辉哥版本：结构化聚类溯源归因与业务报告增强。

移除结构化 audience 字段，强化 voc_业务_2 源评论归因匹配与 Persona 引用展示，更新 README 与流水线默认清理 SQLite。 Co-authored-by: Cursor <cursoragent@cursor.com>
2026-06-17 09:11:04 +08:00 · 2026-06-17 09:11:04 +08:00 · c91e7f8c20
commit c91e7f8c20
parent cb6692c0c0
23 changed files with 1493 additions and 602 deletions
--- a/README.md
+++ b/README.md
@ -1,257 +1,240 @@
-# VOC LLM 结构化分析 (VOC_LLM结构化)
+# 亚马逊评论 VOC 结构化分析

-> 基于大语言模型（[DeepSeek](https://api.deepseek.com) OpenAI 兼容 API）与本地 MLX 向量的亚马逊 VOC（Voice of Customer）评论分析流水线：合并 CSV → 清洗 → LLM 结构化 → 向量化 → 聚类与词频 → 生成 HTML 分析报告。
+> **分支「辉哥版本」**：可结构化、可聚类、可溯源归因的亚马逊站内评论分析流水线。  
+> 基于 DeepSeek Chat API + 本地 MLX 向量，完成：合并 CSV → 清洗 → LLM 结构化 → 向量化 → 聚类 → 词频 → **业务 HTML 报告**（Persona / 差评主题 / 根因 / KANO，每条洞察可回溯到真实评论）。

-## 📖 目录
+**仓库**：https://git.onesvm.com/whoops/amz_review_analyse

- [核心特性](#-核心特性)
- [环境要求](#-环境要求)
- [安装指南](#-安装指南)
- [使用说明](#-使用说明)
- [示例与输出](#-示例与输出)
- [项目结构](#-项目结构)
- [常见问题](#-常见问题)
- [参与贡献](#-参与贡献)
- [开源协议](#-开源协议)
- [联系方式与鸣谢](#-联系方式与鸣谢)
+---

-## ✨ 核心特性
+## 目录

- **七步全流程编排** — `main_voc分析.py` 一键串联：合并、清洗、结构化、向量化、聚类、词频、HTML 报告
- **断点续跑** — 支持 `--from-step` / `--only-step`，从任意步骤恢复，调试时节省 API 成本
- **LLM 结构化提取** — 从评论中抽取受众、痛点、方面、观点、情感等字段（`prompts/schema.yaml` 可配置）
- **本地向量化** — Apple Silicon 上运行 `Qwen3-Embedding-4B-mxfp8`（MLX），无需云端 Embedding API
- **语义聚类** — UMAP + HDBSCAN 多阶段聚类，辅以 LLM 评估簇质量自动调参
- **词频分析** — LLM 归纳专有名词 + spaCy 全量词频统计，报告内嵌词云与六类归类
- **可编辑 Prompt** — `prompts/` 目录下 Markdown / YAML 热加载，产品运营可直接改话术（见 `prompts/README.md`）
- **并行加速** — 聚类与词频在步骤 5–6 由线程池并行执行；结构化批间并行（默认 8 路）
+- [核心能力](#核心能力)
+- [快速开始（推荐）](#快速开始推荐)
+- [主流程说明](#主流程说明)
+- [结构化字段](#结构化字段)
+- [溯源与归因](#溯源与归因)
+- [环境要求](#环境要求)
+- [安装](#安装)
+- [项目结构](#项目结构)
+- [常见问题](#常见问题)

-## 🛠 环境要求
+---
+
+## 核心能力
+
+| 能力 | 说明 |
+|------|------|
+| **LLM 结构化提取** | 从评论抽取 `persona_signals`（画像信号）、`pain_points`（需求痛点）、`product_feedback`（方面/观点/情感/类别） |
+| **本地向量化** | Apple Silicon 上运行 `Qwen3-Embedding-4B-mxfp8`（MLX），无需云端 Embedding |
+| **语义聚类** | UMAP + HDBSCAN：`3a` 全量痛点、`3b` 按情感分桶的 aspect-opinion 聚类 |
+| **Persona 发现** | 绑定聚类簇 + keywords 二次过滤，统计命中数与占比 |
+| **差评主题 & 根因** | LLM 归纳主题；根因引用由系统从真实评论回填（结构化字段优先匹配） |
+| **可溯源 HTML 报告** | Persona 卡片、根因区块展示源评论原文 + ASIN 链接；附录展示结构化/聚类抽样 |
+| **断点续跑** | `main_voc分析.py --from-step` / `run_pipeline.py --from-step` |
+
+---
+
+## 快速开始（推荐）
+
+业务报告入口在 `voc_业务_2/`，一键跑数据管道并生成 HTML：
+
+```bash
+cd voc_业务_2
+
+# 1. 编辑 config.yaml：input_dir 指向原始评论 CSV 目录
+# 2. 全流程（step 1–6 + 自动 build_report）
+../310py/bin/python run_pipeline.py --input-dir "../你的评论CSV目录"
+```
+
+产物示例：
+
+- 根目录 SQLite：`voc_structured.sqlite`、`voc_embeddings.sqlite`、`voc_clustering.sqlite`
+- HTML 报告：`voc_业务_2/output/{产品slug}-voc-report.html`
+
+仅重跑报告（数据库已就绪）：
+
+```bash
+../310py/bin/python build_report.py --product "产品名" --industry "行业"
+```
+
+---
+
+## 主流程说明
+
+### 方式 A：`voc_业务_2/run_pipeline.py`（业务报告）
+
+```
+原始 CSV → main_voc分析 step 1–6 → build_report.py → HTML
+```
+
+- 每步默认**清理旧 SQLite**（不加 `--keep-db`），避免与历史 job 混用
+- 产品名/行业可在 `config.yaml` 留空，由 LLM 从评论样本自动识别
+
+### 方式 B：`main_voc分析.py`（含经典 voc_report）
+
+```bash
+./310py/bin/python main_voc分析.py \
+  --input-dir "reviews_export" \
+  --product "Bikini Trimmer" \
+  --industry "个人护理"
+```
+
+七步：合并 → 清洗 → 结构化 → 向量化 → 聚类 → 词频 → HTML（`voc_report.py`）。
+
+断点示例：
+
+```bash
+./310py/bin/python main_voc分析.py --from-step 4      # 从向量化续跑
+./310py/bin/python main_voc分析.py --only-step 7      # 仅重生成 voc_report
+```
+
+更细步骤见 **[main_voc分析.md](main_voc分析.md)**、**[VOC分析方法论与报告生成逻辑.md](VOC分析方法论与报告生成逻辑.md)**。
+
+---
+
+## 结构化字段
+
+当前 schema（`prompts/schema.yaml`）根字段为 **3 项**（已移除 `audience`）：
+
+```json
+{
+  "persona_signals": ["sensitive skin", "travel grooming"],
+  "pain_points": ["ingrown hair"],
+  "product_feedback": [
+    {
+      "aspect": "battery life",
+      "opinion": "dies after one use",
+      "sentiment": "Negative",
+      "category": "Function"
+    }
+  ]
+}
+```
+
+- 所有字段值须为**自然英文**（多语言评论先理解再英文输出）
+- `product_feedback.category` 优先 8 类标准类别（Trust / Ingredient / Quality / Function / Appearance / Logistics / Customer Service / Price）
+
+Prompt 编辑入口：`prompts/extraction/`、`voc_业务_2/prompts.yaml`。
+
+---
+
+## 溯源与归因
+
+整条链路通过 **`source_row`**（与 `merged_reviews_cleaned.csv` 行号一致）关联：
+
+```
+评论原文 (CSV)
+    ↓ source_row
+comment_extractions (voc_structured.sqlite)
+    ↓ extraction_id / source_row
+embedding_items (voc_embeddings.sqlite)
+    ↓
+cluster_assignments (voc_clustering.sqlite)  → Persona 绑定簇 → source_rows 命中池
+    ↓
+build_report.py HTML
+    ├── Persona 卡片：命中池内结构化/原文匹配，展示 1–3 条源评论
+    ├── 根因分析：Negative product_feedback 优先匹配，展示 1–4 条
+    └── 附录（voc_report）：结构化 JSON + 聚类短语抽样
+```
+
+配置项（`voc_业务_2/config.yaml`）：
+
+```yaml
+persona_quote_max: 3      # Persona 卡片最多展示条数
+rootcause_quote_max: 4    # 每条根因最多展示条数
+```
+
+---
+
+## 环境要求

 | 依赖 | 说明 |
 |------|------|
-| Python | >= 3.10（推荐 3.12，项目内 `310py`） |
-| pip / uv | 安装 `requirements.txt` 中的包 |
-| spaCy 英文模型 | 经 `uv pip` 安装 `en-core-web-sm`（见安装指南，词频步骤必需） |
-| DeepSeek API Key | 结构化、聚类评估、词频、报告等 Chat 步骤 |
-| 本地 Embedding 模型 | 目录 `Qwen3-Embedding-4B-mxfp8/`（约 4GB，已 gitignore，需自行下载） |
-| Apple Silicon | 本地向量化依赖 MLX（M 系列芯片） |
+| Python | ≥ 3.10（推荐 3.12，项目内 `310py`） |
+| DeepSeek API Key | 结构化、Persona/主题/根因/KANO 等 Chat 步骤 |
+| 本地 Embedding 模型 | `Qwen3-Embedding-4B-mxfp8/`（约 4GB，gitignore，需自行下载） |
+| Apple Silicon | 本地 MLX 向量化 |
+| spaCy `en_core_web_sm` | 词频步骤 |

-**Chat API Key**（任选其一，勿提交到 Git）：
+**API Key**（任选其一，勿提交 Git）：

-1. 环境变量 `DEEPSEEK_API_KEY`
-2. 环境变量 `DEEPSEEK_API_KEY_FILE` 指向单行密钥文件
-3. 项目根目录 `.deepseek_key`（单行，无引号）
+```bash
+export DEEPSEEK_API_KEY="sk-xxx"
+# 或项目根 .deepseek_key（已被 .gitignore）
+```

-可选：`DEEPSEEK_MODEL`（默认 `deepseek-v4-pro`）、`DEEPSEEK_BASE_URL`（默认 `https://api.deepseek.com`）。
+---

-## 📦 安装指南
-
-1. 克隆项目到本地：
+## 安装

 ```bash
 git clone https://git.onesvm.com/whoops/amz_review_analyse.git
-cd amz_review_analyse   # 或你的本地目录名
-```
+cd amz_review_analyse

-2. 创建虚拟环境并安装依赖（推荐）：
-
-```bash
 uv venv 310py --python 3.12
 uv pip install --python 310py/bin/python -r requirements.txt
 uv pip install --python 310py/bin/python \
  "en-core-web-sm @ https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl"
 ```

