技术标完善版本V1.2

.deps_installed

@@ -0,0 +1 @@
+

.gitignore

@@ -0,0 +1,6 @@
+.env
+*.log
+__pycache__/
+*.pyc
+.venv/
+venv/

.idea/inspectionProfiles/profiles_settings.xml

@@ -0,0 +1,6 @@
+<component name="InspectionProjectProfileManager">
+  <settings>
+    <option name="USE_PROJECT_PROFILE" value="false" />
+    <version value="1.0" />
+  </settings>
+</component>

.idea/modules.xml

@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ProjectModuleManager">
+    <modules>
+      <module fileurl="file://$PROJECT_DIR$/.idea/tech-bid-manageV1.2_20260424.iml" filepath="$PROJECT_DIR$/.idea/tech-bid-manageV1.2_20260424.iml" />
+    </modules>
+  </component>
+</project>

.idea/tech-bid-manageV1.2_20260424.iml

@@ -0,0 +1,16 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<module type="PYTHON_MODULE" version="4">
+  <component name="NewModuleRootManager">
+    <content url="file://$MODULE_DIR$" />
+    <orderEntry type="inheritedJdk" />
+    <orderEntry type="sourceFolder" forTests="false" />
+  </component>
+  <component name="TemplatesService">
+    <option name="TEMPLATE_CONFIGURATION" value="Jinja2" />
+    <option name="TEMPLATE_FOLDERS">
+      <list>
+        <option value="$MODULE_DIR$/templates" />
+      </list>
+    </option>
+  </component>
+</module>

.idea/vcs.xml

@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="VcsDirectoryMappings">
+    <mapping directory="" vcs="Git" />
+  </component>
+</project>

README.md

@@ -0,0 +1,113 @@
+# 标伙伴 · AI 标书助手
+
+基于大模型的智能标书生成工具（单机版），支持解析招标文件、自动生成技术标书、导出 Word 文档。
+
+## 快速开始
+
+### 方式一：双击启动（Windows）
+
+直接双击 `start.bat`，首次运行会自动安装依赖。
+
+### 方式二：命令行启动
+
+```bash
+# 1. 安装依赖
+pip install -r requirements.txt
+
+# 2. 启动应用
+python app.py
+```
+
+浏览器访问 **http://localhost:5000**
+
+---
+
+## 配置 API Key
+
+首次使用前，点击右上角 ⚙️ 设置图标，选择模型提供商并填入 API Key：
+
+| 提供商 | 推荐模型 | 申请地址 |
+|--------|---------|---------|
+| 通义千问 | qwen-max | https://dashscope.aliyun.com/ |
+| DeepSeek | deepseek-chat (V3) | https://platform.deepseek.com/ |
+| OpenAI | gpt-4o | https://platform.openai.com/ |
+
+> **DeepSeek 说明**：deepseek-chat (V3) 性价比极高，推荐用于生产环境。  
+> 由于 DeepSeek 暂不提供 Embedding API，使用知识库功能时会自动回退到本地 sentence-transformers 模型（首次使用需下载约 90MB）。
+
+也可通过环境变量配置：
+
+```bash
+# 通义千问
+set QWEN_API_KEY=sk-xxxxxxxx
+set MODEL_PROVIDER=qwen
+
+# DeepSeek
+set DEEPSEEK_API_KEY=sk-xxxxxxxx
+set MODEL_PROVIDER=deepseek
+
+python app.py
+```
+
+---
+
+## 使用流程
+
+1. **新建项目** → 输入项目名称
+2. **上传招标文件** → 支持 PDF / DOC / DOCX
+3. **AI 解析** → 自动提取评分要求、资质条件、商务条款
+4. **生成大纲** → 按评分权重生成四级章节目录
+5. **生成内容** → 逐章节或一键全部生成
+6. **合规检查** → 对照招标要求检验覆盖情况
+7. **导出 Word** → 专业排版，直接使用
+
+---
+
+## 目录结构
+
+```
+autorfp/
+├── app.py              # Flask 主程序
+├── config.py           # 配置文件
+├── requirements.txt    # Python 依赖
+├── start.bat           # Windows 一键启动
+├── prompts/            # AI 提示词模板
+├── modules/            # 功能模块
+│   ├── parser.py       # 招标文件解析
+│   ├── generator.py    # 标书内容生成
+│   ├── checker.py      # 合规检查
+│   ├── exporter.py     # Word 导出
+│   └── knowledge.py    # 企业知识库
+├── utils/              # 工具函数
+│   ├── ai_client.py    # AI API 封装
+│   ├── file_utils.py   # 文件处理
+│   └── prompts.py      # 提示词加载
+├── templates/          # HTML 模板
+├── static/             # 静态资源
+└── data/               # 数据目录（自动创建）
+    ├── projects.db     # SQLite 数据库
+    ├── uploads/        # 上传的招标文件
+    ├── exports/        # 导出的标书
+    ├── knowledge/      # 知识库文件
+    └── chroma/         # 向量数据库
+```
+
+---
+
+## 企业知识库
+
+在项目页面切换到「知识库」标签，上传历史标书文件。  
+系统会自动将文件分块存入向量数据库，生成内容时自动检索相关片段，让 AI 更好地体现企业优势。
+
+---
+
+## 常见问题
+
+**Q: 解析速度很慢？**  
+A: 招标文件越长耗时越长，通常 30-120 秒。建议使用 qwen-max 或 gpt-4o。
+
+**Q: 内容生成失败？**  
+A: 检查 API Key 是否正确，以及账户余额是否充足。
+
+**Q: 导出的 Word 文件乱码？**  
+A: 请使用 Microsoft Word 2016 及以上版本打开。

