重构七天训练营老师

l2_training_camp_agent/kb_tools.py

@@ -0,0 +1,204 @@
+"""L2 训练营知识库工具 — 按需加载 + 语义搜索
+
+本文件包含知识库查询工具的实现。修改主提示词时请勿改动此文件。
+"""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+from typing import List
+
+from common.knowledge_loader import list_knowledge_files, load_knowledge_file
+
+# ───────────────────────────── 配置 ─────────────────────────────
+_KB_DIR = Path(__file__).parent / "knowledge_base"
+
+# 知识库元数据（与文件名一一对应）
+_KB_METADATA: dict[str, dict] = {
+    "01_课程层次与问题判断.md": {
+        "title": "课程层次与问题判断",
+        "description": "帮助判断学员问题属于299启蒙课还是1280训练营层次，以及问题卡在哪个阶段",
+        "keywords": [
+            "课程层次", "299", "1280", "启蒙课", "训练营", "问题判断",
+            "阶段", "层次", "基础", "进阶", "学员水平", "课程区别",
+        ],
+    },
+    "02_即时回复表达原则.md": {
+        "title": "即时回复表达原则",
+        "description": "辅导老师回复学员时的表达风格、语气、措辞原则和禁忌",
+        "keywords": [
+            "表达", "回复", "语气", "措辞", "风格", "原则", "禁忌",
+            "怎么说", "话术", "沟通", "表达技巧", "回复模板",
+        ],
+    },
+    "03_常见学员问题回应素材.md": {
+        "title": "常见学员问题回应素材",
+        "description": "学员常见问题的分类、回应策略和具体话术示例",
+        "keywords": [
+            "常见问题", "学员问题", "回应", "话术", "示例", "策略",
+            "情绪", "愤怒", "委屈", "焦虑", "失望", "质疑", "抱怨",
+            "婚姻", "孩子", "工作", "健康", "家庭", "关系",
+        ],
+    },
+    "04_功课与练习反馈边界.md": {
+        "title": "功课与练习反馈边界",
+        "description": "学员功课、练习、打卡的反馈原则和边界，什么该说什么不该说",
+        "keywords": [
+            "功课", "练习", "打卡", "反馈", "边界", "作业",
+            "释放法", "万能释放法", "数息法", "身体感知", "情绪觉察",
+            "练习指导", "功课点评", "训练",
+        ],
+    },
+    "05_安全边界与现实风险.md": {
+        "title": "安全边界与现实风险",
+        "description": "学员安全风险识别、危机干预边界、何时必须上报",
+        "keywords": [
+            "安全", "风险", "危机", "自杀", "自残", "伤害", "暴力",
+            "报警", "上报", "边界", "现实风险", "法律保护", "隐私",
+            "紧急情况", "危险", "保护",
+        ],
+    },
+    "06_1280核心术语词典.md": {
+        "title": "1280核心术语词典",
+        "description": "1280训练营核心术语的标准解释和使用场景",
+        "keywords": [
+            "术语", "词典", "概念", "定义", "解释",
+            "看见自己", "回到自己", "万能释放法", "情绪心钥", "抗拒心钥",
+            "原生数据", "潜意识数据", "数息法", "身体感知", "情绪觉察",
+            "内在模式", "释放", "觉察", "感知",
+        ],
+    },
+    "07_大本营承接与成交话术.md": {
+        "title": "大本营承接与成交话术",
+        "description": "大本营课程承接判断、成交话术、异议处理策略",
+        "keywords": [
+            "大本营", "承接", "成交", "异议", "销售", "推课",
+            "后续课程", "持续训练", "课程信心", "报名", "价格", "权益",
+            "没时间", "没钱", "怕没效果", "家人不同意", "不想继续",
+        ],
+    },
+}
+
+# ───────────────────────────── 工具函数 ─────────────────────────────
+
+def _tokenize(text: str) -> set[str]:
+    """中文分词（使用jieba）"""
+    try:
+        import jieba
+
+        return set(jieba.cut(text))
+    except ImportError:
+        # 降级：按字拆分
+        return set(text)
+
+
+def _calc_relevance(query: str, content: str) -> float:
+    """计算查询与内容的相关度（0-1）"""
+    query_tokens = _tokenize(query.lower())
+    content_tokens = _tokenize(content.lower())