-3. 配置 DeepSeek API Key：
-
-```bash
-export DEEPSEEK_API_KEY="sk-xxx"
-# 或在项目根创建 .deepseek_key（已被 .gitignore 忽略）
-```
-
-4. 准备本地 Embedding 模型（首次向量化前）：
-
-将 `Qwen3-Embedding-4B-mxfp8` 放到项目根，或设置 `VOC_EMBED_MODEL_PATH` 指向模型目录。可从 [Hugging Face](https://huggingface.co/mlx-community/Qwen3-Embedding-4B-mxfp8) 下载。
-
-5. 准备原始评论 CSV 目录（目录内所有 `*.csv` 表头须一致），例如亚马逊导出的 `*_realtime.csv`。
-
-## 🚀 使用说明
-
-### 全流程分析
-
-```bash
-./310py/bin/python main_voc分析.py \
-  --input-dir "reviews_export" \
-  --product "cat deterrent indoor" \
-  --industry "Pet Supplies"
-```
-
- `--input-dir`：原始 CSV 目录
- `--product`：产品名（写入结构化任务与报告路径）
- `--industry`：行业名，默认 `-`（可在步骤 3 写入库）
- `--keep-db`：保留已有结构化 `voc_*.sqlite`，不覆盖删除
-
-### 加速（批间并行，默认已开启）
-
-步骤 3 结构化默认多批并行 Chat 请求；步骤 4 向量为本地 MLX 串行批处理（勿对同一模型多线程）：
-
-```bash
-# 全流程
-./310py/bin/python main_voc分析.py --input-dir "reviews_export" --product "产品名"
-
-# 调低结构化并发（遇 429 时）
-./310py/bin/python main_voc分析.py --input-dir "reviews_export" --product "产品名" \
-  --struct-workers 4
-
-# 环境变量：export VOC_STRUCT_WORKERS=8 VOC_EMBED_BATCH_SIZE=16
-```
-
-### 断点续跑
-
-```bash
-# 从向量化起续跑（步骤 4 起可省略 --product，自动读结构化库）
-./310py/bin/python main_voc分析.py --from-step 4 --keep-db
-
-# 仅重跑词频（复用已有 voc_terms.json）
-./310py/bin/python main_voc分析.py --from-step 6 --skip-wordfreq-llm
-
-# 仅重新生成 HTML 报告
-./310py/bin/python main_voc分析.py --only-step 7
-```
-
-### 其他常用参数
-
-| 参数 | 说明 |
-|------|------|
-| `--clean-intermediates` | 报告成功后删除中间 csv/sqlite，减少磁盘占用 |
-| `--filter-small-clusters` | 报告仅保留簇内评论占比 ≥ 10% 的簇 |
-| `--save-llm-raw` | 将报告 LLM 原文保存为 `report_llm_raw.txt`，调试时使用 |
-
-### 程序式调用
-
-```python
-from pathlib import Path
-from main_voc分析 import run_voc_analysis
-
-result = run_voc_analysis(
-    input_dir=Path("reviews_export"),
-    industry="Pet Supplies",
-    product_name="cat deterrent indoor",
-    from_step=1,
-    clean_databases=True,
-)
-print(result["report_html"])
-```
-
-### Prompt 验收（无需 API Key）
-
-```bash
-./310py/bin/python prompts/smoke.py          # 检查 prompt 能否加载
-./310py/bin/python prompts/smoke.py --live   # 联调模型（需 DEEPSEEK_API_KEY）
-```
-
-### 可选变体：jieba 词频
-
-中文或需 jieba 分词时，可使用 `main_voc分析_jieba.py`（词频走 `词频_jieba.py`，其余步骤与主流程一致）。
+Embedding 模型：从 [Hugging Face mlx-community/Qwen3-Embedding-4B-mxfp8](https://huggingface.co/mlx-community/Qwen3-Embedding-4B-mxfp8) 下载到项目根，或设置 `VOC_EMBED_MODEL_PATH`。

 ---

-更详细的步骤说明、算法与 SQLite 约定见 **[main_voc分析.md](main_voc分析.md)**。
-
-## 📸 示例与输出
-
-流程结束后，主要产物如下：
-
-| 路径 | 说明 |
-|------|------|
-| `merged_reviews.csv` | 多文件合并结果 |
-| `merged_reviews_cleaned.csv` | 清洗、去重后的评论 |
-| `voc_structured.sqlite` | LLM 结构化结果 |
-| `voc_embeddings.sqlite` | 本地 Qwen3 向量（维度见库内 `dimensions` 字段） |
-| `voc_clustering.sqlite` | 多阶段聚类标签 |
-| `output/voc_terms.json` | 专有名词 / 停用词 |
-| `output/word_freq.csv` | 全量词频表 |
-| `output/{product}/{product}_voc_report.html` | **最终 VOC 分析报告**（词云、词频、分簇、AI 正文） |
-
-stdout 会打印 JSON 摘要（含 `report_html` 等键）。
-
-## 📂 项目结构
+## 项目结构

 ```text
-VOC_LLM结构化/
-├── main_voc分析.py          # 主流程编排入口（七步）
-├── main_voc分析_jieba.py    # 词频使用 jieba 的变体入口
-├── main_voc分析.md          # 流程与算法详细说明
-├── voc_llm.py               # DeepSeek Chat 密钥与客户端
-├── local_embedding.py       # 本地 MLX Qwen3 向量化
-├── 合并评论数据.py          # 步骤 1：多 CSV 合并
-├── content清洗.py           # 步骤 2：评论清洗与去重
-├── 结构化_server.py         # 步骤 3：LLM 结构化入库
-├── 结构化_Prompt.py         # 结构化 prompt 组装
-├── 向量化.py                # 步骤 4：本地 Embedding 入库
-├── 聚类.py                  # 步骤 5：UMAP + HDBSCAN
-├── 词频.py / 词频_jieba.py  # 步骤 6：术语提取 + 词频
-├── voc_report.py            # 步骤 7：HTML 报告生成
-├── prompts/                 # 可编辑 prompt、schema、配置
-├── Qwen3-Embedding-4B-mxfp8/ # 本地模型（gitignore，需自行放置）
-├── requirements.txt
-├── output/                  # 报告与词频输出（gitignore）
-└── README.md                # 本文件
+├── main_voc分析.py              # 七步主流程编排
+├── 结构化_server.py             # LLM 结构化 → voc_structured.sqlite
+├── 向量化.py                    # persona_signal / pain / aspect_opinion 向量
+├── 聚类.py                      # 3a 痛点 + 3b 情感分桶聚类
+├── voc_report.py                # 经典 HTML 报告（Dashboard + 附录验证）
+├── prompts/                     # 结构化 & 报告 Prompt（可热加载）
+├── voc_业务_2/
+│   ├── run_pipeline.py          # ★ 业务一键流水线
+│   ├── build_report.py          # ★ 业务 HTML 报告生成
+│   ├── llm_analyzer.py          # Persona / 主题 / KANO / 根因 LLM
+│   ├── report_utils.py          # 统计、引用匹配、HTML 拼装
+│   ├── data_loader.py           # SQLite / CSV 加载
+│   ├── config.yaml              # 产品路径、并发、引用条数等
+│   └── template.html            # 报告模板
+└── output/                      # 报告与词频（gitignore）
 ```

-## ❓ 常见问题
-
-**Q：提示缺少 `DEEPSEEK_API_KEY`？**  
-A：按上文配置环境变量或 `.deepseek_key`，并确认密钥未提交到仓库。
-
-**Q：只有 `.dashscope_key` 报错？**  
-A：Chat 已切换为 DeepSeek，DashScope 密钥不能用于 `api.deepseek.com`，请改用 `.deepseek_key`。
-
-**Q：步骤 4 向量化失败 / 找不到模型？**  
-A：确认 `Qwen3-Embedding-4B-mxfp8/` 在项目根，或设置 `VOC_EMBED_MODEL_PATH`；需在 Apple Silicon + Python 3.10+ 环境。
-
-**Q：步骤 4 报错找不到 `product`？**  
-A：从步骤 1–3 完整跑过，或确保 `voc_structured.sqlite` 中已有最新 job；步骤 4 起可省略 `--product`。
-
-**Q：词频步骤报 spaCy 模型缺失？**  
-A：执行 `uv pip install --python 310py/bin/python "en-core-web-sm @ https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl"`。
-
-**Q：合并 CSV 失败？**  
-A：确保 `--input-dir` 下所有 CSV 表头完全一致。
-
-**Q：修改 LLM 话术后报告解析失败？**  
-A：勿修改 `prompts/schema.yaml` 中 `report.markers` 四段标记名；改完运行 `./310py/bin/python prompts/smoke.py` 验收。
-
-## 🤝 参与贡献
-
-欢迎提交 Issue 与 Pull Request。建议流程：
-
-1. Fork 本仓库
-2. 创建特性分支：`git checkout -b feature/your-feature`
-3. 提交更改：`git commit -m '简要说明变更'`
-4. 推送并发起 Pull Request
-
-修改主流程或 CLI 时，请同步更新 `main_voc分析.md`；修改 `prompts/` 时请遵循 `prompts/README.md` 中的占位符与 schema 约定。
-
-**安全提醒**：勿提交 `.deepseek_key`、`.dashscope_key`、`.env`、真实评论 CSV、`*.sqlite`、`output/` 及本地模型目录（见 `.gitignore`）。
-
-## 📄 开源协议
-
-本项目尚未在仓库中附带 `LICENSE` 文件。若为内部项目，请按组织规范使用；若计划开源，请补充协议文件（如 MIT）并更新本节链接。
-
-## ✉️ 联系方式与鸣谢
-
- **详细技术文档**：[main_voc分析.md](main_voc分析.md)、[prompts/README.md](prompts/README.md)
- **项目仓库**：https://git.onesvm.com/whoops/amz_review_analyse
-
-### 鸣谢
-
- [DeepSeek API](https://api.deepseek.com) — Chat 结构化、聚类评估、词频与报告
- [mlx-community/Qwen3-Embedding-4B-mxfp8](https://huggingface.co/mlx-community/Qwen3-Embedding-4B-mxfp8) — 本地向量化
- [UMAP](https://umap-learn.readthedocs.io/)、[HDBSCAN](https://hdbscan.readthedocs.io/) — 聚类管线
- [spaCy](https://spacy.io/) — 英文词频与 NLP
-
 ---

-*README 与 `main_voc分析.py` 七步流程保持一致；深度说明请参阅 `main_voc分析.md`。*
+## 常见问题
+
+**Q：Persona 显示「命中 N 条」但没有源评论？**  
+A：确保已用最新 `report_utils.pick_persona_quotes` 重跑 `build_report.py`；命中池有数据时会多层回退展示，优先结构化 `persona_signals` / `pain_points` 匹配。
+
+**Q：结构化校验报 persona_signals 错误？**  
+A：「辉哥版本」已移除 `audience` 字段；请重跑 step 3 生成新格式 JSON，勿混用旧库。
+
+**Q：提示缺少 API Key？**  
+A：配置 `DEEPSEEK_API_KEY` 或 `.deepseek_key`（Chat 已用 DeepSeek，DashScope 密钥不可用）。
+
+**Q：向量化找不到模型？**  
+A：确认 `Qwen3-Embedding-4B-mxfp8/` 在项目根或设置 `VOC_EMBED_MODEL_PATH`。
+
+---
+
+## 分支说明
+
+| 分支 | 说明 |
+|------|------|
+| `main` | 基础流水线 + DeepSeek + 本地 Embedding |
+| **`辉哥版本`** | 业务报告（voc_业务_2）、结构化三字段、聚类溯源、Persona/根因源评论归因 |
+
+---
+
+## 鸣谢
+
+- [DeepSeek API](https://api.deepseek.com) — Chat 结构化与分析
+- [mlx-community/Qwen3-Embedding-4B-mxfp8](https://huggingface.co/mlx-community/Qwen3-Embedding-4B-mxfp8) — 本地向量化
+- [UMAP](https://umap-learn.readthedocs.io/)、[HDBSCAN](https://hdbscan.readthedocs.io/) — 聚类
+
+---
+
+*详细算法与 SQLite 表结构见 [main_voc分析.md](main_voc分析.md)。*
--- a/main_voc分析.py
+++ b/main_voc分析.py
@ -60,6 +60,8 @@ DEFAULT_INDUSTRY = "亚马逊电商"
 DEFAULT_PRODUCT = "亚马逊商品"