bid_partner.spec

@@ -0,0 +1,118 @@
+# -*- mode: python ; coding: utf-8 -*-
+"""
+PyInstaller spec for 标伙伴 · AI标书助手
+Build: pyinstaller bid_partner.spec
+
+知识库改用 SQLite + 纯 Python 向量存储，已不依赖 ChromaDB，打包更小。
+"""
+import os
+from PyInstaller.utils.hooks import collect_all, collect_data_files
+
+block_cipher = None
+
+# ── Collect complex packages ─────────────────────────────────────────────────
+openai_datas,   openai_bins,   openai_hidden   = collect_all('openai')
+pydantic_datas, pydantic_bins, pydantic_hidden = collect_all('pydantic')
+
+# tiktoken data (BPE vocab files)
+tiktoken_datas = collect_data_files('tiktoken')
+
+a = Analysis(
+    ['launcher.py'],
+    pathex=['.'],
+    binaries=openai_bins + pydantic_bins,
+    datas=[
+        # ── App assets (read-only, go into _MEIPASS) ──
+        ('templates', 'templates'),
+        ('static',    'static'),
+        # ── Package data ──
+        *openai_datas,
+        *pydantic_datas,
+        *tiktoken_datas,
+    ],
+    hiddenimports=[
+        # Flask / Werkzeug
+        'flask', 'flask_cors', 'werkzeug', 'werkzeug.serving',
+        'werkzeug.routing', 'werkzeug.middleware.proxy_fix',
+        'jinja2', 'jinja2.ext',
+        # SQLite (stdlib, always present)
+        'sqlite3',
+        # OpenAI
+        *openai_hidden,
+        # Pydantic
+        *pydantic_hidden,
+        # Document processing
+        'PyPDF2', 'pypdf', 'pypdf.errors',
+        'pdfminer', 'pdfminer.high_level', 'pdfminer.layout',
+        'pdfminer.pdfpage', 'pdfminer.pdfinterp', 'pdfminer.converter',
+        'docx', 'docx.oxml', 'docx.oxml.ns', 'docx.shared',
+        'docx.enum', 'docx.enum.text', 'docx.enum.style',
+        'python_docx',
+        # tiktoken
+        'tiktoken', 'tiktoken.core', 'tiktoken.model',
+        'tiktoken_ext', 'tiktoken_ext.openai_public',
+        # Network / encoding
+        'requests', 'chardet', 'httpx', 'httpcore',
+        'anyio', 'anyio.streams', 'anyio.streams.memory',
+        'sniffio', 'certifi',
+        # Stdlib extras
+        'importlib.metadata', 'importlib.resources',
+        'pkg_resources', 'json', 'math', 'threading',
+        # Local project modules (explicitly include all)
+        'config', 'app',
+        'utils', 'utils.ai_client', 'utils.file_utils',
+        'utils.prompts', 'utils.settings', 'utils.boq_parser', 'utils.bill_analysis',
+        'modules', 'modules.parser', 'modules.generator',
+        'modules.checker', 'modules.exporter', 'modules.knowledge',
+    ],
+    hookspath=[],
+    hooksconfig={},
+    runtime_hooks=[],
+    excludes=[
+        # Heavy packages not used in this app
+        'matplotlib', 'pandas', 'scipy', 'numpy',
+        'IPython', 'jupyter', 'notebook',
+        'PIL', 'Pillow',
+        'cv2', 'torch', 'tensorflow',
+        'pytest', 'unittest',
+        # ChromaDB 及其依赖（已移除，改用 SQLite 内置存储）
+        'chromadb', 'hnswlib', 'posthog', 'pypika',
+        'mmh3', 'overrides', 'monotonic',
+        'sentence_transformers', 'onnxruntime',
+    ],
+    win_no_prefer_redirects=False,
+    win_private_assemblies=False,
+    cipher=block_cipher,
+    noarchive=False,
+)
+
+pyz = PYZ(a.pure, a.zipped_data, cipher=block_cipher)
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    [],
+    exclude_binaries=True,
+    name='bid_partner',
+    debug=False,
+    bootloader_ignore_signals=False,
+    strip=False,
+    upx=False,
+    console=False,       # no black console window — GUI launcher takes over
+    disable_windowed_traceback=False,
+    argv_emulation=False,
+    target_arch=None,
+    codesign_identity=None,
+    entitlements_file=None,
+)
+
+coll = COLLECT(
+    exe,
+    a.binaries,
+    a.zipfiles,
+    a.datas,
+    strip=False,
+    upx=False,
+    upx_exclude=[],
+    name='BidPartner',
+)

