init: RAG 知识库服务初始提交

- 后端 API（Flask + Gunicorn）
- RAG 引擎（混合检索 + 云端 Reranker + 引用溯源）
- 文档解析（MinerU + 多格式支持）
- Docker 生产部署配置
- 排除前端项目、敏感配置、模型文件

core/agentic_quality.py

@@ -0,0 +1,137 @@
+"""
+Agentic RAG - 质量评估 Mixin
+
+包含置信度门控、质量评估、推理反思等方法
+"""
+
+import logging
+
+from .agentic_base import logger
+
+logger = logging.getLogger(__name__)
+
+
+class QualityMixin:
+    """质量评估方法"""
+
+    def _check_confidence_gate(self, query: str, docs: list, verbose: bool = True,
+                               precomputed_scores: list = None):
+        """检查置信度门控
+
+        Args:
+            query: 用户查询
+            docs: 文档列表
+            verbose: 是否详细输出
+            precomputed_scores: 预计算的 Rerank 分数（可选，避免重复推理）
+        """
+        if not self.confidence_gate:
+            return {"passed": True, "reason": "no_gate"}
+
+        try:
+            result = self.confidence_gate.evaluate(query, docs,
+                                                    precomputed_scores=precomputed_scores)
+            return result
+        except Exception as e:
+            logger.warning(f"置信度门控检查失败: {e}")
+            return {"passed": True, "reason": "error"}
+
+    def _assess_quality(self, query: str, docs: list, metas: list = None,
+                       verbose: bool = True) -> dict:
+        """多维质量评估"""
+        if not self.quality_assessor:
+            return {"overall_score": 0.5, "dimensions": {}}
+
+        try:
+            result = self.quality_assessor.assess(query, docs, metas)
+            return result
+        except Exception as e:
+            logger.warning(f"质量评估失败: {e}")
+            return {"overall_score": 0.5, "dimensions": {}}
+
+    def _reflect_on_answer(self, query: str, answer: str, contexts: list,
+                          verbose: bool = True) -> dict:
+        """推理反思"""
+        if not self.reasoning_reflector:
+            return {"needs_reflection": False, "issues": []}
+
+        try:
+            result = self.reasoning_reflector.reflect(query, answer, contexts)
+            return result
+        except Exception as e:
+            logger.warning(f"推理反思失败: {e}")
+            return {"needs_reflection": False, "issues": []}
+
+    def _think(self, original_query: str, current_query: str,
+               iteration: int, contexts: list, verbose: bool = True) -> dict:
+        """
+        Agent 思考：决定下一步行动
+
+        Returns:
+            {
+                "action": "answer" | "rewrite" | "search_web" | "decompose",
+                "reason": "...",
+                "rewrite_query": "..."  # 如果 action == "rewrite"
+            }
+        """
+        from core.llm_utils import call_llm, parse_json_from_response
+        from .agentic_base import MODEL
+
+        # 构建思考提示
+        context_summary = ""
+        if contexts:
+            for i, ctx in enumerate(contexts[:3], 1):
+                meta = ctx.get('meta', {})
+                source = meta.get('source', '未知')
+                doc_preview = ctx.get('doc', '')[:100]
+                context_summary += f"{i}. [{source}] {doc_preview}...\n"
+
+        prompt = f"""你是一个 RAG 系统的决策 Agent，需要判断下一步行动。
+
+【原始问题】
+{original_query}
+
+【当前问题】
+{current_query}
+
+【迭代轮次】
+{iteration} / {self.max_iterations}
+
+【已检索到的上下文】
+{context_summary if context_summary else "（无）"}
+
+【可选行动】
+1. answer - 已有足够信息，可以回答
+2. rewrite - 查询不够清晰，需要重写
+3. search_web - 知识库信息不足，需要网络搜索
+4. decompose - 问题太复杂，需要分解
+
+【决策要求】
+- 如果上下文足够回答问题，选择 answer
+- 如果上下文不足且迭代未超限，选择 search_web 或 rewrite
+- 返回 JSON 格式
+
+请决策："""
+
+        try:
+            result = call_llm(
+                self.client, prompt, MODEL,
+                temperature=0.3,
+                max_tokens=200
+            )
+
+            decision = parse_json_from_response(result) if result else {}
+
+            # 默认决策
+            if not decision or "action" not in decision:
+                if contexts and len(contexts) >= 2:
+                    decision = {"action": "answer", "reason": "有足够上下文"}
+                else:
+                    decision = {"action": "rewrite", "reason": "上下文不足"}
+
+            return decision
+
+        except Exception as e:
+            logger.warning(f"Agent 思考失败: {e}")
+            if contexts:
+                return {"action": "answer", "reason": "默认回答"}
+            return {"action": "rewrite", "reason": "默认重写"}