+
+
 def _safe_product_dir_name(product_name: str) -> str:
    """用于 output 子目录名；去除路径非法字符。"""
    name = (product_name or "product").strip()
--- a/prompts/README.md
+++ b/prompts/README.md
@ -16,7 +16,7 @@
 ## 占位符

 `.md` 文件使用 Python `format` 语法，例如 `{product_name}`、`{industry}`。  
-正文里需要字面量花括号时写双花括号：`{{"audience": "unknown"}}`。
+正文里需要字面量花括号时写双花括号：`{{"persona_signals": []}}`。

 ## 改完怎么验收

--- a/prompts/extraction/batch_output_format.md
+++ b/prompts/extraction/batch_output_format.md
@ -2,6 +2,6 @@
 本批共 {n_keys} 条评论，用户消息中每条评论以 [C1]、[C2]… 前缀标识。
 - 只输出一个 JSON 对象；顶层键必须且仅能是：{keys_literal}
 - 每个顶层键对应一条同前缀评论，不得遗漏、不得新增其他顶层键
- 每个键的值是单条结构化对象，仅含 audience、pain_points、product_feedback 三个字段
+- 每个键的值是单条结构化对象，仅含 persona_signals、pain_points、product_feedback 三个字段
 - 不要用 JSON 数组作为顶层；不要把多条评论合并进一个对象；不要用 results、data 等包裹层
 - 不要输出 markdown 代码围栏或任何解释文字
--- a/prompts/extraction/examples.yaml
+++ b/prompts/extraction/examples.yaml
@ -2,7 +2,9 @@
 - instruction: "示例 1 教学：展示正常长评论如何标准提取，如何准确分类产品优点、缺陷以及物流问题。"
  review: "Bought this for my 12yo lab who struggles with stairs. It is very soft and helps her get onto the bed easily. But the zipper broke after a week and the shipping box was damaged."
  output:
-    audience: "12yo lab"
+    persona_signals:
+      - "large breed dog"
+      - "mobility limited pet"
    pain_points:
      - "struggles with stairs"
      - "difficulty getting onto bed"
@ -23,7 +25,8 @@
 - instruction: "示例 2 教学：防混淆与隐性属性提炼。什么是「产品缺陷（如讲太快）」，它属于 product_feedback，绝不是 pain_points（用户需求）！并提炼出 teaching speed 这个隐性反馈对象。"
  review: "The course content is good, the teacher speaks too fast, can't keep up."
  output:
-    audience: "student"
+    persona_signals:
+      - "online course learner"
    pain_points: []
    product_feedback:
      - aspect: "course content"
@ -38,7 +41,9 @@
 - instruction: "示例 3 教学：深度上下文推理。如何通过动作（抓跳蚤）推理出使用者（pet），如何通过「全家能睡觉」推理出用户需求（失眠），以及如何将「Worth every penny」准确归类为对 Price（价格）的产品反馈。"
  review: "Finally found something that stops the midnight flea scratching! The chemical smell is a bit strong initially, but it fades. Worth every penny since our family can finally sleep."
  output:
-    audience: "pet"
+    persona_signals:
+      - "flea problem pet"
+      - "household sleep disruption"
    pain_points:
      - "midnight flea scratching"
      - "sleep deprivation"
@ -55,7 +60,8 @@
 - instruction: "示例 4 教学：1) 非英语评论（例如西班牙语）输出必须全英文，禁止保留原文非英文片段，先理解语义再用英文短语表达；2) 「随附配件」不属于 8 个标准类别，演示如何补充一个简洁的英文新类别 'Accessories'。"
  review: "El producto en sí está muy bien, viene con 2 cristales templados y 2 grips para los joycon. Pero es lo que he leido en algunos comentarios, la funda huele como a tabaco jajaja es raro pero es así"
  output:
-    audience: "self"
+    persona_signals:
+      - "joycon grip user"
    pain_points: []
    product_feedback:
      - aspect: "product"
@ -74,7 +80,9 @@
 - instruction: "示例 5 教学：保留关键成分与症状等核心具体词汇。绝不能将带有具体成分/病症的词泛化提取。例如遇到「chicken flavor」（鸡肉风味）或「joint pain」（关节疼痛）时，必须保留核心修饰词，绝不能错误缩减提取为「flavor」或「pain」。"
  review: "My elderly cat suffers from severe joint pain. But he is very attracted by the chicken flavor of this supplement! It really helps him walk better."
  output:
-    audience: "elderly cat"
+    persona_signals:
+      - "senior cat"
+      - "joint pain cat"
    pain_points:
      - "severe joint pain"
    product_feedback:
--- a/prompts/extraction/field_rules.md
+++ b/prompts/extraction/field_rules.md
@ -1,12 +1,19 @@
 ## 语言规则（红线）:
 - 评论输入可能是英语、西班牙语、法语、德语、日语等**任意语言**。
- **所有输出字段值**（audience、pain_points 每一条、aspect、opinion）**必须是自然英文**。
+- **所有输出字段值**（persona_signals 每一条、pain_points 每一条、aspect、opinion）**必须是自然英文**。
 - **禁止**保留原文非英文片段（如 está muy bien、huele como a tabaco、très bon）；先理解语义，再用英文短语表达。

 ## 分析要求:

-1. audience (为谁购买):
-   - 提取出实际的使用者，使用简短的英文名词。若无明确提及，请根据上下文推理；若完全无法推理则输出 'unknown'。
+1. persona_signals (用户画像信号):
+   - **字符串数组**；每条为 2–8 词的英文短语，描述身份、体质、场景或使用背景。
+   - 从评论原文推断，典型信号：
+     - 体质/生理：`sensitive skin`、`coarse thick hair`、`pregnant`、`elderly`
+     - 场景/行为：`travel grooming`、`shower use`、`bikini area`、`first-time buyer`
+     - 自我标注句式：`I have …` / `As a …` / `My skin is …` 须提炼为短语写入此处
+   - **禁止**仅写 `self` / `user` / `unknown` 等泛化词。
+   - 不得把产品缺陷/差评（如 `pulls hair`、`battery dead`）写入 persona_signals；产品体验属于 product_feedback。
+   - 若无任何可区分信号，可输出 `[]`。

 2. pain_points (用户需求):
   - 仅限提取用户在购买前遇到的外部困扰、疾病、或具体场景（购买前尚未被本产品解决的需求）。
--- a/prompts/extraction/filter_irrelevant.md
+++ b/prompts/extraction/filter_irrelevant.md
@ -1,3 +1,3 @@
 5. 无关评论过滤：
-   - 若某条评论明显与{product_name}无关（其他品类、其他 SKU 或完全跑题），该条输出：{{"audience": "unknown", "pain_points": [], "product_feedback": []}}。
-   - 不得将无关内容填入 audience、pain_points 或 product_feedback。
+   - 若某条评论明显与{product_name}无关（其他品类、其他 SKU 或完全跑题），该条输出：{{"persona_signals": [], "pain_points": [], "product_feedback": []}}。
+   - 不得将无关内容填入 persona_signals、pain_points 或 product_feedback。
--- a/prompts/extraction/format_constraints_batch.md
+++ b/prompts/extraction/format_constraints_batch.md
@ -1,3 +1,3 @@
 4. 单条评论对象内的格式与字段约束：
-   - 每个评论对象只能包含 `audience`, `pain_points`, `product_feedback` 这 3 个字段。
+   - 每个评论对象只能包含 `persona_signals`, `pain_points`, `product_feedback` 这 3 个字段。
   - **绝对不要**在 JSON 中输出 `instruction`、`教学说明` 或其他任何多余字段。
--- a/prompts/extraction/format_constraints_single.md
+++ b/prompts/extraction/format_constraints_single.md
@ -1,4 +1,4 @@
 4. 格式与字段约束：
-   - **你的 JSON 输出只能包含 `audience`, `pain_points`, `product_feedback` 这 3 个根字段。**
+   - **你的 JSON 输出只能包含 `persona_signals`, `pain_points`, `product_feedback` 这 3 个根字段。**
   - **绝对不要**在 JSON 中输出 `instruction`、`教学说明` 或其他任何多余字段。
   - 必须以纯 JSON 格式输出结果，不要包含任何 markdown 标记（如 ```json ）或其他解释性文字。
--- a/prompts/schema.yaml
+++ b/prompts/schema.yaml
@ -3,7 +3,7 @@

 extraction:
  root_fields:
-    - audience
+    - persona_signals
    - pain_points
    - product_feedback
  product_feedback_fields:
--- a/voc_report.py
+++ b/voc_report.py
@ -321,7 +321,7 @@ def _extraction_to_display_zh(ext: Dict[str, Any]) -> Dict[str, Any]:
    if not isinstance(pains, list):
        pains = []
    return {
-        "受众": str(ext.get("audience", "")).strip(),
+        "画像信号": [str(s).strip() for s in ext.get("persona_signals") or [] if str(s).strip()],
        "需求/痛点": [str(p).strip() for p in pains if str(p).strip()],
        "产品反馈": feedback,
    }
--- a/voc_业务_2/build_report.py
+++ b/voc_业务_2/build_report.py
@ -39,7 +39,7 @@ from llm_analyzer import (
 from report_utils import (
    enrich_theme_keywords, recalc_neg_priorities, compute_persona_pcts,
    assign_persona_clusters, validate_persona_physiological_labels,
-    enrich_persona_catalog_physio_counts,
+    prepare_persona_catalog_for_llm,
    pick_persona_quotes, fix_jtbd_fields, filter_matrix_rows,
    enrich_matrix_scene_evidence, build_matrix_table_rows_html,
    enrich_rootcause_quotes, build_executive_summary, enhanced_market_judgment,
@ -53,6 +53,7 @@ from report_utils import (
    build_product_stopwords, build_kano_grid_html, build_neg_theme_summary_note,
    build_neg_theme_table_rows, build_pos_theme_table_rows,
    build_sentiment_keyword_groups, prepare_keyword_display, build_keyword_tables_html,
+    build_quote_blocks_html,
 )
 from echarts_builder import (
    get_echarts_script, build_all_charts, calc_theme_freq, calc_per_asin_theme_freq,
@ -134,13 +135,20 @@ def build_report_data(loader: DataLoader, cfg: dict) -> Dict[str, Any]:
    data["cluster_data"] = cluster_data
    all_clusters = loader.load_cluster_data()
    reviews = loader.load_reviews()
+    extractions = loader.load_comment_extractions()
+    data["extractions"] = extractions
+    persona_quote_max = int(cfg.get("persona_quote_max", 3))
+    rootcause_quote_max = int(cfg.get("rootcause_quote_max", 4))
    physio_min = int(cfg.get("persona_physio_min_reviews", 5))
-    enrich_persona_catalog_physio_counts(
+    min_hit_count = int(cfg.get("persona_min_hit_count", 5))
+    data["persona_min_hit_count"] = min_hit_count
+    prepare_persona_catalog_for_llm(
        cluster_data["persona_cluster_catalog"], all_clusters, reviews,
+        min_physio_reviews=physio_min,
    )

    # ── Persona ──
-    personas = discover_personas(cluster_data)
+    personas = discover_personas(cluster_data, product_name=product, industry=industry)
    personas = assign_persona_clusters(personas, all_clusters)
    personas = validate_persona_physiological_labels(
        personas, cluster_data, reviews, all_clusters, min_review_count=physio_min,
@ -156,11 +164,15 @@ def build_report_data(loader: DataLoader, cfg: dict) -> Dict[str, Any]:
    data["neg_themes"] = neg_themes
    data["pos_themes"] = pos_themes