bill-worker.js

@@ -0,0 +1,672 @@
+/**
+ * bill-worker.js — PDF 清单解析调度器（Worker Thread）
+ *
+ * 架构（v3 — SharedArrayBuffer 零拷贝）：
+ *   Phase 1 — 并行文本提取
+ *     将 PDF 数据写入 SharedArrayBuffer（一次分配，所有子线程共享读）
+ *     启动 N 个 page-worker，每个负责固定 20 页
+ *
+ *   Phase 2 — 清单页筛选 + 文本解析（纯正则，毫秒级）
+ *     汇总全部页面文本 → 关键字筛选清单页 → 多行合并 → 逐行解析
+ */
+'use strict';
+const { parentPort } = require('worker_threads');
+const { Worker } = require('worker_threads');
+const path = require('path');
+
+const PAGES_PER_CHUNK = 20;
+
+parentPort.on('message', async (msg) => {
+    if (msg.type !== 'parse') return;
+    const t0 = Date.now();
+    try {
+        // 立即做一次干净的拷贝，确保拥有独立的 ArrayBuffer
+        const raw = msg.buffer;
+        const buf = Buffer.alloc(raw.byteLength);
+        Buffer.from(raw).copy(buf);
+
+        if (buf.length === 0) {
+            parentPort.postMessage({ type: 'done', ok: false, error: '收到空 PDF 数据' });
+            return;
+        }
+
+        // ── 获取总页数 ──
+        const pdfjsModule = await import('pdfjs-dist/build/pdf.mjs');
+        const pdfjsLib = pdfjsModule.default || pdfjsModule;
+        // 给 pdfjs 一份独立拷贝（pdfjs 内部可能 detach buffer）
+        const pdfData = new Uint8Array(buf.length);
+        buf.copy(Buffer.from(pdfData.buffer));
+        const pdf = await pdfjsLib.getDocument({ data: pdfData, isEvalSupported: false }).promise;
+        const totalPages = pdf.numPages;
+
+        // ── 将 PDF 数据写入 SharedArrayBuffer（一次分配，所有子线程共享读）──
+        const sab = new SharedArrayBuffer(buf.length);
+        const sabView = new Uint8Array(sab);
+        buf.copy(Buffer.from(sabView.buffer)); // 从独立 buf 拷贝到共享内存
+
+        const workerCount = Math.ceil(totalPages / PAGES_PER_CHUNK);
+        console.log(`[BillWorker] PDF ${totalPages} 页, ${workerCount} 路并行 (SharedArrayBuffer ${(buf.length/1024/1024).toFixed(1)}MB)`);
+
+        // Phase 1: 并行文本提取
+        const pageTexts = await parallelExtract(sab, buf.length, totalPages, workerCount);
+        const t1 = Date.now();
+
+        const extractedCount = pageTexts.filter(t => t.length > 0).length;
+        console.log(`[BillWorker] Phase1 完成: ${t1 - t0}ms, ${extractedCount}/${totalPages} 页有文本`);
+
+        // 扫描件判断
+        const totalChars = pageTexts.reduce((s, t) => s + t.length, 0);
+        if (totalChars < 50) {
+            parentPort.postMessage({ type: 'done', ok: true, data: { scanned: true, reason: 'noText', totalPages } });
+            return;
+        }
+
+        // Phase 2: 筛选清单页（宽松策略 + 连续页补全）
+        const BILL_KW = ['项目编码', '项目名称', '工程量', '计量单位', '综合单价', '清单编码'];
+        const SEC_KW  = ['分部分项', '分类分项', '措施项目', '其他项目', '工程量清单计价'];
+        // 第一轮：标记确定的清单页
+        const billFlags = new Array(pageTexts.length).fill(false);
+        for (let i = 0; i < pageTexts.length; i++) {
+            const t = pageTexts[i];
+            if (!t.trim()) continue;
+            const hHits = BILL_KW.filter(k => t.includes(k)).length;
+            const sHit  = SEC_KW.some(k => t.includes(k));
+            const hasCode = /\d{9}/.test(t);
+            // 放宽：有9位编码即可（不再要求同时命中表头关键字）
+            if (hHits >= 2 || sHit || hasCode) {
+                billFlags[i] = true;
+            }
+        }
+        // 第二轮：连续页补全 — 两个清单页之间的非空页也视为清单页（续页无表头）
+        // 但排除纯费用/税金页面（它们不含施工清单项）
+        const FEE_PAGE_KW = ['规费', '税金', '社会保险费', '住房公积金', '养老保险',
+                             '工伤保险', '失业保险', '医疗保险', '教育费附加', '城市维护建设税'];
+        const firstBill = billFlags.indexOf(true);
+        const lastBill  = billFlags.lastIndexOf(true);
+        if (firstBill >= 0 && lastBill > firstBill) {
+            for (let i = firstBill; i <= lastBill; i++) {
+                if (!billFlags[i] && pageTexts[i] && pageTexts[i].trim().length > 30) {
+                    const t = pageTexts[i];
+                    const feeHits = FEE_PAGE_KW.filter(kw => t.includes(kw)).length;
+                    // 命中 2+ 个费用关键字且没有9位工程编码 → 纯费用页，排除
+                    if (feeHits >= 2 && !/\d{9}/.test(t)) continue;
+                    billFlags[i] = true;
+                }
+            }
+        }
+        const billTexts = [];
+        for (let i = 0; i < pageTexts.length; i++) {
+            if (billFlags[i]) billTexts.push(pageTexts[i]);
+        }
+
+        if (!billTexts.length) {
+            parentPort.postMessage({ type: 'done', ok: true, data: { scanned: false, noBillPages: true, totalPages } });
+            return;
+        }
+
+        console.log(`[BillWorker] ${totalPages} 页 → ${billTexts.length} 页清单 (原始识别 ${billFlags.filter(f=>f).length - (lastBill - firstBill >= 0 ? 0 : 0)} / 补全后 ${billTexts.length})`);
+
+        // Phase 3: 文本解析
+        const merged = billTexts.join('\n');
+        const parsed = parseBillText(merged);
+        const t2 = Date.now();
+        console.log(`[BillWorker] Phase2+3: ${t2 - t1}ms, 总耗时: ${t2 - t0}ms`);
+
+        parentPort.postMessage({
+            type: 'done', ok: true,
+            data: {
+                scanned: false,
+                ...parsed,
+                _meta: {
+                    method: 'local-parallel',
+                    workers: workerCount,
+                    billPages: billTexts.length,
+                    totalPages,
+                    extractMs: t1 - t0,
+                    parseMs: t2 - t1,
+                    totalMs: t2 - t0,
+                }
+            }
+        });
+    } catch (err) {
+        console.error('[BillWorker] 错误:', err.message);
+        parentPort.postMessage({ type: 'done', ok: false, error: err.message });
+    }
+});
+
+// ================================================================
+// Phase 1: 多 Worker 并行提取（SharedArrayBuffer 零拷贝）
+// ================================================================
+
+function parallelExtract(sab, dataLength, totalPages, workerCount) {
+    return new Promise((resolve) => {
+        const workerPath = path.join(__dirname, 'page-worker.js');
+        const allPageTexts = new Array(totalPages).fill('');
+        const workerStatus = new Array(workerCount).fill('pending'); // pending, done, failed
+        let resolved = false;
+
+        const checkComplete = () => {
+            if (resolved) return;
+            const doneCount = workerStatus.filter(s => s === 'done' || s === 'failed').length;
+            if (doneCount >= workerCount) {
+                resolved = true;
+                // 检查是否有失败的worker，打印警告
+                const failedCount = workerStatus.filter(s => s === 'failed').length;
+                if (failedCount > 0) {
+                    console.warn(`[BillWorker] ${failedCount}/${workerCount} 个worker失败，可能导致部分页面无内容`);
+                }
+                resolve(allPageTexts);
+            }
+        };
+
+        for (let i = 0; i < workerCount; i++) {
+            const startPage = i * PAGES_PER_CHUNK + 1;
+            const endPage = Math.min((i + 1) * PAGES_PER_CHUNK, totalPages);
+
+            // workerData 传 SharedArrayBuffer（跨线程共享，不会被清空）
+            const w = new Worker(workerPath, {
+                workerData: { sab, dataLength, startPage, endPage }
+            });
+
+            let workerDone = false;
+
+            const markDone = (status) => {
+                if (workerDone) return;
+                workerDone = true;
+                workerStatus[i] = status;
+                checkComplete();
+            };
+
+            w.on('message', (msg) => {
+                if (msg.ok && msg.results) {
+                    for (const r of msg.results) {
+                        allPageTexts[r.page - 1] = r.text;
+                    }
+                    markDone('done');
+                } else if (!msg.ok) {
+                    console.warn(`[BillWorker] page-worker[${startPage}-${endPage}] 失败: ${msg.error}`);
+                    markDone('failed');
+                }
+            });
+
+            w.on('error', (err) => {
+                console.warn(`[BillWorker] page-worker[${startPage}-${endPage}] 异常: ${err.message}`);
+                markDone('failed');
+            });
+
+            w.on('exit', (code) => {
+                // exit 在 message 之后触发，但如果 worker 崩溃没发 message 则在这里兜底
+                if (code !== 0 && !workerDone) {
+                    console.warn(`[BillWorker] page-worker[${startPage}-${endPage}] 意外退出(code=${code})`);
+                    markDone('failed');
+                } else if (!workerDone) {
+                    markDone('done');
+                }
+            });
+        }
+
+        if (workerCount <= 0) {
+            resolved = true;
+            resolve(allPageTexts);
+        }
+    });
+}
+
+// ================================================================
+// Phase 3: 清单文本解析（纯正则 + 字符串处理，毫秒级）
+// ================================================================
+
+function parseBillText(text) {
+    const rawLines = text.split(/\n/).map(l => {
+        let line = l.replace(/\t/g, ' ').trim();
+        // 规范化带横杠的编码：如 "010-101-001-001" → "010101001001"
+        line = line.replace(/(\d{2,4})[-‐–](\d{2,4})[-‐–](\d{2,4})(?:[-‐–](\d{2,4}))?/g,
+            (m, a, b, c, d) => {