+
+    if not query_tokens:
+        return 0.0
+
+    overlap = query_tokens & content_tokens
+    return len(overlap) / len(query_tokens)
+
+
+def _extract_relevant_sections(content: str, query: str, max_sections: int = 3) -> list[str]:
+    """从知识库内容中提取与查询最相关的章节"""
+    # 按二级标题分割
+    sections = re.split(r"\n(?=##\s+)", content)
+
+    scored = []
+    for section in sections:
+        if not section.strip():
+            continue
+        score = _calc_relevance(query, section)
+        scored.append((score, section))
+
+    scored.sort(key=lambda x: x[0], reverse=True)
+    return [s for _, s in scored[:max_sections]]
+
+
+# ───────────────────────────── 公开工具 ─────────────────────────────
+
+def list_kb() -> str:
+    """列出所有可用的知识库文件
+
+    当你不确定该查哪个知识库文件时，先调用此工具查看列表。
+    """
+    lines = ["# 可用知识库文件\n"]
+    for fname, meta in _KB_METADATA.items():
+        lines.append(f"## {meta['title']}")
+        lines.append(f"- 文件名: `{fname}`")
+        lines.append(f"- 描述: {meta['description']}")
+        lines.append(f"- 关键词: {', '.join(meta['keywords'][:5])}...\n")
+
+    return "\n".join(lines)
+
+
+def load_kb(filename: str) -> str:
+    """加载指定知识库文件的完整内容
+
+    Args:
+        filename: 知识库文件名（如 "01_课程层次与问题判断.md"）
+
+    当你已确定需要哪个文件的内容时，调用此工具加载完整内容。
+    """
+    if filename not in _KB_METADATA:
+        available = ", ".join(_KB_METADATA.keys())
+        return f"错误: 文件 '{filename}' 不存在。可用文件: {available}"
+
+    content = load_knowledge_file(_KB_DIR, filename)
+    if not content:
+        return f"错误: 无法读取文件 '{filename}'"
+
+    return content
+
+
+def search_kb(query: str, top_k: int = 3) -> str:
+    """根据关键词搜索所有知识库文件，返回最相关的片段
+
+    Args:
+        query: 搜索关键词（中文）
+        top_k: 返回结果数量（默认3条）
+
+    当你不确定具体文件，但知道要查什么内容时，调用此工具。
+    """
+    all_files = list_knowledge_files(_KB_DIR)
+    results: list[tuple[float, str, str]] = []
+
+    for fname in all_files:
+        content = load_knowledge_file(_KB_DIR, fname)
+        if not content:
+            continue
+
+        # 计算文件级相关度
+        file_score = _calc_relevance(query, content)
+
+        # 提取最相关的章节
+        sections = _extract_relevant_sections(content, query, max_sections=2)
+
+        for section in sections:
+            section_score = _calc_relevance(query, section)
+            # 综合得分：文件级 + 章节级
+            final_score = (file_score + section_score) / 2
+            results.append((final_score, fname, section))
+
+    # 排序并取 top_k
+    results.sort(key=lambda x: x[0], reverse=True)
+    top_results = results[:top_k]
+
+    if not top_results:
+        return f"未找到与 '{query}' 相关的知识库内容。"
+
+    lines = [f"# 知识库搜索结果: '{query}'\n"]
+    for i, (score, fname, section) in enumerate(top_results, 1):
+        meta = _KB_METADATA.get(fname, {})
+        title = meta.get("title", fname)
+        lines.append(f"## 结果 {i}（相关度: {score:.2f}）")
+        lines.append(f"**来源**: {title} (`{fname}`)")
+        lines.append(f"**内容**:\n{section[:800]}...\n")
+
+    return "\n".join(lines)