-    # ── 主题频次统计 ──
-    neg_freq = calc_theme_freq(neg_themes, reviews, is_neg=True)
-    pos_freq = calc_theme_freq(pos_themes, reviews, is_neg=False)
-    per_asin_neg = calc_per_asin_theme_freq(neg_themes, reviews, is_neg=True)
-    per_asin_pos = calc_per_asin_theme_freq(pos_themes, reviews, is_neg=False)
+    # ── 主题频次统计（结构化 category+aspect 优先） ──
+    neg_freq = calc_theme_freq(neg_themes, reviews, is_neg=True, extractions=extractions)
+    pos_freq = calc_theme_freq(pos_themes, reviews, is_neg=False, extractions=extractions)
+    per_asin_neg = calc_per_asin_theme_freq(
+        neg_themes, reviews, is_neg=True, extractions=extractions,
+    )
+    per_asin_pos = calc_per_asin_theme_freq(
+        pos_themes, reviews, is_neg=False, extractions=extractions,
+    )
    neg_themes = recalc_neg_priorities(
        neg_themes, neg_freq, stats.neg_review_count, len(stats.asins), per_asin_neg,
    )
@ -174,7 +186,10 @@ def build_report_data(loader: DataLoader, cfg: dict) -> Dict[str, Any]:
    personas = compute_persona_pcts(personas, reviews, all_clusters)
    personas = sort_personas_by_evidence(personas)
    data["personas"] = personas
-    data["persona_quotes"] = pick_persona_quotes(personas, reviews)
+    data["persona_quotes"] = pick_persona_quotes(
+        personas, reviews, extractions,
+        max_quotes=persona_quote_max,
+    )

    # ── KANO + JTBD + 情感关键词（三者并发） ──
    neg_kw_groups = build_sentiment_keyword_groups(neg_themes, reviews, limit=6, is_neg=True)
@ -212,21 +227,16 @@ def build_report_data(loader: DataLoader, cfg: dict) -> Dict[str, Any]:
    data["matrix"] = matrix

    # ── 根因分析（各 Persona 并发） ──
-    per_aud = loader.load_per_audience_clusters()
-    per_aud_prompt = {}
-    for al, d in per_aud.items():
-        per_aud_prompt[al] = {
-            "pains": [{"label": c.cluster_label, "top_phrases": c.top_phrases} for c in d["pain"]],
-            "negative": [{"label": c.cluster_label, "top_phrases": c.top_phrases} for c in d["negative"]],
-            "positive": [{"label": c.cluster_label, "top_phrases": c.top_phrases} for c in d["positive"]],
-        }
    rootcauses = analyze_all_rootcauses(
-        personas, per_aud_prompt, neg_themes, max_workers=llm_workers,
-        product_name=product, industry=industry, min_hit_count=10,
+        personas, neg_themes, all_clusters=all_clusters,
+        max_workers=llm_workers,
+        product_name=product, industry=industry, min_hit_count=min_hit_count,
    )
    rootcauses = enrich_rootcause_quotes(
-        rootcauses, personas, per_aud, loader, reviews,
+        rootcauses, personas, loader, reviews,
        persona_quotes=data["persona_quotes"],
+        extractions=extractions,
+        max_quotes=rootcause_quote_max,
    )
    rootcauses = normalize_rootcauses(rootcauses, neg_themes)
    rootcauses = sort_rootcauses_by_evidence(rootcauses, personas)
@ -411,21 +421,20 @@ def render_html(template_path: Path, data: Dict[str, Any], cfg: dict) -> str:
    persona_cards = []
    for i, p in enumerate(data["personas"]):
        pq = next((q for q in data["persona_quotes"] if q["persona"] == p["name"]), None)
-        quote_html = ""
-        if pq and pq.get("quote"):
-            q_cls = "neg" if pq.get("is_neg") else "pos"
-            star_hint = f" ({int(pq.get('rating', 0))}★)" if pq.get("rating") else ""
-            cn = pq.get("cn_summary") or quote_cn_summary(pq["quote"])
-            quote_html = (
-                f'<div class="quote {q_cls}">"{pq["quote"]}" '
-                f'— <a href="{pq["amazon_url"]}" target="_blank" rel="noopener">'
-                f'{asin_labels.get(pq["asin"], pq["asin"])}</a>{star_hint}'
-                f'<div class="quote-cn">摘要：{cn}</div></div>'
-            )
-        elif pq:
-            quote_html = '<div class="quote-empty">暂无与画像 keywords 匹配的代表性评论</div>'
+        quote_items = (pq.get("quotes") or []) if pq else []
+        if not quote_items and pq and pq.get("quote"):
+            quote_items = [pq]
+        quote_html = build_quote_blocks_html(
+            quote_items,
+            asin_labels,
+            empty_msg="暂无命中池内的代表性评论",
+        )
        normalize_persona_dimension(p)
-        meta = persona_display_meta(p, stats.total_reviews)
+        min_hit = int(
+            data.get("persona_min_hit_count")
+            or cfg.get("persona_min_hit_count", 5)
+        )
+        meta = persona_display_meta(p, stats.total_reviews, min_hit_count=min_hit)
        persona_cards.append(f"""<div class="persona">
  <div class="p-name">{p.get("name", "?")}</div>
  <div class="p-meta">{meta}</div>
@ -651,26 +660,41 @@ def main():
            raw = json.load(f)
        stats = loader.load_basic_stats()
        reviews = loader.load_reviews()
+        extractions = loader.load_comment_extractions()
+        persona_quote_max = int(cfg.get("persona_quote_max", 3))
+        rootcause_quote_max = int(cfg.get("rootcause_quote_max", 4))
        all_clusters = loader.load_cluster_data()
        personas = assign_persona_clusters(raw.get("personas", []), all_clusters)
        cluster_data_rr = raw.get("cluster_data") or loader.build_cluster_prompt_data(
            persona_sample_reviews=int(cfg.get("persona_sample_reviews", 5)),
        )
-        enrich_persona_catalog_physio_counts(
+        physio_min = int(cfg.get("persona_physio_min_reviews", 5))
+        prepare_persona_catalog_for_llm(
            cluster_data_rr.get("persona_cluster_catalog") or [],
            all_clusters, reviews,
+            min_physio_reviews=physio_min,
        )
-        physio_min = int(cfg.get("persona_physio_min_reviews", 5))
        personas = validate_persona_physiological_labels(
            personas, cluster_data_rr, reviews, all_clusters, min_review_count=physio_min,
        )
        personas = compute_persona_pcts(personas, reviews, all_clusters)
        personas = sort_personas_by_evidence(personas)
-        persona_quotes = pick_persona_quotes(personas, reviews)
-        neg_freq = calc_theme_freq(raw.get("neg_themes", []), reviews, is_neg=True)
-        pos_freq = calc_theme_freq(raw.get("pos_themes", []), reviews, is_neg=False)
-        per_asin_neg = calc_per_asin_theme_freq(raw.get("neg_themes", []), reviews, is_neg=True)
-        per_asin_pos = calc_per_asin_theme_freq(raw.get("pos_themes", []), reviews, is_neg=False)
+        persona_quotes = pick_persona_quotes(
+            personas, reviews, extractions,
+            max_quotes=persona_quote_max,
+        )
+        neg_freq = calc_theme_freq(
+            raw.get("neg_themes", []), reviews, is_neg=True, extractions=extractions,
+        )
+        pos_freq = calc_theme_freq(
+            raw.get("pos_themes", []), reviews, is_neg=False, extractions=extractions,
+        )
+        per_asin_neg = calc_per_asin_theme_freq(
+            raw.get("neg_themes", []), reviews, is_neg=True, extractions=extractions,
+        )
+        per_asin_pos = calc_per_asin_theme_freq(
+            raw.get("pos_themes", []), reviews, is_neg=False, extractions=extractions,
+        )
        neg_themes = recalc_neg_priorities(
            raw.get("neg_themes", []), neg_freq, stats.neg_review_count, len(stats.asins), per_asin_neg,
        )
@ -710,14 +734,14 @@ def main():
        market_title, market_desc = enhanced_market_judgment(stats)
        data["market_title"] = market_title
        data["market_desc"] = market_desc
-        per_aud = loader.load_per_audience_clusters()
        data["rootcauses"] = enrich_rootcause_quotes(
            data.get("rootcauses") or [],
            personas,
-            per_aud,
            loader,
            reviews,
            persona_quotes=persona_quotes,
+            extractions=extractions,
+            max_quotes=rootcause_quote_max,
        )
    elif args.no_llm:
        logger.warning("--no-llm 模式：跳过 LLM 调用，仅渲染模板")
--- a/voc_业务_2/config.yaml
+++ b/voc_业务_2/config.yaml
@ -30,6 +30,15 @@ prompts_file: "./prompts.yaml"
 persona_sample_reviews: 5
 # Persona 生理标签：绑定簇内至少 N 条评论含对应英文词方可保留（如 pregnant≥5 才可用「孕妇」）
 persona_physio_min_reviews: 5
+# Persona 命中门槛：hit_count 低于此值跳过根因 LLM，卡片置信度显示「低样本」
+persona_min_hit_count: 5
+# Persona 卡片展示的代表性评论条数（结构化字段匹配 + 原文回退）
+persona_quote_max: 3
+# 每条根因展示的代表性评论条数
+rootcause_quote_max: 4
+# 进入 step 2 细粒度聚类的 audience 簇最小占比（相对总评论数）
+audience_coverage_threshold: 0.05
+

 # ── LLM 配置 ──
 # 复用根目录 voc_llm.py 的配置（DEEPSEEK_API_KEY 环境变量 或 .deepseek_key）
--- a/voc_业务_2/data_loader.py
+++ b/voc_业务_2/data_loader.py
@ -7,11 +7,12 @@ from __future__ import annotations
 import csv
 import json
 import logging
+import math
 import sqlite3
 from collections import Counter, defaultdict
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Set, Tuple

 logger = logging.getLogger("voc.data_loader")

@ -22,6 +23,52 @@ CLEANED_CSV_NAME = "merged_reviews_cleaned.csv"
 WORD_FREQ_CSV_NAME = "output/word_freq.csv"
 TERMS_JSON_NAME = "output/voc_terms.json"

+# 与 report_utils._PERSONA_PHRASE_STOPWORDS 保持一致（避免循环 import）
+_PERSONA_PHRASE_STOPWORDS = frozenset({
+    "self", "unknown", "user", "customer", "buyer", "myself", "the user",
+    "a user", "consumer", "reviewer", "amazon", "product", "item",
+})
+
+
+def _phrase_segments_local(phrases: List[str]) -> List[str]:
+    """将聚类短语拆成可匹配的英文片段。"""
+    segs: List[str] = []
+    for p in phrases:
+        p = (p or "").lower().strip()
+        if not p:
+            continue
+        segs.append(p)
+        for part in p.split(","):
+            part = part.strip()
+            if len(part) >= 3:
+                segs.append(part)
+    return list(dict.fromkeys(segs))
+
+
+def _cluster_semantic_label(
+    cluster: "ClusterData",
+    phrase_doc_freq: Dict[str, int],
+    n_clusters_in_stage: int,
+) -> str:
+    """从 top_phrases 生成簇短语义标签（频次 × IDF，同类 stage 内去重）。"""
+    scored: List[Tuple[float, str]] = []
+    for rank, phrase in enumerate(cluster.top_phrases):
+        tf = 1.0 / (1 + rank)
+        for seg in _phrase_segments_local([phrase]):
+            if seg in _PERSONA_PHRASE_STOPWORDS or len(seg) < 4:
+                continue
+            df = phrase_doc_freq.get(seg, 1)
+            idf = math.log((n_clusters_in_stage + 1) / (df + 0.5))
+            scored.append((tf * idf, seg))
+    if scored:
+        scored.sort(key=lambda x: (-x[0], -len(x[1])))
+        return scored[0][1][:48]
+    for phrase in cluster.top_phrases:
+        p = (phrase or "").strip()
+        if len(p) >= 4:
+            return p[:48]
+    return f"cluster_{cluster.cluster_label}"
+