+                const combined = a + b + c + (d || '');
+                return (combined.length >= 9 && combined.length <= 12) ? combined : m;
+            });
+        return line;
+    });
+
+    // ── Step 1: 多行合并成逻辑行 ──
+    // pdfjs 按 Y 坐标分行，表格一行通常 = 一条文本行
+    // 但有时 项目特征/名称 会折行，需要合并
+    //
+    // 新逻辑行的起始标志（任一命中即切断）：
+    //   a) 序号模式：1.1.1.1.5 开头
+    //   b) 清单编码：9-12位数字 或 B+5-6位数字 开头
+    //   c) 中文大标题：一 二 三 ... 或 （一）（二）...
+    //   d) 表头行内容（跳过）
+    //   e) 纯数字序号 + 空格 + 编码（如 "5 500101004001"）
+
+    const ITEM_START    = /^\d+(\.\d+)+\s/;              // 1.1 或 1.1.1 等序号
+    const CODE_INLINE   = /(?:^|\s)(\d{9,12}|(?<![A-Za-z])B\d{5,6})\s/; // 行内含清单编码（排除 GB/DB 等标准号）
+    const CODE_START_RE = /^(\d{9,12}|B\d{5,6})\s/;     // 行首就是清单编码（行首 B 不会有前缀字母）
+    const SEQ_CODE_RE   = /^\d{1,4}\s+(\d{9,12}|(?<![A-Za-z])B\d{5,6})\s/; // "序号 编码"格式
+    const PAGE_MARK     = /^--\s*\d+\s+of\s+\d+\s*--/;
+    const HEADER_RE     = /^序号\s+(项目编码|项目名称)/;
+    const HEADER_KW     = /^(项目编码|项目名称|清单编码|计量单位|综合单价|工程量|合\s*价|金额|序号)\s/;
+    const CATEGORY_MARKERS = ['一', '二', '三', '四', '五', '六', '七', '八', '九', '十',
+                              '（一）', '（二）', '（三）', '（四）', '（五）'];
+
+    const logicLines = [];
+    let currentLine = '';
+
+    function isNewLineTrigger(raw) {
+        if (ITEM_START.test(raw)) return true;
+        if (CODE_START_RE.test(raw)) return true;
+        if (SEQ_CODE_RE.test(raw)) return true;
+        if (CATEGORY_MARKERS.some(m => raw.startsWith(m + ' ') || raw.startsWith(m + '\u3000'))) return true;
+        return false;
+    }
+
+    for (const raw of rawLines) {
+        if (!raw || PAGE_MARK.test(raw)) continue;
+        if (HEADER_RE.test(raw) || HEADER_KW.test(raw)) continue;
+        if (/^（元）|^款章节号|^备注$|^第\d+页/.test(raw)) continue;
+
+        if (isNewLineTrigger(raw)) {
+            if (currentLine) logicLines.push(currentLine);
+            currentLine = raw;
+        } else if (CODE_INLINE.test(raw) && raw.length > 15) {
+            // 行内包含编码且够长（像是完整的表格行）→ 也开新行
+            if (currentLine) logicLines.push(currentLine);
+            currentLine = raw;
+        } else {
+            // 续行（项目特征折行等短文本）
+            // 安全阀：已合并行过长时强制切断，防止整页吞并
+            if (currentLine && currentLine.length > 300) {
+                logicLines.push(currentLine);
+                currentLine = raw;
+            } else {
+                currentLine = currentLine ? currentLine + ' ' + raw : raw;
+            }
+        }
+    }
+    if (currentLine) logicLines.push(currentLine);
+
+    console.log(`[BillWorker] 合并后 ${logicLines.length} 条逻辑行（原始 ${rawLines.length} 行）`);
+    // 打印前5条逻辑行供调试
+    for (let i = 0; i < Math.min(5, logicLines.length); i++) {
+        console.log(`[BillWorker]   L${i}: ${logicLines[i].substring(0, 120)}`);
+    }
+
+    const categories = [];
+    let curCat = null, curItem = null;
+
+    // 编码匹配：支持行内任意位置的9-12位数字或B编码（排除 GB/DB 等标准号前缀）
+    const CODE_RE = /(?<![A-Za-z])(\d{9,12}|(?<![A-Za-z])B\d{5,6})/;
+    const UNIT_TOKENS = ['m³','m²','m3','m2','km','hm2','㎡','㎥','t','kg',
+                         '个','台','套','组','根','块','片','张','只','吨','项',
+                         '处','座','件','段','条','把','扇','口','圈','道','孔',
+                         '对','副','樘','方','延m','株','棵','m'];
+    const UNIT_SET = new Set(UNIT_TOKENS);
+    const unitEscaped = UNIT_TOKENS.map(u => u.replace(/[.*+?^${}()|[\]\\]/g, '\\$&'));
+    const UNIT_RE = new RegExp(`(?:^|\\s)(${unitEscaped.join('|')})(?=\\s|\\d|$)`);
+    const SKIP_RE = /合\s*计|小\s*计|本页小计|总\s*计|价税合计/;
+
+    for (const line of logicLines) {
+        if (SKIP_RE.test(line)) continue;
+
+        // 去掉行首的序号部分（"1.1.1.1.5 " 或 "5 " 等纯序号前缀）
+        let stripped = line.replace(/^\d+(\.\d+)*\s+/, '').trim();
+        if (!stripped) stripped = line.trim();
+        if (!stripped) continue;
+
+        const cm = stripped.match(CODE_RE);
+        if (cm) {
+            if (curItem && curCat) curCat.items.push(curItem);
+            if (!curCat) { curCat = { name: '未分类', items: [] }; categories.push(curCat); }
+
+            const code = cm[1];
+            let rest = stripped.substring(cm.index + cm[0].length).trim();
+            let name = '', unit = '', quantity = '', spec = '';
+
+            const unitMatch = rest.match(UNIT_RE);
+            if (unitMatch) {
+                const ui = rest.indexOf(unitMatch[0]);
+                let rawName = rest.substring(0, ui).trim();
+                unit = unitMatch[1];
+                const afterUnit = rest.substring(ui + unitMatch[0].length).trim();
+                const qm = afterUnit.match(/^([\d,.]+)/);
+                if (qm) {
+                    quantity = qm[1];
+                    // 提取 quantity 之后的尾部文本，跳过纯数字字段（综合单价、合价等）
+                    let tail = afterUnit.substring(qm.index + qm[0].length).trim();
+                    if (tail) {
+                        const tailTokens = tail.split(/\s+/);
+                        let si = 0;
+                        while (si < tailTokens.length && /^[\d,.%\-]+$/.test(tailTokens[si])) si++;
+                        const specTail = tailTokens.slice(si).join(' ').trim();
+                        if (specTail) spec = specTail;
+                    }
+                }
+                // 分离 rawName 中的"项目名称"和内联"项目特征"
+                const ns = splitNameAndSpec(rawName);
+                name = ns.name;
+                if (ns.spec) spec = ns.spec + (spec ? ';' + spec : '');
+            } else {
+                const tokens = rest.split(/\s+/).filter(t => t);
+                let foundUnitIdx = -1;
+                for (let ti = tokens.length - 1; ti >= 1; ti--) {
+                    if (UNIT_SET.has(tokens[ti])) { foundUnitIdx = ti; break; }
+                }
+                if (foundUnitIdx >= 1) {
+                    const rawNameStr = tokens.slice(0, foundUnitIdx).join(' ');
+                    const ns = splitNameAndSpec(rawNameStr);
+                    name = ns.name;
+                    if (ns.spec) spec = ns.spec;
+                    unit = tokens[foundUnitIdx];
+                    const afterTokens = tokens.slice(foundUnitIdx + 1);
+                    if (afterTokens.length && /^[\d,.]+$/.test(afterTokens[0])) {
+                        quantity = afterTokens[0];
+                        let si = 1;
+                        while (si < afterTokens.length && /^[\d,.%\-]+$/.test(afterTokens[si])) si++;
+                        const specTail = afterTokens.slice(si).join(' ').trim();
+                        if (specTail) spec = spec ? spec + ';' + specTail : specTail;
+                    }
+                } else {
+                    name = rest;
+                }
+            }
+
+            name = name.replace(/\s+/g, '').trim();
+            for (const u of UNIT_TOKENS) {
+                if (name.endsWith(u) && name.length > u.length) {
+                    unit = unit || u;
+                    name = name.substring(0, name.length - u.length);
+                    break;
+                }
+            }
+
+            curItem = { code, name, unit, quantity, spec };
+            continue;
+        }
+
+        // ── 回退：无标准编码但有 "名称 单位 数量" 结构 → 也视为清单项 ──
+        // 常见于措施项目、未编码的补充清单项
+        if (!cm && stripped.length > 4) {
+            const uniMatch = stripped.match(UNIT_RE);
+            if (uniMatch) {
+                const ui = stripped.indexOf(uniMatch[0]);
+                const beforeUnit = stripped.substring(0, ui).trim();
+                const afterUnit  = stripped.substring(ui + uniMatch[0].length).trim();
+                const hasQty = /^[\d,.]+/.test(afterUnit);
+                // 名称 2-50 字、含中文、有数量、不是分部标题
+                if (beforeUnit.length >= 2 && beforeUnit.length <= 50 && hasQty
+                    && /[\u4e00-\u9fff]/.test(beforeUnit)) {
+                    if (curItem && curCat) curCat.items.push(curItem);
+                    if (!curCat) { curCat = { name: '未分类', items: [] }; categories.push(curCat); }
+                    const unit = uniMatch[1];
+                    const qm = afterUnit.match(/^([\d,.]+)/);
+                    const quantity = qm ? qm[1] : '';
+                    const ns = splitNameAndSpec(beforeUnit);
+                    const name = ns.name.replace(/\s+/g, '').trim();
+                    const spec = ns.spec || '';
+                    curItem = { code: '', name, unit, quantity, spec };
+                    continue;
+                }
+            }
+        }
+
+        // 分部标题判断：不含编码、较短的文本、含工程关键字
+        // 关键守卫：如果行里有计量单位，说明是清单项，不是标题
+        if (stripped.length > 2 && stripped.length < 60 && !CODE_RE.test(stripped)) {
+            if (UNIT_RE.test(stripped) && /\d+\.?\d*\s*$/.test(stripped)) {
+                if (curItem) curItem.spec = curItem.spec ? curItem.spec + ';' + stripped : stripped;
+                continue;
+            }
+            if (isCatTitle(stripped) && !UNIT_RE.test(stripped) && !isFeeCatTitle(stripped)) {
+                if (curItem && curCat) { curCat.items.push(curItem); curItem = null; }
+                const cleanTitle = stripped.replace(/\s+(座|个|项|处|m|km|段|条)\s+\d+[\d.]*\s*$/, '').trim();
