+        "使用注意",
+    }
+)
+
+
+def demote_sentiment_inner_h4_to_h5_for_matrix_group(md: str) -> str:
+    """将情感归纳四个固定小节从 ``####`` 降为 ``#####``，以便嵌在 ``#### 细类`` 下仍能被按细类抽取。"""
+    out_lines: list[str] = []
+    for line in (md or "").splitlines():
+        m = re.match(r"^####\s+(.+)$", line)
+        if m:
+            title = m.group(1).strip()
+            if title in _SENTIMENT_INNER_H4_TITLES:
+                out_lines.append(f"##### {title}")
+                continue
+        out_lines.append(line)
+    return "\n".join(out_lines)
+
 
 def generate_comment_sentiment_analysis_llm(payload: dict[str, Any]) -> str:
     """基于 lexicon 统计 + 语义池与按词表归类的抽样，生成评价情感归纳段落（Markdown）；**默认不**嵌入竞品报告正文。"""
@@ -60,7 +85,10 @@ def generate_comment_sentiment_analysis_llm(payload: dict[str, Any]) -> str:
     if len(raw) > 88_000:
         raw = raw[:82_000] + "\n\n…（输入过长已截断，请勿编造截断外内容）\n"
     user = "请根据以下 JSON 按系统说明输出 Markdown：" + scope_note + "\n\n" + raw
-    return call_llm(SENTIMENT_LLM_SYSTEM, user)
+    out = call_llm(SENTIMENT_LLM_SYSTEM, user)
+    if isinstance(mg, str) and mg.strip():
+        out = demote_sentiment_inner_h4_to_h5_for_matrix_group(out)
+    return out
 
 
 def split_competitor_report_for_bridges(
diff --git a/backend/pipeline/reporting/report_matrix_group_evidence.py b/backend/pipeline/reporting/report_matrix_group_evidence.py
index bf75156..98d91f7 100644
--- a/backend/pipeline/reporting/report_matrix_group_evidence.py
+++ b/backend/pipeline/reporting/report_matrix_group_evidence.py
@@ -4,6 +4,8 @@
 
 报告生成侧约定：矩阵/价盘/促销/评论/场景等 LLM 小节均以 ``#### `` + 与矩阵一致的细类名为小节标题
 （见 ``generate_group_summaries`` 系统提示）。
+**8.3 评价正/负向主题**：细类下内层小节使用 ``#####``（由 ``demote_sentiment_inner_h4_to_h5_for_matrix_group`` 处理），
+避免与外层 ``#### 细类名`` 同级导致本节正文抽取为空。
 """
 from __future__ import annotations
 
diff --git a/backend/pipeline/tests/test_report_matrix_group_evidence.py b/backend/pipeline/tests/test_report_matrix_group_evidence.py
index bf0e0c5..54b95da 100644
--- a/backend/pipeline/tests/test_report_matrix_group_evidence.py
+++ b/backend/pipeline/tests/test_report_matrix_group_evidence.py
@@ -3,6 +3,7 @@ from __future__ import annotations
 
 from django.test import SimpleTestCase
 
+from pipeline.llm.generate_sections import demote_sentiment_inner_h4_to_h5_for_matrix_group
 from pipeline.reporting.report_matrix_group_evidence import (
     extract_level4_sections_by_group_title,
 )
@@ -37,3 +38,24 @@ B 段评论归纳。
             extract_level4_sections_by_group_title("## 二\n", "饼干"),
             [],
         )
+
+    def test_extract_8_3_sentiment_when_inner_subsections_are_h5(self) -> None:
+        """8.3 在 #### 细类 下须用 ##### 子标题，否则抽取在首个 #### 子节处断开。"""
+        inner = demote_sentiment_inner_h4_to_h5_for_matrix_group(
+            "#### 正向体验主题\n\n正文A\n\n#### 负向评价主题归因\n\n正文B\n"
+        )
+        md = (
+            "### 8.3 评价正/负向主题\n\n"
+            "#### 饼干\n\n"
+            f"{inner}\n"
+        )
+        parts = extract_level4_sections_by_group_title(md, "饼干")
+        self.assertEqual(len(parts), 1)
+        self.assertIn("正文A", parts[0])
+        self.assertIn("正文B", parts[0])
+
+    def test_demote_known_h4_titles_only(self) -> None:
+        raw = "#### 正向体验主题\nx\n#### 饼干\ny\n"
+        out = demote_sentiment_inner_h4_to_h5_for_matrix_group(raw)
+        self.assertIn("##### 正向体验主题", out)
+        self.assertIn("#### 饼干", out)