@dataclass
 class ReviewRecord:
@ -69,6 +116,63 @@ class MarketStats:
    asins: List[ASINStats] = field(default_factory=list)


+# Persona A/B/C 维度信号词（用于运行时判定簇的 suggested_dimension）
+_A_DIM_SIGNALS = frozenset({
+    # 生理/物理特征 — 跨品类通用
+    "sensitive skin", "sensitive", "allergy", "allergic",
+    "elderly", "senior", "older",
+    "kid", "child", "children", "toddler", "infant", "baby", "newborn",
+    "pregnant", "pregnancy", "nursing", "breastfeeding",
+    "diabetic", "diabetes",
+    "my dog", "my cat", "my pet", "puppy", "kitten", "pet owner",
+    "professional", "beginner", "first time user",
+    "large breed", "small breed", "small dog", "large dog",
+    "oily skin", "dry skin", "acne", "eczema", "psoriasis",
+    "curly hair", "fine hair", "color treated",
+})
+_B_DIM_SIGNALS = frozenset({
+    # 行为/使用场景 — 跨品类通用
+    "travel", "on the go", "portable", "lightweight", "compact",
+    "daily", "everyday", "first time", "first-time", "beginner",
+    "outdoor", "indoor", "home", "office", "kitchen", "car",
+    "quick", "easy to use", "simple", "convenient",
+    "maintenance", "cleaning", "storage", "organization",
+    "camping", "hiking", "gym", "workout",
+    "cooking", "baking", "cleaning house",
+})
+_C_DIM_SIGNALS = frozenset({
+    # 购买动机/背景 — 跨品类通用
+    "gift", "present", "birthday", "christmas", "holiday",
+    "for my wife", "for my husband", "for my mom", "for my daughter",
+    "replacement", "replace", "upgrade", "switched from",
+    "price", "cheap", "expensive", "affordable", "worth the money",
+    "waste of money", "not worth", "overpriced", "good value",
+    "recommend", "recommended", "saw on", "social media",
+    "amazon", "online", "review", "reviews", "rating",
+    "bought", "purchased", "ordered", "arrived",
+})
+
+
+def _suggest_dimension(stage: str, top_phrases: List[str]) -> str:
+    """根据簇的 top_phrases 内容判定建议的 Persona 维度（A/B/C）。"""
+    blob = " ".join(top_phrases).lower()
+    a_hits = sum(1 for w in _A_DIM_SIGNALS if w in blob)
+    b_hits = sum(1 for w in _B_DIM_SIGNALS if w in blob)
+    c_hits = sum(1 for w in _C_DIM_SIGNALS if w in blob)
+    if a_hits >= 2:
+        return "A"
+    if a_hits >= 1 and (b_hits + c_hits) == 0:
+        return "A"
+    if b_hits > c_hits and b_hits >= 2:
+        return "B"
+    if c_hits > b_hits and c_hits >= 2:
+        return "C"
+    # 默认按 stage 推理
+    if stage.startswith("3a"):
+        return "C"
+    return "B"
+
+
 class DataLoader:
    def __init__(self, project_root: Path, product_name: str = "", industry: str = ""):
        self.project_root = Path(project_root).resolve()
@ -142,6 +246,44 @@ class DataLoader:
                return r
        return None

+    def load_comment_extractions(
+        self,
+        job_id: Optional[int] = None,
+    ) -> Dict[int, Dict[str, Any]]:
+        """加载 source_row → extraction_json 映射（最新 job 或指定 job_id）。"""
+        if not self.structured_db.is_file():
+            logger.warning("未找到结构化库 %s，引用匹配将仅使用原文关键词", self.structured_db)
+            return {}
+        conn = sqlite3.connect(self.structured_db)
+        try:
+            if job_id is None:
+                row = conn.execute(
+                    "SELECT id FROM analysis_jobs ORDER BY id DESC LIMIT 1"
+                ).fetchone()
+                if not row:
+                    logger.warning("voc_structured.sqlite 中无结构化任务")
+                    return {}
+                job_id = int(row[0])
+            cur = conn.execute(
+                """
+                SELECT source_row, extraction_json
+                FROM comment_extractions
+                WHERE job_id = ?
+                ORDER BY source_row
+                """,
+                (job_id,),
+            )
+            out: Dict[int, Dict[str, Any]] = {}
+            for sr, js in cur.fetchall():
+                try:
+                    out[int(sr)] = json.loads(js)
+                except json.JSONDecodeError as e:
+                    logger.debug("跳过无效 extraction source_row=%s: %s", sr, e)
+            logger.info("已加载 %s 条结构化提取 (job_id=%s)", len(out), job_id)
+            return out
+        finally:
+            conn.close()
+
    def load_basic_stats(self) -> MarketStats:
        reviews = self.load_reviews()
        if not reviews:
@ -241,33 +383,46 @@ class DataLoader:
        logger.info("已加载聚类: %s stages, %s 簇", len(result), sum(len(v) for v in result.values()))
        return dict(result)

+
+
    def build_persona_cluster_catalog(
        self,
        sample_reviews_per_cluster: int = 5,
        max_review_chars: int = 320,
    ) -> List[Dict[str, Any]]:
-        """供 Persona LLM 绑定的聚类簇目录（含每簇代表性评论）。"""
+        """供 Persona LLM 绑定的聚类簇目录（含每簇代表性评论和运行时维度建议）。"""
        all_c = self.load_cluster_data()
-        dim_hint = {
-            "1_audience": "A",
-            "3a_pain_global": "C",
-            "3b_aspect_opinion_negative": "B",
-            "3b_aspect_opinion_positive": "B",
-        }
        catalog: List[Dict[str, Any]] = []
        for stage in (
-            "1_audience",
            "3a_pain_global",
            "3b_aspect_opinion_negative",
            "3b_aspect_opinion_positive",
        ):
-            for c in sorted(all_c.get(stage, []), key=lambda x: -x.review_count):
+            stage_clusters = sorted(all_c.get(stage, []), key=lambda x: -x.review_count)
+            phrase_doc_freq: Dict[str, int] = defaultdict(int)
+            for c in stage_clusters:
+                segs = set(_phrase_segments_local(c.top_phrases))
+                for seg in segs:
+                    if seg not in _PERSONA_PHRASE_STOPWORDS and len(seg) >= 4:
+                        phrase_doc_freq[seg] += 1
+            n_stage = max(len(stage_clusters), 1)
+            used_labels: Set[str] = set()
+            for c in stage_clusters:
+                label = _cluster_semantic_label(c, phrase_doc_freq, n_stage)
+                if label in used_labels:
+                    for rank, phrase in enumerate(c.top_phrases[1:], start=1):
+                        alt = (phrase or "").strip().lower()[:48]
+                        if alt and len(alt) >= 4 and alt not in used_labels:
+                            label = alt
+                            break
+                used_labels.add(label)
                entry: Dict[str, Any] = {
                    "stage": stage,
                    "label": c.cluster_label,
                    "review_count": c.review_count,
+                    "semantic_label": label,
                    "top_phrases": c.top_phrases[:12],
-                    "suggested_dimension": dim_hint.get(stage, "B"),
+                    "suggested_dimension": _suggest_dimension(stage, c.top_phrases),
                }
                if sample_reviews_per_cluster > 0:
                    quotes = self.get_representative_quotes(
@ -318,10 +473,8 @@ class DataLoader:
        self,
        persona_sample_reviews: int = 5,
    ) -> Dict[str, Any]:
-        audience = self.load_audience_clusters()
        global_pain = self.load_global_pain_clusters()
        global_fb = self.load_global_feedback_clusters()
-        per_aud = self.load_per_audience_clusters()
        stats = self.load_basic_stats()

        def _sum(clusters: List[ClusterData]) -> List[Dict[str, Any]]:
@ -332,18 +485,14 @@ class DataLoader:
                for c in sorted(clusters, key=lambda x: -x.phrase_count)
            ]
        return {
-            "audience_clusters": _sum(audience),
+            "audience_clusters": [],
            "global_pains": _sum(global_pain),
            "global_negative": _sum(global_fb["negative"]),
            "global_positive": _sum(global_fb["positive"]),
            "persona_cluster_catalog": self.build_persona_cluster_catalog(
                sample_reviews_per_cluster=persona_sample_reviews,
            ),
-            "per_audience": {
-                al: {"pains": _sum(d["pain"]), "negative": _sum(d["negative"]),
-                     "positive": _sum(d["positive"])}
-                for al, d in sorted(per_aud.items())
-            },
+            "per_audience": {},
            "total_reviews": stats.total_reviews,
            "neg_review_count": stats.neg_review_count,
            "pos_review_count": stats.pos_review_count,
--- a/voc_业务_2/echarts_builder.py
+++ b/voc_业务_2/echarts_builder.py
@ -683,12 +683,34 @@ def calc_keyword_group_freq(
    return count


+def _theme_matches_review(
+    keywords: List[str],
+    review: Any,
+    extractions: Dict[int, Any],
+    *,
+    is_neg: bool,
+) -> bool:
+    """主题命中：优先结构化 category+aspect，回退原文 keyword。"""
+    from report_utils import match_theme_extraction
+
+    kws = [kw.lower().strip() for kw in keywords if kw and kw.strip()]
+    if not kws:
+        return False
+    ext = extractions.get(review.source_row)
+    if ext and match_theme_extraction(kws, ext, is_neg=is_neg):
+        return True
+    text = (review.title + " " + review.content).lower()
+    return _match_theme_in_text(kws, text)
+
+
 def calc_theme_freq(
    themes: List[Dict[str, Any]],
    reviews: List[Any],
    is_neg: bool = True,
+    extractions: Dict[int, Any] | None = None,
 ) -> Dict[str, int]:
-    """用主题 keywords（英文聚类短语 + LLM 关键词）统计频次。"""
+    """用主题 keywords 统计频次（结构化 category+aspect 优先，原文回退）。"""
+    ext_map = extractions or {}
    freq: Dict[str, int] = {}
    for theme in themes:
        keywords = [kw.lower().strip() for kw in theme.get("keywords", []) if kw.strip()]
@ -698,8 +720,7 @@ def calc_theme_freq(
                continue
            if not is_neg and r.rating < 4:
                continue
-            text = (r.title + " " + r.content).lower()
-            if _match_theme_in_text(keywords, text):
+            if _theme_matches_review(keywords, r, ext_map, is_neg=is_neg):
                count += 1
        freq[theme["name"]] = count
    return freq
@ -709,9 +730,12 @@ def calc_per_asin_theme_freq(
    themes: List[Dict[str, Any]],
    reviews: List[Any],
    is_neg: bool = True,
+    extractions: Dict[int, Any] | None = None,
 ) -> Dict[str, Dict[str, int]]:
-    """按 ASIN 分别统计各主题频次。"""
+    """按 ASIN 分别统计各主题频次（结构化优先）。"""
    from collections import defaultdict
+
+    ext_map = extractions or {}
    result: Dict[str, Dict[str, int]] = defaultdict(lambda: defaultdict(int))
    for theme in themes:
        keywords = [kw.lower().strip() for kw in theme.get("keywords", []) if kw.strip()]
@ -721,7 +745,6 @@ def calc_per_asin_theme_freq(
                continue
            if not is_neg and r.rating < 4:
                continue
-            text = (r.title + " " + r.content).lower()
-            if _match_theme_in_text(keywords, text):
+            if _theme_matches_review(keywords, r, ext_map, is_neg=is_neg):
                result[r.asin][name] += 1
    return {asin: dict(counts) for asin, counts in result.items()}
--- a/voc_业务_2/llm_analyzer.py
+++ b/voc_业务_2/llm_analyzer.py
@ -131,10 +131,12 @@ def _j(obj: Any) -> str:

 # ── Persona ──

-def build_persona_prompt(cluster_data):
+def build_persona_prompt(cluster_data, product_name: str = "", industry: str = ""):
    catalog = cluster_data.get("persona_cluster_catalog") or []
    return user_prompt(
        "persona",
+        product_name=product_name or "主产品",
+        industry=industry or "当前品类",
        catalog_json=_j(catalog),
        audience_clusters_json=_j(cluster_data.get("audience_clusters", [])),
        global_pains_json=_j(cluster_data.get("global_pains", [])),
@ -143,11 +145,17 @@ def build_persona_prompt(cluster_data):
    )


-def discover_personas(cluster_data):
-    logger.info("LLM: Persona发现...")
+def discover_personas(
+    cluster_data,
+    product_name: str = "",
+    industry: str = "",
+):
+    logger.info("LLM: Persona发现（产品=%s）...", product_name or "主产品")
    params = llm_params("persona")
    raw = _call_llm(
-        system_prompt("persona"), build_persona_prompt(cluster_data), **params,
+        system_prompt("persona"),
+        build_persona_prompt(cluster_data, product_name, industry),
+        **params,
    )
    result = _parse_json(raw)
    personas = result.get("personas", [])
@ -277,8 +285,8 @@ def analyze_matrix(personas, kano, cluster_data, total_reviews, market_avg: floa

 def analyze_rootcause_per_persona(
    persona,
-    per_aud_data,
    neg_themes,
+    bound_clusters: Optional[Dict[str, Any]] = None,
    product_name: str = "",
    industry: str = "",
 ):
@ -287,13 +295,15 @@ def analyze_rootcause_per_persona(
    product_ctx = product_name or "主产品"
    industry_ctx = industry or "当前品类"
    logger.info("LLM: 根因-%s...", name)
+    # 构建 Persona 自身绑定簇的摘要数据（替代旧 per_aud_data）
+    bc_data = bound_clusters or {}
    user = user_prompt(
        "rootcause",
        persona_name=name,
        product_name=product_ctx,
        industry=industry_ctx,
        persona_json=_j(persona),
-        per_aud_data_json=_j(per_aud_data),
+        bound_clusters_json=_j(bc_data),
        theme_names_json=_j(theme_names),
    )
    params = llm_params("rootcause")
@ -304,27 +314,35 @@ def analyze_rootcause_per_persona(

 def analyze_all_rootcauses(
    personas: List[Dict[str, Any]],
-    per_audience: Dict[Any, Any],
    neg_themes: List[Dict[str, Any]],
+    all_clusters: Optional[Dict[str, List[Any]]] = None,
    max_workers: int = DEFAULT_LLM_WORKERS,
    product_name: str = "",
    industry: str = "",
-    min_hit_count: int = 10,
+    min_hit_count: int = 5,
 ) -> List[Dict[str, Any]]:
-    """并发按 Persona 根因分析；跳过命中不足的 Persona。"""
-    aud_keys = sorted(per_audience.keys())
+    """并发按 Persona 根因分析；跳过命中不足的 Persona。不再依赖 per_audience。"""

-    def _audience_data_for(persona: Dict[str, Any]) -> Dict[str, Any]:
-        aud_label = persona.get("audience_label")
-        if aud_label is None or aud_label < 0:
-            ref = persona.get("cluster_ref") or {}
-            if ref.get("stage") == "1_audience" and ref.get("label") is not None:
-                aud_label = int(ref["label"])
-        if aud_label is not None and aud_label >= 0 and aud_label in per_audience:
-            return per_audience[aud_label]
-        if aud_keys:
-            return per_audience.get(aud_keys[0], {})
-        return {}
+    def _bound_clusters_for(persona: Dict[str, Any]) -> Dict[str, Any]:
+        """从 Persona 自身 cluster_refs 构建绑定簇的摘要数据。"""
+        refs = persona.get("cluster_refs") or {}
+        bc_data: Dict[str, Any] = {}
+        if isinstance(refs, dict):
+            for dim, ref in refs.items():
+                if isinstance(ref, dict) and ref.get("stage") and ref.get("label") is not None:
+                    stage = ref["stage"]
+                    label = ref["label"]
+                    # 尝试从 all_clusters 中查找实际簇数据
+                    if all_clusters and stage in all_clusters:
+                        for c in all_clusters[stage]:
+                            if int(getattr(c, "cluster_label", -1)) == int(label):
+                                bc_data[f"{dim}_dim"] = {
+                                    "stage": stage,
+                                    "label": label,
+                                    "top_phrases": getattr(c, "top_phrases", [])[:10],
+                                }
+                                break
+        return bc_data

    eligible = [
        (i, p) for i, p in enumerate(personas)
@ -344,9 +362,10 @@ def analyze_all_rootcauses(
    logger.info("LLM: 并发根因分析 %s/%s 个 Persona（workers=%s）...", n, len(personas), workers)

    def _one(idx: int, persona: Dict[str, Any]) -> Dict[str, Any]:
-        aud_data = _audience_data_for(persona)
+        bc_data = _bound_clusters_for(persona)
        rc = analyze_rootcause_per_persona(
-            persona, aud_data, neg_themes,
+            persona, neg_themes,
+            bound_clusters=bc_data,
            product_name=product_name, industry=industry,
        )
        rc["persona_name"] = persona.get("name", f"P{idx}")
--- a/voc_业务_2/prompts.yaml
+++ b/voc_业务_2/prompts.yaml
@ -48,34 +48,38 @@ persona:
  user_template: |
    ## 任务：基于聚类数据按三维度发现用户画像（Persona），4-7 个。

+    ## 分析产品（product_detect / config 已识别，Persona 须针对该品类）
+    产品：{{product_name}} | 行业：{{industry}}
+    purchase_motivation 用「雇佣{{product_name}}完成…」JTBD 句式；Persona 痛点/需求须与该产品使用场景一致，禁止脱离品类写无关人群。
+
    ## 三维度强制覆盖（缺一不可）
-    A. 物理/生理/受众特征 → 必须绑定 stage=1_audience 的簇
-       典型信号：特定体质/肤质/年龄/体型/使用对象（如 sensitive / elderly / for kids / large breed）
-    B. 行为/使用场景 → 绑定 3b_aspect_opinion_positive / 3b_aspect_opinion_negative / 3a_pain_global
-       典型信号：travel / daily use / outdoor / first time / gift
-    C. 购买动机/背景 → 绑定 3a_pain_global 或 opinion 簇
-       典型信号：switched from / saw on social media / first time / too expensive
-    → 不能只有 B 和 C；A 维至少 1 个。若 A 维缺失，说明物理特征群体被遗漏，必须补建。
+    A. 物理/生理/受众特征 → 从 catalog 中 suggested_dimension="A" 的簇中选择
+       典型信号：特定体质/肤质/年龄/体型/使用对象（如 sensitive skin / coarse hair / pregnant / bikini area）
+    B. 行为/使用场景 / 具体痛点 → 从 catalog 中 suggested_dimension="B" 的簇中选择
+       典型信号：travel / first time / pulls hair / battery dead / durability
+    C. 购买动机/背景 → 从 catalog 中 suggested_dimension="C" 的簇中选择
+       典型信号：switched from / saw on social media / too expensive
+    → 每个 Persona 必须同时绑 B 或 C 维的具体痛点/需求簇，用于精准计算占比。

    ## 自我标注信号强制检查（归纳后必做）
-    在绑定簇的 top_phrases 中搜索以下模式，出现 ≥5 条则必须单独建 Persona：
-    - "I have [adj] [noun]"（如 I have sensitive skin / I have a large dog）
-    - "My [noun] is/are [adj]"（如 My pet is very anxious）
-    - "As a [noun]"（如 As a first-time buyer / As a pet owner）
-    - 受众/体质/使用对象相关形容词（elderly / sensitive / indoor / outdoor）
+    在绑定簇的 top_phrases / sample_reviews 中搜索以下模式，出现 ≥5 条且 eligible_physio_labels 达标时才可建对应生理 Persona：
+    - "I have [adj] [noun]"（如 I have sensitive skin）
+    - "My [noun] is/are [adj]"
+    - "As a [noun]"
+    - 若 eligible_physio_labels 为空或不包含对应维度，禁止在名称/core_pain 写生理标签

    ## 其他遗漏检查
-    - 长期使用/复购用户：含 after months / after a while / second bottle / bought again 的差评是否形成独立群体
+    - 长期使用/复购用户：含 after months / second bottle / bought again 的差评是否形成独立群体
    - 占比低（~5%）但痛点独特、无法被其他 Persona 代表的群体，仍须单独列出

-    ## 可绑定的聚类簇目录（cluster_ref 必须从中选择；每簇含 top_phrases + sample_reviews 最多 5 条原文）
+    ## 可绑定的聚类簇目录（cluster_refs 必须从中选择）
+    每簇含：semantic_label、suggested_dimension（A/B/C建议维度）、eligible_physio_labels（仅≥5条佐证的生理维度）、top_phrases、sample_reviews
    {{catalog_json}}

-    ## 生理标签硬规则（名称 + core_pain，违反则系统会剔除）
-    - 名称或 core_pain 中的生理/体质类中文标签（如孕妇、老年、儿童、敏感肌、粗硬发、疤痕等），绑定簇内须有 ≥5 条评论原文含对应英文词（catalog 中 physio_review_counts 字段，如 pregnancy: 8）
-    - top_phrases 仅出现 1–4 次不算数；不得以短语偶然出现代替评论条数门槛
-    - 无达标评论佐证时禁止写入该标签；不得凭品类常识或方法论示例脑补
-    - sample_reviews 为溯源原文，core_pain 只能归纳其中明确出现的内容，禁止推断未出现的生理状态
+    ## 生理标签硬规则（进 LLM 前已过滤，仅 eligible_physio_labels 中的维度可用于命名）
+    - catalog 仅展示 eligible_physio_labels（评论佐证≥5条）；未出现的生理标签禁止写入名称或 core_pain
+    - top_phrases 仅出现 1–4 次不算数；不得凭品类常识脑补
+    - sample_reviews 为溯源原文，core_pain 只能归纳其中明确出现的内容

    ## 补充聚类摘要
    audience_clusters: {{audience_clusters_json}}
@ -84,17 +88,22 @@ persona:
    global_positive: {{global_positive_json}}

    ## 输出JSON
-    {"personas":[{"name":"≤6中文字","dimension":"A","cluster_ref":{"stage":"1_audience","label":0},"keywords":["从绑定簇 top_phrases 复制"],"core_pain":"核心痛点（≤3条，分号分隔）","core_need":"核心需求（≤3条，分号分隔）","purchase_motivation":"雇佣产品做什么（动词+宾语，JTBD句式）"}]}
+    {"personas":[{"name":"≤6中文字","dimension":"A","cluster_refs":{"A":{"stage":"3b_aspect_opinion_negative","label":2},"B":{"stage":"3b_aspect_opinion_negative","label":5},"C":{"stage":"3a_pain_global","label":1}},"keywords":["从绑定簇 top_phrases 复制"],"core_pain":"核心痛点（≤3条，分号分隔）","core_need":"核心需求（≤3条，分号分隔）","purchase_motivation":"雇佣产品做什么（动词+宾语，JTBD句式）"}]}