+                curCat = { name: cleanTitle, items: [] };
+                categories.push(curCat);
+                continue;
+            }
+        }
+
+        if (/^[一二三四五六七八九十]+\s/.test(stripped) || /^（[一二三四五六七八九十\d]+）/.test(stripped)) {
+            // 中文序号标题也需要排除费用类
+            const cleanTitle = stripped.replace(/\s+(座|个|项|处)\s+\d+[\d.]*\s*$/, '').trim();
+            if (isFeeCatTitle(cleanTitle)) {
+                // 费用类标题：跳过，不建分部（其下的行会作为续行处理）
+                continue;
+            }
+            if (curItem && curCat) { curCat.items.push(curItem); curItem = null; }
+            curCat = { name: cleanTitle, items: [] };
+            categories.push(curCat);
+            continue;
+        }
+
+        if (curItem && stripped.length > 1) {
+            curItem.spec = curItem.spec ? curItem.spec + ';' + stripped : stripped;
+        }
+    }
+
+    if (curItem && curCat) curCat.items.push(curItem);
+
+    // 过滤费用项：只保留需要写入技术标的施工清单项
+    let feeFiltered = 0;
+    for (const cat of categories) {
+        if (cat.items) {
+            const before = cat.items.length;
+            cat.items = cat.items.filter(it => !isFeeItem(it.name));
+            feeFiltered += before - cat.items.length;
+        }
+    }
+    if (feeFiltered > 0) console.log(`[BillWorker] 费用项过滤: 移除 ${feeFiltered} 项`);
+
+    // ========== 按项目名称合并（核心去重，大幅减少清单项数量）==========
+    // 规则：同一分部内，name 相同的清单项合并为一条
+    //   - code: 保留第一个非空编码
+    //   - unit: 保留第一个非空单位
+    //   - quantity: 尝试数值求和，否则用分号拼接
+    //   - spec: 去重后用分号拼接（截断过长的）
+    let totalBeforeMerge = 0, totalAfterMerge = 0;
+    for (const cat of categories) {
+        if (!cat.items || !cat.items.length) continue;
+        totalBeforeMerge += cat.items.length;
+
+        const nameMap = new Map(); // name → merged item
+        for (const item of cat.items) {
+            const key = (item.name || '').replace(/\s+/g, '').trim();
+            if (!key) continue;
+
+            if (!nameMap.has(key)) {
+                nameMap.set(key, {
+                    code: item.code || '',
+                    name: item.name,
+                    unit: item.unit || '',
+                    quantity: item.quantity || '',
+                    spec: item.spec || '',
+                    _count: 1,
+                    _quantities: item.quantity ? [item.quantity] : [],
+                    _specs: item.spec ? [item.spec] : [],
+                });
+            } else {
+                const m = nameMap.get(key);
+                m._count++;
+                // code: 取第一个非空的
+                if (!m.code && item.code) m.code = item.code;
+                // unit: 取第一个非空的
+                if (!m.unit && item.unit) m.unit = item.unit;
+                // quantity: 收集所有
+                if (item.quantity) m._quantities.push(item.quantity);
+                // spec: 收集不重复的
+                if (item.spec && !m._specs.includes(item.spec)) {
+                    m._specs.push(item.spec);
+                }
+            }
+        }
+
+        // 后处理：合成最终字段
+        const merged = [];
+        for (const [, m] of nameMap) {
+            // quantity: 尝试数值求和
+            if (m._quantities.length > 1) {
+                const nums = m._quantities.map(q => parseFloat(q.replace(/,/g, '')));
+                if (nums.every(n => !isNaN(n))) {
+                    const sum = nums.reduce((a, b) => a + b, 0);
+                    m.quantity = sum % 1 === 0 ? String(sum) : sum.toFixed(2);
+                } else {
+                    m.quantity = m._quantities.join('; ');
+                }
+            } else if (m._quantities.length === 1) {
+                m.quantity = m._quantities[0];
+            }
+            // spec: 拼接去重后的 spec，每条最多120字
+            if (m._specs.length > 0) {
+                const trimmed = m._specs.map(s => s.length > 120 ? s.substring(0, 120) + '...' : s);
+                m.spec = trimmed.join('; ');
+                // 总 spec 上限 300 字
+                if (m.spec.length > 300) m.spec = m.spec.substring(0, 300) + '...';
+            }
+            // 清理临时字段
+            delete m._count; delete m._quantities; delete m._specs;
+            merged.push(m);
+        }
+        cat.items = merged;
+        totalAfterMerge += merged.length;
+    }
+
+    const mergedCount = totalBeforeMerge - totalAfterMerge;
+    if (mergedCount > 0) {
+        console.log(`[BillWorker] 按名称合并: ${totalBeforeMerge} → ${totalAfterMerge} 项（合并 ${mergedCount} 个重复项）`);
+    }
+
+    const valid = categories.filter(c => c.items && c.items.length > 0);
+    const totalItems = valid.reduce((s, c) => s + c.items.length, 0);
+    const withSpec = valid.reduce((s, c) => s + c.items.filter(it => it.spec).length, 0);
+    const withCode = valid.reduce((s, c) => s + c.items.filter(it => it.code).length, 0);
+    console.log(`[BillWorker] 最终结果: ${valid.length} 分部, ${totalItems} 清单项 (${withCode} 有编码, ${withSpec} 有spec)`);
+    // 打印前 3 个 item 供调试
+    let debugCount = 0;
+    for (const cat of valid) {
+        for (const it of cat.items) {
+            if (debugCount < 3) {
+                console.log(`[BillWorker]   样例: [${it.code}] ${it.name} | ${it.unit} | qty=${it.quantity} | spec=${(it.spec||'').substring(0, 80)}`);
+                debugCount++;
+            }
+        }
+    }
+
+    return {
+        project_summary: { remark: `本地解析：${valid.length} 个分部，${totalItems} 个清单项（合并前 ${totalBeforeMerge} 项）` },
+        categories: valid,
+    };
+}
+
+/**
+ * 判断清单项是否为"费用项"（非施工内容，不写入技术标）
+ * 如：安全文明措施费、规费、税金、暂列金额等
+ */
+function isFeeItem(name) {
+    if (!name) return false;
+    const n = name.replace(/\s+/g, '');
+
+    // ── 1. 精确匹配 ──
+    const EXACT = [
+        '规费', '税金', '利润', '增值税', '暂列金额', '暂估价', '计日工',
+        '总承包服务费', '企业管理费', '甲供材料保管费', '价税合计',
+    ];
+    if (EXACT.includes(n)) return true;
+
+    // ── 2. 包含匹配：措施费/规费/保险/行政类 ──
+    const FEE_KW = [
+        '安全文明', '文明施工费', '环境保护费', '临时设施费',
+        '夜间施工增加费', '夜间施工费',
+        '冬雨季施工增加费', '冬雨季施工费',
+        '二次搬运费', '大型机械设备进出场', '大型机械进出场',
+        '施工排水降水', '排水降水费',
+        '已完工程及设备保护', '已完工程保护费',
+        '工程排污费', '社会保障费', '住房公积金',
+        '工伤保险', '劳动保险', '意外伤害保险', '建筑工程保险',
+        '城市维护建设税', '城市建设维护税',
+        '教育费附加', '地方教育附加',
+        '材料暂估', '专业工程暂估',
+        '超高施工增加费', '安全防护费',
+        '措施项目费', '其他项目费', '不可竞争费',
+    ];
+    for (const kw of FEE_KW) {
+        if (n.includes(kw)) return true;
+    }
+
+    return false;
+}
+
+/**
+ * 将 rawName 中的"项目名称"与内联"项目特征描述"分离
+ * 例: "土方开挖 1.土壤类别：普通土" → { name: "土方开挖", spec: "1.土壤类别：普通土" }
+ */
+function splitNameAndSpec(rawName) {
+    if (!rawName) return { name: '', spec: '' };
+    // Pattern 1: 数字+点+中文（如 "1.土壤类别" "2、强度等级"）
+    const m = rawName.match(/\d+[.、．)\uFF09]\s*[\u4e00-\u9fff]/);
+    if (m && m.index > 0) {
+        return {
+            name: rawName.substring(0, m.index).trim(),
+            spec: rawName.substring(m.index).trim()
+        };
+    }
+    // Pattern 2: 特征关键字+冒号（如 "材质：" "规格："）
+    const SPEC_KW_RE = /(材质|规格|型号|品牌|颜色|尺寸|厚度|直径|管径|强度|等级|类别|类型|做法|要求|标准|内容|工作内容|土壤|含量|配合比|工艺|方式|形式|范围|部位|位置|高度|宽度|长度|深度|坡度|截面|跨度|运距|开挖|回填|混凝土|钢筋|压实)[：:]/;
+    const kw = rawName.match(SPEC_KW_RE);
+    if (kw && kw.index > 0) {
+        return {
+            name: rawName.substring(0, kw.index).trim(),
+            spec: rawName.substring(kw.index).trim()
+        };
+    }
+    // Pattern 3: 括号开头的特征描述 "（1）" "(1)"
+    const paren = rawName.match(/[（(]\d+[）)]/);
+    if (paren && paren.index > 0) {
+        return {
+            name: rawName.substring(0, paren.index).trim(),
+            spec: rawName.substring(paren.index).trim()
+        };
+    }
+    return { name: rawName, spec: '' };
+}
+
+function isCatTitle(text) {
+    const KW = [
+        '土建','建筑','结构','装饰','装修','安装','给排水','暖通','空调','通风',
+        '电气','强电','弱电','消防','智能化','幕墙','门窗','园林','绿化','景观',
+        '市政','道路','桥梁','管网','基础','地基','桩基','主体','屋面','防水',
+        '保温','钢结构','排水','给水','照明','动力','防雷','电梯','人防','室外',
+        '附属','分部','工程','措施','清单','土石方','混凝土','砌筑','模板','脚手架',
+        '水利','河道','管道','阀门','设备','仪表','自动化','通信','网络',
+        '拆除','外墙','内墙','楼地面','天棚','吊顶','栏杆','屋顶','涂料','抹灰',
+        '廊道','阀门井','蓄水池','泵站','供水','引水','水源','渠道','闸门',
+        '围栏','警示','检修','管线','配电','水池','水塔','取水','净水',
+    ];
+    return KW.some(k => text.includes(k));
+}
+
+/**
+ * 判断分部标题是否为"费用类"（不应创建分部分类）
+ * 如：规费、税金、措施项目费、其他项目费 等非施工类分部
+ */
+function isFeeCatTitle(text) {
+    if (!text) return false;
+    const t = text.replace(/\s+/g, '');
+    // 精确匹配整个标题
+    const EXACT = [
+        '规费', '税金', '利润', '增值税', '暂列金额', '暂估价', '计日工',
+        '总承包服务费', '企业管理费', '价税合计',
+        '措施项目费', '其他项目费', '不可竞争费',
+    ];
+    if (EXACT.includes(t)) return true;
+    // 包含匹配
+    const FEE_CAT_KW = [
+        '措施项目费', '其他项目费', '不可竞争费',
+        '规费汇总', '税金汇总', '费率', '费用汇总', '费用合计',
+        '暂列金额', '暂估价', '计日工', '总承包服务费',
+        '安全文明施工费', '社会保障费', '住房公积金',
+        '工伤保险', '教育费附加', '城市维护建设税',
+    ];
+    for (const kw of FEE_CAT_KW) {
+        if (t.includes(kw)) return true;
+    }
+    return false;
+}

build.bat

@@ -0,0 +1,95 @@
+@echo off
+chcp 65001 >nul 2>&1
+setlocal
+
+echo ============================================================
+echo   BidPartner - Build Desktop EXE
+echo ============================================================
+echo.
+
+:: ── 1. Check Python ────────────────────────────────────────────────────────
+python --version >nul 2>&1
+if errorlevel 1 (
+    echo [ERROR] Python not found. Please install Python 3.9+.
+    pause & exit /b 1
+)
+
+:: ── 2. Install / upgrade PyInstaller ───────────────────────────────────────
+echo [Step 1/4] Installing PyInstaller...
+pip install --quiet --upgrade pyinstaller
+if errorlevel 1 (
+    echo [ERROR] Failed to install PyInstaller.
+    pause & exit /b 1
+)
+
+:: ── 3. Install project dependencies (if not already installed) ─────────────
+echo [Step 2/4] Checking dependencies...
+pip install --quiet -r requirements.txt
+if errorlevel 1 (
+    echo [ERROR] Failed to install dependencies.
+    pause & exit /b 1
+)
+
+:: ── 4. Sanitize settings.json - REMOVE API KEYS before build ───────────────
+echo [Step 3/4] Sanitizing settings (removing API keys from build)...
+if exist "data\settings.json" (
+    :: Back up real settings
+    copy /y "data\settings.json" "data\settings.json.bak" >nul
+)
+:: Write a clean settings file with no real keys
+(
+    echo {
+    echo   "model_provider": "deepseek",
+    echo   "qwen_api_key": "sk-your-qwen-key",
+    echo   "qwen_model": "qwen3.6-plus",
+    echo   "openai_api_key": "sk-your-openai-key",
+    echo   "openai_model": "gpt-4o",
+    echo   "deepseek_api_key": "sk-your-deepseek-key",
+    echo   "deepseek_model": "deepseek-chat",
+    echo   "max_concurrent": 5,
+    echo   "content_volume": "standard"
+    echo }
+) > "data\settings_clean.tmp"
+
+:: ── 5. Build ────────────────────────────────────────────────────────────────
+echo [Step 4/4] Building EXE with PyInstaller...
+echo   (This may take 3-10 minutes on first run)
+echo.
+
+:: Clean previous build artifacts
+if exist "build" rd /s /q "build" >nul 2>&1
+if exist "dist\BidPartner" rd /s /q "dist\BidPartner" >nul 2>&1
+
+pyinstaller bid_partner.spec --noconfirm
+set BUILD_RESULT=%errorlevel%
+
+:: ── Restore real settings ───────────────────────────────────────────────────
+if exist "data\settings.json.bak" (
+    copy /y "data\settings.json.bak" "data\settings.json" >nul
+    del /f /q "data\settings.json.bak" >nul 2>&1
+)
+del /f /q "data\settings_clean.tmp" >nul 2>&1
+
+if %BUILD_RESULT% neq 0 (
+    echo.
+    echo [ERROR] PyInstaller build failed. See output above for details.
+    pause & exit /b 1
+)
+
+:: ── 6. Result ───────────────────────────────────────────────────────────────
+echo.
+echo ============================================================
+echo   Build SUCCESSFUL!
+echo   Output: dist\BidPartner\bid_partner.exe
+echo ============================================================
+echo.
+echo   The 'dist\BidPartner' folder is your distributable package.
+echo   Users only need this folder - no Python installation required.
+echo   Each user must set their own API key in the app settings.
+echo.
+
+:: Open the output folder
+explorer "dist\BidPartner" >nul 2>&1
+
+endlocal
+pause