    ## 硬性要求
-    - 每个 Persona 必须有 cluster_ref；stage 只能是：1_audience / 3a_pain_global / 3b_aspect_opinion_negative / 3b_aspect_opinion_positive
-    - cluster_ref.label 必须是上方目录中存在的 label；系统用该簇的 review_count 作为命中规模（非 keywords 扫全文）
-    - dimension 只能是 A、B 或 C；A 维 Persona 的 cluster_ref.stage 必须是 1_audience
-    - keywords 至少 5 个，必须从绑定簇的 top_phrases 复制英文片段（≥3字符），禁止编造不在簇中的词
-    - 4-7 个 Persona，三维度 A/B/C 均至少覆盖 1 个；优先选 review_count 较大的簇，小簇（<15条）仅在有明确短语证据时使用
+    - 每个 Persona 必须有 cluster_refs；A/B/C 三维均从 suggested_dimension 匹配的簇中选择
+    - A 维 stage 允许: 3a_pain_global / 3b_aspect_opinion_negative / 3b_aspect_opinion_positive（从 suggested_dimension="A" 的簇中选）
+    - B 维 stage 允许: 3a_pain_global / 3b_aspect_opinion_negative / 3b_aspect_opinion_positive（从 suggested_dimension="B" 的簇中选）
+    - C 维 stage 允许: 3a_pain_global / 3b_aspect_opinion_negative / 3b_aspect_opinion_positive（从 suggested_dimension="C" 的簇中选）
+    - 每个 Persona 必须至少绑 1 个 B 或 C 维簇（具体痛点/需求/场景），用于 keyword 过滤占比
+    - cluster_refs 各维 label 必须是上方 catalog 中存在的 label；绑定前先看 semantic_label 与 suggested_dimension
+    - 系统以 A 维锚点簇 + B/C 维 keywords 过滤计算占比
+    - dimension 表示该 Persona 主类型（A/B/C 之一）；cluster_refs.A 始终必填
+    - keywords 至少 5 个，必须从所有已绑定簇的 top_phrases 复制英文片段（≥3字符），禁止编造不在簇中的词
+    - 4-7 个 Persona，三维度 A/B/C 均至少覆盖 1 个；小簇（<15条）仅在有明确短语证据时使用
    - core_pain 来自该群体差评语义；core_need 来自该群体好评或诉求；purchase_motivation 用「雇佣产品完成…」句式
-    - 命名示例（格式参考，须从聚类归纳）：女性群体、敏感体质用户、首次购买用户、粗硬发质疤痕体质用户、旅行护理用户
-    - 禁止在绑定簇 physio_review_counts 未达 ≥5 条时将「孕妇/孕期」等生理标签写入名称或 core_pain
+    - 命名须从聚类归纳；生理类命名仅当 eligible_physio_labels 含对应维度时才允许
+    - 禁止在 eligible_physio_labels 未列出的生理标签写入名称或 core_pain

 # ── 2. 差评/好评主题 ──
 theme:
@ -304,7 +313,7 @@ rootcause:
    - ✅ 机制层：「密封/接口设计不足导致进水腐蚀」「关键部件角度/间距不当导致效果未达预期」

    {{persona_json}}
-    {{per_aud_data_json}}
+    {{bound_clusters_json}}

    ## 可用差评主题名（affected_themes 只能从中选择）
    {{theme_names_json}}
@ -316,7 +325,7 @@ rootcause:
    - quotes 字段不要输出（引用由系统从真实评论回填）
    - quote_keywords 每条根因 2-4 个**买家评论常见英文词/短语**（如 pull, dull, broke, charge, shower, waterproof, loud, trim, smooth, irritation, snag, waste），须与 mechanism 语义相关
    - quote_keywords 禁止生僻工程术语（如 O-ring、IPX7、DLC、magnetic charging、martensitic）；mechanism 可写工程细节，quote_keywords 必须像亚马逊买家口语
-    - 优先从 per_aud_data 聚类 top_phrases 中选取真实出现的英文片段
+    - 优先从 bound_clusters 聚类 top_phrases 中选取真实出现的英文片段
    - affected_themes 只能使用上方差评主题名，优先 P0/P1 主题
    - dev_direction 须针对 {{product_name}} 可改进点
    - 每个 Persona 最多 3 条 root_causes
--- a/voc_业务_2/report_utils.py
+++ b/voc_业务_2/report_utils.py
--- a/voc_业务_2/run_pipeline.py
+++ b/voc_业务_2/run_pipeline.py
@ -22,8 +22,8 @@

 内部调用等价于（以全流程为例）::

-  ../310py/bin/python ../main_voc分析.py --only-step 1 --input-dir "..." --product "..." --industry "..." --keep-db
-  ../310py/bin/python ../main_voc分析.py --only-step 2 --product "..." --industry "..." --keep-db
+  ../310py/bin/python ../main_voc分析.py --only-step 1 --input-dir "..." --product "..." --industry "..."
+  ../310py/bin/python ../main_voc分析.py --only-step 2 --product "..." --industry "..."
  ...（依此类推到 step 6）
 """
 from __future__ import annotations
@ -57,7 +57,6 @@ def run_step(python_bin: str, main_script: str, step: int, **kwargs) -> bool:
        "--only-step", str(step),
        "--product", kwargs.get("product", "亚马逊商品"),
        "--industry", kwargs.get("industry", "亚马逊电商"),
-        "--keep-db",
    ]
    if step == 1 and "input_dir" in kwargs:
        cmd.extend(["--input-dir", kwargs["input_dir"]])
--- a/voc_业务_2/提示词编辑稿.md
+++ b/voc_业务_2/提示词编辑稿.md
@ -114,6 +114,10 @@ Persona 命名用简洁中文（≤6字），避免营销化夸张名称。输
 ```
 ## 任务：基于聚类数据按三维度发现用户画像（Persona），4-7 个。

+## 分析产品（product_detect / config 已识别）
+产品：{{product_name}} | 行业：{{industry}}
+purchase_motivation 用「雇佣{{product_name}}完成…」句式；Persona 须与该产品使用场景一致。
+
 ## 三维度强制覆盖（缺一不可）
 A. 物理/生理/受众特征 → 必须绑定 stage=1_audience 的簇
   典型信号：特定体质/肤质/年龄/体型/使用对象（如 sensitive / elderly / for kids / large breed）
@ -162,6 +166,7 @@ global_positive: {{global_positive_json}}
 ```

 **占位符说明：**
+- `{{product_name}}` / `{{industry}}` — 产品识别结果（自动填入）
 - `{{catalog_json}}` — 聚类簇目录（自动填入）
 - `{{audience_clusters_json}}` 等 — 聚类摘要（自动填入）

--- a/向量化.py
+++ b/向量化.py
@ -1,5 +1,5 @@
 """
-从 voc_structured.sqlite 最新 job 展开 audience / pain_point / aspect_opinion
+从 voc_structured.sqlite 最新 job 展开 persona_signal / pain_point / aspect_opinion
 （与聚类.py 一致，不向量化单独的 aspect、opinion），使用本地 MLX 写入 voc_embeddings.sqlite。

 溯源：source_row 与结构化时一致（CSV 第 1 条数据行=1）；对应 merged_reviews_cleaned.csv
@ -59,11 +59,56 @@ def _resolve_embed_workers(explicit: int | None = None) -> int:
        logger.warning("本地 embedding 仅支持串行，--workers 已忽略（使用 1）")
    return 1
 ENTITY_TYPES = (
-    "audience",
+    "persona_signal",
    "pain_point",
    "aspect_opinion",
 )

+_GENERIC_PERSONA_SIGNALS = frozenset({
+    "self", "unknown", "user", "customer", "buyer", "myself", "the user", "consumer",
+})
+
+
+def _normalize_persona_signals(raw: Any) -> List[str]:
+    if raw is None:
+        items: List[Any] = []
+    elif isinstance(raw, list):
+        items = raw
+    elif isinstance(raw, str) and raw.strip():
+        items = [raw.strip()]
+    else:
+        items = []
+    out: List[str] = []
+    seen: set[str] = set()
+    for item in items:
+        text = str(item).strip()
+        if not text or text.lower() in _GENERIC_PERSONA_SIGNALS:
+            continue
+        key = text.lower()
+        if key in seen:
+            continue
+        seen.add(key)
+        out.append(text)
+    return out
+
+
+def _derive_persona_signals_fallback(
+    persona_signals: List[str],
+    pain_points: List[Any],
+) -> List[str]:
+    """无 persona_signals 时：用 pain_points 前 2 条作兜底。"""
+    if persona_signals:
+        return persona_signals
+    return [str(p).strip() for p in pain_points if str(p).strip()][:2]
+
+
+def _should_skip_extraction(
+    persona_signals: List[str],
+    pain_points: List[Any],
+    feedback: List[Any],
+) -> bool:
+    return not persona_signals and not pain_points and not feedback
+

@dataclass
 class EmbedTask:
@ -73,7 +118,7 @@ class EmbedTask:
    entity_type: str
    entity_index: int
    embed_text: str
-    audience: str
+    audience: str | None
    aspect: str | None
    opinion: str | None
    category: str | None
@ -151,11 +196,6 @@ def _init_embed_schema(conn: sqlite3.Connection) -> None:
    )


-def _should_skip_extraction(audience: str, pain_points: List[Any], feedback: List[Any]) -> bool:
-    aud = (audience or "").strip().lower()
-    return aud == "unknown" and not pain_points and not feedback
-
-
 def _build_tasks(
    job_id: int,
    rows: List[sqlite3.Row],
@ -171,7 +211,7 @@ def _build_tasks(
            logger.warning("跳过无效 JSON extraction_id=%s: %s", ext_id, e)
            continue

-        audience = str(data.get("audience", "unknown")).strip() or "unknown"
+        persona_signals = _normalize_persona_signals(data.get("persona_signals"))
        pain_points = data.get("pain_points") or []
        feedback = data.get("product_feedback") or []
        if not isinstance(pain_points, list):
@ -179,7 +219,7 @@ def _build_tasks(
        if not isinstance(feedback, list):
            feedback = []

-        if _should_skip_extraction(audience, pain_points, feedback):
+        if _should_skip_extraction(persona_signals, pain_points, feedback):
            continue

        content = content_map.get(source_row, "")
@ -189,17 +229,18 @@ def _build_tasks(
                source_row,
            )

-        aud_norm = audience.strip()
-        if aud_norm.lower() != "unknown":
+        signals = _derive_persona_signals_fallback(persona_signals, pain_points)
+
+        for i, sig in enumerate(signals):
            tasks.append(
                EmbedTask(
                    job_id=job_id,
                    extraction_id=ext_id,
                    source_row=source_row,
-                    entity_type="audience",
-                    entity_index=0,
-                    embed_text=aud_norm,
-                    audience=aud_norm,
+                    entity_type="persona_signal",
+                    entity_index=i,
+                    embed_text=sig,
+                    audience=None,
                    aspect=None,
                    opinion=None,
                    category=None,
@ -220,7 +261,7 @@ def _build_tasks(
                    entity_type="pain_point",
                    entity_index=i,
                    embed_text=text,
-                    audience=aud_norm,
+                    audience=None,
                    aspect=None,
                    opinion=None,
                    category=None,
@ -247,7 +288,7 @@ def _build_tasks(
                    entity_type="aspect_opinion",
                    entity_index=i,
                    embed_text=merged,
-                    audience=aud_norm,
+                    audience=None,
                    aspect=aspect,
                    opinion=opinion,
                    category=category,
--- a/结构化_server.py
+++ b/结构化_server.py
@ -588,19 +588,55 @@ def _normalize_product_feedback(items: Any) -> List[Dict[str, str]]:
    return out