config.py

@@ -0,0 +1,145 @@
+import os
+import sys
+import threading
+import random
+import time
+from contextlib import contextmanager
+
+# When running as a PyInstaller bundle:
+#   sys._MEIPASS  →  read-only bundle dir (templates, static, prompts)
+#   sys.executable dir  →  writable dir next to the .exe (data, settings, db)
+if getattr(sys, 'frozen', False):
+    _BUNDLE_DIR = sys._MEIPASS                        # bundled app files
+    BASE_DIR    = os.path.dirname(sys.executable)     # writable runtime dir
+else:
+    _BUNDLE_DIR = os.path.dirname(os.path.abspath(__file__))
+    BASE_DIR    = _BUNDLE_DIR
+
+DATA_DIR     = os.path.join(BASE_DIR,    'data')
+UPLOAD_DIR   = os.path.join(DATA_DIR,   'uploads')
+EXPORT_DIR   = os.path.join(DATA_DIR,   'exports')
+KNOWLEDGE_DIR= os.path.join(DATA_DIR,   'knowledge')
+DB_PATH      = os.path.join(DATA_DIR,   'projects.db')
+CHROMA_DIR   = os.path.join(DATA_DIR,   'chroma')
+PROMPTS_DIR  = os.path.join(_BUNDLE_DIR, 'prompts')
+
+# ==================== AI 模型配置 ====================
+# 模型选择：'openai' | 'qwen' | 'deepseek' | 'ollama'
+MODEL_PROVIDER = os.environ.get('MODEL_PROVIDER', 'qwen')
+
+# OpenAI
+OPENAI_API_KEY  = os.environ.get('OPENAI_API_KEY',  'sk-your-openai-key')
+OPENAI_MODEL    = os.environ.get('OPENAI_MODEL',    'gpt-4.1')
+OPENAI_BASE_URL = os.environ.get('OPENAI_BASE_URL', 'https://api.openai.com/v1')
+
+# 阿里云通义千问
+QWEN_API_KEY  = os.environ.get('QWEN_API_KEY',  'sk-your-qwen-key')
+QWEN_MODEL    = os.environ.get('QWEN_MODEL',    'qwen3.6-plus')
+QWEN_BASE_URL = os.environ.get('QWEN_BASE_URL', 'https://dashscope.aliyuncs.com/compatible-mode/v1')
+# 文生图（多模态生成）与 Chat 接口不同：北京地域见 Model Studio 文档
+QWEN_MULTIMODAL_BASE = os.environ.get(
+    'QWEN_MULTIMODAL_BASE',
+    'https://dashscope.aliyuncs.com/api/v1',
+)
+# 投标附件插图默认模型（同步文生图）
+QWEN_IMAGE_MODEL = os.environ.get('QWEN_IMAGE_MODEL', 'qwen-image-2.0-pro')
+QWEN_IMAGE_SIZE = os.environ.get('QWEN_IMAGE_SIZE', '1536*1024')
+QWEN_IMAGE_PROMPT_EXTEND = os.environ.get('QWEN_IMAGE_PROMPT_EXTEND', 'true').lower() in (
+    '1', 'true', 'yes',
+)
+QWEN_IMAGE_WATERMARK = os.environ.get('QWEN_IMAGE_WATERMARK', 'false').lower() in (
+    '1', 'true', 'yes',
+)
+QWEN_IMAGE_NEGATIVE_PROMPT = os.environ.get(
+    'QWEN_IMAGE_NEGATIVE_PROMPT',
+    '低分辨率, 模糊, 畸形肢体, 过度饱和, 蜡像感, 杂乱构图, 扭曲文字, 公司商标, LOGO, 投标人名称, '
+    '塔吊, 塔式起重机, 起重机, 吊车, 挖掘机, 装载机, 压路机, 泵车, 搅拌车, 推土机, 施工机械, 工程车辆, '
+    'crane, excavator, tower crane, bulldozer',
+)
+
+# DeepSeek
+DEEPSEEK_API_KEY  = os.environ.get('DEEPSEEK_API_KEY',  'sk-your-deepseek-key')
+DEEPSEEK_MODEL    = os.environ.get('DEEPSEEK_MODEL',    'deepseek-chat')
+DEEPSEEK_BASE_URL = os.environ.get('DEEPSEEK_BASE_URL', 'https://api.deepseek.com/v1')
+
+# Ollama 本地（OpenAI 兼容接口）
+OLLAMA_BASE_URL = os.environ.get('OLLAMA_BASE_URL', 'http://localhost:11434/v1')
+OLLAMA_MODEL    = os.environ.get('OLLAMA_MODEL',    'qwen3:8b')
+
+# 豆包 / 火山引擎（字节跳动，OpenAI 兼容接口）
+DOUBAO_API_KEY  = os.environ.get('DOUBAO_API_KEY',  'sk-your-doubao-key')
+DOUBAO_MODEL    = os.environ.get('DOUBAO_MODEL',    'doubao-1-5-pro-32k')
+DOUBAO_BASE_URL = os.environ.get('DOUBAO_BASE_URL', 'https://ark.cn-beijing.volces.com/api/v3')
+
+# Kimi / Moonshot AI（OpenAI 兼容接口，支持 Embedding）
+KIMI_API_KEY  = os.environ.get('KIMI_API_KEY',  'sk-your-kimi-key')
+KIMI_MODEL    = os.environ.get('KIMI_MODEL',    'moonshot-v1-32k')
+KIMI_BASE_URL = os.environ.get('KIMI_BASE_URL', 'https://api.moonshot.cn/v1')
+
+# Embedding 模型
+OPENAI_EMBEDDING_MODEL = 'text-embedding-3-small'
+QWEN_EMBEDDING_MODEL   = 'text-embedding-v3'
+KIMI_EMBEDDING_MODEL   = 'moonshot-v1-embedding'
+
+# ==================== 应用配置 ====================
+MAX_FILE_SIZE_MB = 50
+ALLOWED_EXTENSIONS = {'pdf', 'doc', 'docx'}
+SECRET_KEY = 'bidhuo-partner-secret-2024'
+
+# ==================== 生成配置 ====================
+MAX_RETRIES = 3
+REQUEST_TIMEOUT = int(os.environ.get('REQUEST_TIMEOUT', '180'))
+# 大纲生成单次提示词长、输出大，适当延长读超时（秒），避免接口未返回即被客户端断开
+OUTLINE_REQUEST_TIMEOUT = int(os.environ.get('OUTLINE_REQUEST_TIMEOUT', '300'))
+CHUNK_SIZE = 2000      # 知识库文本分块大小（字符数）
+CHUNK_OVERLAP = 200    # 分块重叠大小
+TOP_K_KNOWLEDGE = 3   # 知识库检索数量
+CONTENT_VOLUME = os.environ.get('CONTENT_VOLUME', 'standard')  # 篇幅档位: concise / standard / detailed / full
+TARGET_PAGES = int(os.environ.get('TARGET_PAGES', '0') or '0')  # 目标页数（0=不启用）
+PAGE_CHAR_ESTIMATE = int(os.environ.get('PAGE_CHAR_ESTIMATE', '700') or '700')  # 粗略每页字数估算
+
+# ==================== 并发控制 (极速优化核心) ====================
+# 全局 LLM 调用信号量与章节生成线程池共用同一上限，默认 40 路真实并发。
+# 可通过环境变量 LLM_CONCURRENCY_CAP 提高硬上限（界面保存仍不可超过该值）。
+LLM_CONCURRENCY_CAP = int(os.environ.get('LLM_CONCURRENCY_CAP', '40'))
+_env_def = str(LLM_CONCURRENCY_CAP)
+_raw_llm = int(os.environ.get('LLM_CONCURRENCY_LIMIT', _env_def))
+_raw_sec = int(os.environ.get('MAX_CONCURRENT_SECTIONS', _env_def))
+_cv = max(1, min(_raw_llm, _raw_sec, LLM_CONCURRENCY_CAP))
+LLM_CONCURRENCY_LIMIT = _cv
+MAX_CONCURRENT_SECTIONS = _cv
+_llm_semaphore = threading.Semaphore(_cv)
+
+
+def refresh_llm_semaphore() -> None:
+    """按当前 LLM/章节并发配置重建信号量（加载 settings.json 后须调用）。"""
+    global _llm_semaphore, LLM_CONCURRENCY_LIMIT, MAX_CONCURRENT_SECTIONS
+    v = max(1, min(int(LLM_CONCURRENCY_LIMIT), int(MAX_CONCURRENT_SECTIONS), int(LLM_CONCURRENCY_CAP)))
+    LLM_CONCURRENCY_LIMIT = v
+    MAX_CONCURRENT_SECTIONS = v
+    _llm_semaphore = threading.Semaphore(v)
+
+
+def set_concurrency(n: int) -> None:
+    """统一设置章节池并发与 LLM 信号量上限，并重置信号量。"""
+    global MAX_CONCURRENT_SECTIONS, LLM_CONCURRENCY_LIMIT, _llm_semaphore
+    v = max(1, min(int(n), int(LLM_CONCURRENCY_CAP)))
+    MAX_CONCURRENT_SECTIONS = v
+    LLM_CONCURRENCY_LIMIT = v
+    _llm_semaphore = threading.Semaphore(v)
+
+
+@contextmanager
+def llm_call():
+    """全局LLM调用信号量。所有 ai_client.chat / embedding 必须使用。
+    针对Qwen云API增加轻微jitter避免429。超时60s防止死锁。"""
+    acquired = _llm_semaphore.acquire(blocking=True, timeout=60.0)
+    if not acquired:
+        raise TimeoutError(f"LLM并发已达上限({LLM_CONCURRENCY_LIMIT})，请稍后重试")
+    try:
+        # Qwen RPM敏感，增加极小jitter (0-0.08s) 避免429限流
+        if MODEL_PROVIDER == 'qwen':
+            time.sleep(random.uniform(0, 0.08))
+        yield
+    finally:
+        _llm_semaphore.release()

data/attachment_section_rules.json

@@ -0,0 +1,22 @@
+{
+  "_meta": "附件类章节：stack_charts_only 为默认，叶节点按 diagram 意图栈只输出 [FIGURE]/[TABLE] 块、无叙述正文；full 为长文；single_chart_only 为栈顶单块。修改后重启生效。",
+  "_field_docs": {
+    "title_regex": "标题任一则正则匹配即视为附件节（Python re 语法）",
+    "table_hint_keywords": "标题含此类子串且双开关均开时倾向表格",
+    "figure_hint_keywords": "标题含此类子串且双开关均开时倾向图示",
+    "default_kind_when_ambiguous": "双开且标题无倾向词时的默认：figure 或 table",
+    "attachment_leaf_body_mode": "stack_charts_only：意图栈只生成图/表块；full：与常规章节相同长文；single_chart_only：仅栈顶一块图或表"
+  },
+  "schema_version": 1,
+  "attachment_leaf_body_mode": "stack_charts_only",
+  "title_regex": [
+    "附件\\s*[一二三四五六七八九十0-9A-Za-z、：:.]",
+    "附\\s*图",
+    "附\\s*表",
+    "附\\s*件\\s*\\(",
+    "^\\s*[\\d一二三四五六七八九十\\.．、]+\\s*附件"
+  ],
+  "table_hint_keywords": ["附表", "一览表", "清单表", "表", "统计表", "明细表"],
+  "figure_hint_keywords": ["附图", "示意图", "平面图", "流程图", "布置图", "组织图", "横道"],
+  "default_kind_when_ambiguous": "table"
+}

data/diagram_intent_rules.json

@@ -0,0 +1,51 @@
+{
+  "_meta": "章节级图/表意图：标题与大纲窗口关键词计分，阈值入栈，按栈序拼接图示/表格生成规范。修改后重启服务生效。",
+  "_field_docs": {
+    "threshold_figure": "图示倾向分达到此值才入栈",
+    "threshold_table": "表格倾向分达到此值才入栈",
+    "title_weight": "标题命中的权重乘子",
+    "context_weight": "大纲上下文窗口命中的权重乘子",
+    "outline_context_lines": "before/after 为相对匹配行上下扩展行数",
+    "stack_order_when_both": "figure_first | table_first | score_desc（两者同时入栈时的顺序，栈顶为 index 0）",
+    "figure_keywords": "字符串或 {text,weight} 对象列表",
+    "table_keywords": "同上"
+  },
+  "schema_version": 1,
+  "threshold_figure": 1.0,
+  "threshold_table": 1.0,
+  "title_weight": 1.0,
+  "context_weight": 0.6,
+  "outline_context_lines": {"before": 4, "after": 6},
+  "stack_order_when_both": "score_desc",
+  "figure_keywords": [
+    {"text": "组织", "weight": 1.0},
+    {"text": "架构", "weight": 1.0},
+    {"text": "流程", "weight": 1.2},
+    {"text": "工序", "weight": 1.0},
+    {"text": "进度", "weight": 1.2},
+    {"text": "横道", "weight": 1.5},
+    {"text": "网络图", "weight": 1.5},
+    {"text": "平面", "weight": 1.0},
+    {"text": "布置", "weight": 0.8},
+    {"text": "监测", "weight": 0.8},
+    {"text": "示意", "weight": 0.8},
+    {"text": "应急", "weight": 0.8}
+  ],
+  "table_keywords": [
+    {"text": "一览表", "weight": 1.5},
+    {"text": "人员", "weight": 1.0},
+    {"text": "配置", "weight": 0.8},
+    {"text": "设备", "weight": 1.0},
+    {"text": "机械", "weight": 0.9},
+    {"text": "劳动力", "weight": 1.2},
+    {"text": "工种", "weight": 1.0},
+    {"text": "检验", "weight": 1.0},
+    {"text": "验收", "weight": 0.9},
+    {"text": "材料", "weight": 1.0},
+    {"text": "供应", "weight": 0.9},
+    {"text": "风险", "weight": 1.0},
+    {"text": "措施", "weight": 0.6},
+    {"text": "清单", "weight": 0.8},
+    {"text": "计划", "weight": 0.7}
+  ]
+}

data/settings.json

@@ -0,0 +1,25 @@
+{
+  "model_provider": "qwen",
+  "qwen_api_key": "sk-999173b3ca7f425a97cc4b12a2d3575f",
+  "qwen_model": "qwen-turbo-latest",
+  "qwen_base_url": "https://dashscope.aliyuncs.com/compatible-mode/v1",
+  "openai_api_key": "sk-your-openai-key",
+  "openai_model": "gpt-4.1",
+  "openai_base_url": "https://api.openai.com/v1",
+  "deepseek_api_key": "sk-your-deepseek-key",
+  "deepseek_model": "deepseek-chat",
+  "deepseek_base_url": "https://api.deepseek.com/v1",
+  "ollama_base_url": "http://localhost:11434/v1",
+  "ollama_model": "qwen3:8b",
+  "doubao_api_key": "sk-your-doubao-key",
+  "doubao_model": "doubao-1-5-pro-32k",
+  "doubao_base_url": "https://ark.cn-beijing.volces.com/api/v3",
+  "kimi_api_key": "sk-your-kimi-key",
+  "kimi_model": "moonshot-v1-32k",
+  "kimi_base_url": "https://api.moonshot.cn/v1",
+  "max_concurrent": 40,
+  "llm_concurrency_limit": 40,
+  "content_volume": "concise",
+  "target_pages": 120,
+  "page_char_estimate": 700
+}

data/style_presets.json

@@ -0,0 +1,35 @@
+{
+  "standard": {
+    "marginTop": 2.54,
+    "marginBottom": 2.54,
+    "marginLeft": 3.18,
+    "marginRight": 3.18,
+    "paperOrientation": "纵",
+    "heading1Font": "黑体",
+    "heading1Size": "三号",
+    "heading1Bold": true,
+    "heading2Font": "宋体",
+    "heading2Size": "小四",
+    "tableHeaderFont": "宋体",
+    "tableHeaderSize": "小四",
+    "tableHeaderLineSpacing": "24",
+    "tableHeaderAlign": "center",
+    "tableHeaderBold": false,
+    "tableBodyFont": "宋体",
+    "tableBodySize": "小四",
+    "tableBodyLineSpacing": "24",
+    "tableBodyAlign": "center",
+    "tableBodyBold": false,
+    "tocEnabled": true,
+    "tocTitleFont": "黑体",
+    "tocTitleSize": "三号",
+    "tocTitleLineSpacing": "24",
+    "tocTitleAlign": "center",
+    "tocTitleBold": true,
+    "tocBodyFont": "宋体",
+    "tocBodySize": "四号",
+    "tocBodyLineSpacing": "24",
+    "tocBodyAlign": "left",
+    "tocBodyBold": false
+  }
+}