+_GENERIC_PERSONA_SIGNALS = frozenset({
+    "self", "unknown", "user", "customer", "buyer", "myself", "the user", "consumer",
+})
+
+
+def _normalize_persona_signals(raw: Any) -> List[str]:
+    if raw is None:
+        items: List[Any] = []
+    elif isinstance(raw, list):
+        items = raw
+    elif isinstance(raw, str) and raw.strip():
+        items = [raw.strip()]
+    else:
+        items = []
+    out: List[str] = []
+    seen: set[str] = set()
+    for item in items:
+        text = str(item).strip()
+        if not text or text.lower() in _GENERIC_PERSONA_SIGNALS:
+            continue
+        key = text.lower()
+        if key in seen:
+            continue
+        seen.add(key)
+        out.append(text)
+    return out
+
+
 def _validate_extraction_strict(obj: Any, key_label: str = "") -> Dict[str, Any]:
    """校验结构化结果；失败时抛出带键名的 ValueError，供回传模型修正。"""
    prefix = f"{key_label}: " if key_label else ""
    if not isinstance(obj, dict):
        raise ValueError(f"{prefix}必须是 JSON 对象")

-    for field in ("audience", "pain_points", "product_feedback"):
+    for field in ("persona_signals", "pain_points", "product_feedback"):
        if field not in obj:
            raise ValueError(f"{prefix}缺少必填字段 {field}")

-    audience = obj.get("audience")
-    if not isinstance(audience, str) or not audience.strip():
-        raise ValueError(f"{prefix}audience 必须为非空字符串")
+    persona_signals = obj.get("persona_signals")
+    if not isinstance(persona_signals, list):
+        raise ValueError(f"{prefix}persona_signals 必须为数组")
+    for i, sig in enumerate(persona_signals):
+        if not isinstance(sig, str) or not str(sig).strip():
+            raise ValueError(f"{prefix}persona_signals[{i}] 必须为非空字符串")
+        if str(sig).strip().lower() in _GENERIC_PERSONA_SIGNALS:
+            raise ValueError(
+                f"{prefix}persona_signals[{i}] 禁止为泛化词 {sig!r}，"
+                "须写具体身份/体质/场景短语"
+            )

    pain_points = obj.get("pain_points")
    if not isinstance(pain_points, list):
@ -639,11 +675,7 @@ def _normalize_extraction(obj: Any) -> Dict[str, Any]:
    if not isinstance(obj, dict):
        raise ValueError("Each extraction must be a JSON object")

-    audience = obj.get("audience", "unknown")
-    if not isinstance(audience, str) or not str(audience).strip():
-        audience = "unknown"
-    else:
-        audience = str(audience).strip()
+    persona_signals = _normalize_persona_signals(obj.get("persona_signals", []))

    pain_points = obj.get("pain_points", [])
    if pain_points is None:
@ -656,14 +688,14 @@ def _normalize_extraction(obj: Any) -> Dict[str, Any]:

    missing = [
        f
-        for f in ("audience", "pain_points", "product_feedback")
+        for f in ("persona_signals", "pain_points", "product_feedback")
        if f not in obj
    ]
    if missing:
        logger.info("Normalized missing fields %s in extraction", missing)

    return {
-        "audience": audience,
+        "persona_signals": persona_signals,
        "pain_points": pain_points,
        "product_feedback": product_feedback,
    }
@ -677,7 +709,8 @@ def _build_batch_correction_message(keys: List[str], errors: List[str]) -> str:
        f"校验错误：\n{err_block}\n\n"
        "要求：\n"
        f"- 输出一个 JSON 对象，顶层键必须且仅能是：{keys_literal}\n"
-        "- 每个键的值必须包含 audience、pain_points、product_feedback\n"
+        "- 每个键的值必须包含 persona_signals、pain_points、product_feedback\n"
+        "- persona_signals 禁止仅写 self/user/unknown 等泛化词\n"
        "- product_feedback 每条须含 aspect、opinion、sentiment"
        "（仅 Positive/Negative/Neutral，禁止 Mixed/Ambiguous；褒贬交织选主倾向或拆条）、"
        "category（禁止 Value，性价比用 Price）\n"
--- a/聚类.py
+++ b/聚类.py
@ -3,8 +3,8 @@

 流程：
  1. audience（LLM 自动调参 n_neighbors）
-  2a. 前两 audience 簇各自独立：簇内 pain_point（LLM 自动调参）
-  2b. 前两 audience 簇各自独立：簇内 aspect_opinion 按 Positive/Negative/Neutral 分桶聚类
+  2a. 占比达阈值的 audience 簇各自独立：簇内 pain_point（LLM 自动调参）
+  2b. 占比达阈值的 audience 簇各自独立：簇内 aspect_opinion 按 Positive/Negative/Neutral 分桶聚类
  3a. 全量 pain_point（LLM 自动调参）
  3b. 全量 aspect_opinion 按 Positive/Negative/Neutral 分桶聚类（各自 LLM 自动调参）

@ -110,7 +110,10 @@ SAMPLE_RATIO = 0.6
 STEP2_N_NEIGHBORS = 8

 # 簇内去重评论数 / 本步骤参与聚类的去重评论总数 < 该比例则不写入库、不进入报告
-CLUSTER_MIN_REVIEW_RATIO = 0.10
+CLUSTER_MIN_REVIEW_RATIO = 0.05
+# step 2 入选 audience 簇：去重评论数 / 全量清洗评论总数 ≥ 该比例
+DEFAULT_AUDIENCE_COVERAGE_THRESHOLD = 0.01
+VOC_BUSINESS_CONFIG = PROJECT_ROOT / "voc_业务_2" / "config.yaml"


@dataclass
@ -1093,10 +1096,24 @@ def _save_tuning_logs(
        )


-def _top2_audience_clusters(
+def _config_threshold(key: str, default: float) -> float:
+    try:
+        import yaml
+        if VOC_BUSINESS_CONFIG.is_file():
+            with VOC_BUSINESS_CONFIG.open(encoding="utf-8") as f:
+                cfg = yaml.safe_load(f) or {}
+            val = float(cfg.get(key, default))
+            if 0 < val <= 1:
+                return val
+    except (TypeError, ValueError, OSError):
+        pass
+    return default
+
+
+def _audience_cluster_ranking(
    conn: sqlite3.Connection, run_id: int
-) -> Tuple[List[int], Dict[int, int]]:
-    """返回 (前两簇标签列表, source_row -> audience簇标签)。"""
+) -> Tuple[Dict[int, set[int]], Dict[int, int]]:
+    """返回 (cluster_label -> source_rows, source_row -> audience簇标签)。"""
    cur = conn.execute(
        """
        SELECT cluster_label, source_row
@ -1113,11 +1130,46 @@ def _top2_audience_clusters(
        sr = int(sr)
        row_to_cluster[sr] = lab
        cluster_rows.setdefault(lab, set()).add(sr)
-    ranked = sorted(
-        cluster_rows.items(), key=lambda x: len(x[1]), reverse=True
-    )
-    top2 = [lab for lab, _ in ranked[:2]]
-    return top2, row_to_cluster
+    return cluster_rows, row_to_cluster
+
+
+def _top2_audience_clusters(
+    conn: sqlite3.Connection, run_id: int
+) -> Tuple[List[int], Dict[int, int]]:
+    """兼容旧逻辑：返回前两大的 audience 簇。"""
+    cluster_rows, row_to_cluster = _audience_cluster_ranking(conn, run_id)
+    ranked = sorted(cluster_rows.items(), key=lambda x: len(x[1]), reverse=True)
+    return [lab for lab, _ in ranked[:2]], row_to_cluster
+
+
+def _eligible_audience_clusters(
+    conn: sqlite3.Connection,
+    run_id: int,
+    total_reviews: int,
+    threshold: float = DEFAULT_AUDIENCE_COVERAGE_THRESHOLD,
+) -> Tuple[List[int], Dict[int, int]]:
+    """返回 (占比达阈值的 audience 簇列表, source_row -> audience簇标签)。"""
+    cluster_rows, row_to_cluster = _audience_cluster_ranking(conn, run_id)
+    if total_reviews <= 0:
+        return [], row_to_cluster
+    ranked = sorted(cluster_rows.items(), key=lambda x: len(x[1]), reverse=True)
+    eligible = [
+        lab for lab, rows in ranked
+        if len(rows) / total_reviews >= threshold
+    ]
+    if len(eligible) > 4:
+        logger.warning(
+            "eligible audience 簇数量=%s（阈值=%.0f%%），可能过度拆分 step 2",
+            len(eligible),
+            threshold * 100,
+        )
+    if not eligible and ranked:
+        logger.warning(
+            "无 audience 簇达到 step2 阈值 %.0f%%（总评论 %s），跳过 step 2",
+            threshold * 100,
+            total_reviews,
+        )
+    return eligible, row_to_cluster


 def run_clustering(
@ -1128,12 +1180,14 @@ def run_clustering(
    cluster_db: Path = CLUSTER_DB,
    reset_db: bool = True,
 ) -> dict:
+    """仅运行 3a（全量 pain）和 3b（全量 aspect_opinion × 情感）聚类。
+    已移除步骤1 audience 和步骤2 per-audience 聚类。
+    """
    require_chat_api_key()

    econn = sqlite3.connect(embed_db)
    try:
        jid = _resolve_job_id(job_id, econn, structured_db)
-        audience_rows = _load_embed_rows(econn, jid, "audience")
        pain_rows = _load_embed_rows(econn, jid, "pain_point")
        ao_rows = _load_embed_rows(econn, jid, "aspect_opinion")
    finally:
@ -1153,50 +1207,6 @@ def run_clustering(
        total_reviews = _count_csv_reviews(csv_path)
        logger.info("清洗后评论总数: %s", total_reviews)

-        # --- 1 audience ---
-        labels1, meta1 = _cluster_stage(
-            audience_rows, stage="1_audience", use_llm_tune=True, client=client
-        )
-        filt1 = _save_assignments(
-            cconn,
-            run_id,
-            "1_audience",
-            audience_rows,
-            labels1,
-            n_neighbors=meta1.get("n_neighbors"),
-            outlier_participates=False,
-        )
-        meta1["cluster_filter"] = filt1
-        _save_stage_meta(cconn, run_id, "1_audience", meta1)
-        if meta1.get("tuning_log"):
-            _save_tuning_logs(cconn, run_id, "1_audience", meta1["tuning_log"])
-
-        top2, row_to_aud = _top2_audience_clusters(cconn, run_id)
-        meta_top2 = {
-            "top2_audience_clusters": top2,
-            "per_cluster_source_rows": {
-                str(lab): len(_source_rows_for_audience_cluster(row_to_aud, lab))
-                for lab in top2
-            },
-            "step2_mode": "per_audience_cluster_llm_auto_tune",
-        }
-        _save_stage_meta(cconn, run_id, "step2_filter", meta_top2)
-        logger.info("Audience 前两簇（将分别聚类）: %s", top2)
-
-        step2_counts: Dict[str, int] = {}
-        for aud_lab in top2:
-            step2_counts.update(
-                _cluster_step2_for_audience(
-                    cconn,
-                    run_id,
-                    audience_cluster=aud_lab,
-                    pain_rows=pain_rows,
-                    ao_rows=ao_rows,
-                    row_to_aud=row_to_aud,
-                    client=client,
-                )
-            )
-
        # --- 3a pain global ---
        labels3a, meta3a = _cluster_stage(
            pain_rows, stage="3a_pain_global", use_llm_tune=True, client=client
@ -1234,10 +1244,7 @@ def run_clustering(
            "total_reviews": total_reviews,
            "min_cluster_review_ratio": CLUSTER_MIN_REVIEW_RATIO,
            "cluster_db": str(cluster_db.resolve()),
-            "top2_audience_clusters": top2,
            "counts": {
-                "1_audience": len(audience_rows),
-                **step2_counts,
                "3a_pain_global": len(pain_rows),
                **step3b_counts,
            },