diff --git a/backend/app/main_graph/nodes/rag_nodes.py b/backend/app/main_graph/nodes/rag_nodes.py index 87893fe..a88856a 100644 --- a/backend/app/main_graph/nodes/rag_nodes.py +++ b/backend/app/main_graph/nodes/rag_nodes.py @@ -45,7 +45,7 @@ def _get_rag_tool() -> Optional[callable]: # ========== RAG 检索核心逻辑 ========== async def _rag_retrieve_core(state: MainGraphState, pipeline) -> MainGraphState: - """执行 RAG 检索的核心逻辑""" + info(f"[RAG Core] _rag_retrieve_core 开始") retrieval_query = state.user_query # 优先使用推理结果中的优化查询 - 从新的结构化字段获取 @@ -55,9 +55,14 @@ async def _rag_retrieve_core(state: MainGraphState, pipeline) -> MainGraphState: if cfg and cfg.retrieval_query: retrieval_query = cfg.retrieval_query + info(f"[RAG Core] 使用检索查询: {retrieval_query[:50]}...") # 直接调用 pipeline 获取文档和上下文 + info(f"[RAG Core] 调用 pipeline.aretrieve") documents = await pipeline.aretrieve(retrieval_query) + info(f"[RAG Core] pipeline.aretrieve 返回，得到 {len(documents)} 个文档") + info(f"[RAG Core] 调用 pipeline.format_context") rag_context = pipeline.format_context(documents) + info(f"[RAG Core] pipeline.format_context 返回") info(f"[RAG Core] 获取到 rag_context: {type(rag_context)}, 长度={len(rag_context) if rag_context else 0}") info(f"[RAG Core] 获取到 rag_docs: {len(documents)} 个文档") @@ -69,15 +74,17 @@ async def _rag_retrieve_core(state: MainGraphState, pipeline) -> MainGraphState: state.rag_attempts = getattr(state, 'rag_attempts', 0) + 1 # 移除对 debug_info 的依赖，不再保存 rag_scores + info(f"[RAG Core] _rag_retrieve_core 结束") return state # ========== RAG 检索节点 ========== async def rag_retrieve_node(state: MainGraphState, config: Optional[RunnableConfig] = None) -> MainGraphState: - """RAG 检索节点：检索 + 置信度评估""" + info(f"[RAG] rag_retrieve_node 开始") state.current_phase = "rag_retrieving" start_time = time.time() + info(f"[RAG] 调用 _get_rag_pipeline") pipeline = _get_rag_pipeline() await dispatch_custom_event( @@ -87,9 +94,12 @@ async def rag_retrieve_node(state: MainGraphState, config: Optional[RunnableConf ) try: + info(f"[RAG] 调用 _rag_retrieve_core") state = await _rag_retrieve_core(state, pipeline) + info(f"[RAG] _rag_retrieve_core 返回") # 评估置信度 + info(f"[RAG] 调用 _evaluate_rag_confidence") confidence = await _evaluate_rag_confidence(state) state.rag_confidence = confidence @@ -111,10 +121,11 @@ async def rag_retrieve_node(state: MainGraphState, config: Optional[RunnableConf ) except Exception as e: - info(f"[RAG] 检索失败: {e}") + info(f"[RAG] 检索失败: {e}", exc_info=True) state.rag_confidence = 0.0 state.rag_retrieved = False + info(f"[RAG] rag_retrieve_node 结束") return state diff --git a/backend/app/rag/fusion.py b/backend/app/rag/fusion.py index ddf8f42..9f1e464 100644 --- a/backend/app/rag/fusion.py +++ b/backend/app/rag/fusion.py @@ -1,8 +1,11 @@ # rag/fusion.py +import logging from typing import List, Dict from langchain_core.documents import Document +logger = logging.getLogger(__name__) + def reciprocal_rank_fusion( doc_lists: List[List[Document]], k: int = 60 @@ -17,12 +20,14 @@ def reciprocal_rank_fusion( Returns: 融合后按 RRF 得分降序排列的文档列表 """ + logger.info(f"[RRF] reciprocal_rank_fusion 开始: {len(doc_lists)} 组文档") # 使用文档内容作为唯一标识（如果内容相同但 metadata 不同，视为同一文档） # 更好的做法是用 docstore 的 ID，这里简化处理：用内容 hash doc_to_score: Dict[str, float] = {} doc_map: Dict[str, Document] = {} - for docs in doc_lists: + for list_idx, docs in enumerate(doc_lists): + logger.info(f"[RRF] 处理第 {list_idx} 组: {len(docs)} 个文档") for rank, doc in enumerate(docs, start=1): # 生成唯一标识符（内容+来源组合，避免不同文件相同内容混淆） doc_id = f"{doc.page_content[:200]}_{doc.metadata.get('source', '')}" @@ -31,6 +36,9 @@ def reciprocal_rank_fusion( score = doc_to_score.get(doc_id, 0.0) + 1.0 / (k + rank) doc_to_score[doc_id] = score + logger.info(f"[RRF] 去重后共 {len(doc_map)} 个唯一文档") # 按得分排序 sorted_ids = sorted(doc_to_score.keys(), key=lambda x: doc_to_score[x], reverse=True) - return [doc_map[doc_id] for doc_id in sorted_ids] \ No newline at end of file + result = [doc_map[doc_id] for doc_id in sorted_ids] + logger.info(f"[RRF] reciprocal_rank_fusion 结束: 返回 {len(result)} 个文档") + return result \ No newline at end of file diff --git a/backend/app/rag/pipeline.py b/backend/app/rag/pipeline.py index a161cc4..1d85bcb 100644 --- a/backend/app/rag/pipeline.py +++ b/backend/app/rag/pipeline.py @@ -62,31 +62,40 @@ class RAGPipeline: return self._last_scores async def aretrieve(self, query: str) -> List[Document]: + logger.info(f"[Pipeline] aretrieve 开始: query={query[:50]}...") # Step 1: 检索 + logger.info(f"[Pipeline] Step 1: 调用 _retrieve") child_docs = await self._retrieve(query) - logger.info(f"[Pipeline] 检索到 {len(child_docs)} 个子文档") + logger.info(f"[Pipeline] Step 1 完成: 检索到 {len(child_docs)} 个子文档") # 调试：打印子文档长度 for i, doc in enumerate(child_docs[:5]): content_len = len(doc.page_content) logger.info(f"[Pipeline] 子文档[{i}] 长度={content_len}字符") # Step 2: 重排 + logger.info(f"[Pipeline] Step 2: 开始重排") if self.reranker: try: child_docs = self.reranker.compress_documents(child_docs, query, self.rerank_top_n) - logger.info(f"[Pipeline] 重排后 {len(child_docs)} 个") + logger.info(f"[Pipeline] Step 2 完成: 重排后 {len(child_docs)} 个") except Exception as e: logger.warning(f"[Pipeline] 重排失败: {e}") child_docs = child_docs[:self.rerank_top_n] + else: + logger.info(f"[Pipeline] Step 2 跳过: 未启用 reranker") # Step 3: 获取父文档 + logger.info(f"[Pipeline] Step 3: 开始获取父文档") if self.return_parent_docs: parent_docs = await self._get_parents(child_docs) + logger.info(f"[Pipeline] Step 3 完成: 获取到 {len(parent_docs)} 个父文档") # 保存分数信息到 last_scores 供外部访问 self._last_scores = self._extract_scores(parent_docs) + logger.info(f"[Pipeline] aretrieve 结束: 返回父文档") return parent_docs self._last_scores = self._extract_scores(child_docs) + logger.info(f"[Pipeline] aretrieve 结束: 返回子文档") return child_docs def _extract_scores(self, docs: List[Document]) -> List[dict]: @@ -100,14 +109,27 @@ class RAGPipeline: return scores async def _retrieve(self, query: str) -> List[Document]: + logger.info(f"[Pipeline] _retrieve 开始: query={query[:50]}...") if self.query_generator: + logger.info(f"[Pipeline] _retrieve: 调用 query_generator.agenerate") queries = await self.query_generator.agenerate(query) queries = [query] + [q for q in queries if q != query] + logger.info(f"[Pipeline] _retrieve: 生成 {len(queries)} 个查询: {queries}") + logger.info(f"[Pipeline] _retrieve: 开始 asyncio.gather 并行检索") doc_lists = await asyncio.gather(*[self.retriever.ainvoke(q) for q in queries]) - return reciprocal_rank_fusion(doc_lists) - return await self.retriever.ainvoke(query) + logger.info(f"[Pipeline] _retrieve: asyncio.gather 完成，得到 {len(doc_lists)} 组结果") + logger.info(f"[Pipeline] _retrieve: 开始 reciprocal_rank_fusion") + result = reciprocal_rank_fusion(doc_lists) + logger.info(f"[Pipeline] _retrieve: RRF 完成，得到 {len(result)} 个文档") + logger.info(f"[Pipeline] _retrieve 结束") + return result + logger.info(f"[Pipeline] _retrieve: query_generator 未启用，直接单次检索") + result = await self.retriever.ainvoke(query) + logger.info(f"[Pipeline] _retrieve 结束") + return result async def _get_parents(self, child_docs: List[Document]) -> List[Document]: + logger.info(f"[Pipeline] _get_parents 开始: {len(child_docs)} 个子文档") # 收集 parent_id 和对应的分数 parent_map = {} # parent_id -> (embedding_score, rerank_score) @@ -120,14 +142,18 @@ class RAGPipeline: rerank_score = doc.metadata.get("rerank_score", 0.0) parent_map[pid] = (embedding_score, rerank_score) + logger.info(f"[Pipeline] _get_parents: 收集到 {len(parent_map)} 个 unique parent_id") if not parent_map: logger.warning("[Pipeline] 未找到 parent_id，返回子文档") return child_docs try: + logger.info(f"[Pipeline] _get_parents: 调用 create_docstore") from backend.rag_core import create_docstore docstore, _ = create_docstore() + logger.info(f"[Pipeline] _get_parents: 调用 docstore.amget") parent_docs =await docstore.amget(list(parent_map.keys())) + logger.info(f"[Pipeline] _get_parents: docstore.amget 返回 {len(parent_docs)} 个结果") # 构建结果，保持分数信息 result = [] @@ -142,20 +168,25 @@ class RAGPipeline: result.sort(key=lambda x: x[1], reverse=True) docs = [d for d, _ in result] - logger.info(f"[Pipeline] 获取到 {len(docs)} 个父文档") + logger.info(f"[Pipeline] _get_parents: 最终得到 {len(docs)} 个父文档") + logger.info(f"[Pipeline] _get_parents 结束") return docs except Exception as e: - logger.warning(f"[Pipeline] 获取父文档失败: {e}") + logger.warning(f"[Pipeline] 获取父文档失败: {e}", exc_info=True) return child_docs def format_context(self, documents: List[Document]) -> str: + logger.info(f"[Pipeline] format_context 开始: {len(documents)} 个文档") if not documents: + logger.info(f"[Pipeline] format_context: 无文档，返回空字符串") return "" parts = [] for i, doc in enumerate(documents, 1): source = doc.metadata.get("source", "未知来源") parts.append(f"【资料 {i}】来源：{source}\n{doc.page_content}\n---\n") - return "\n".join(parts) + result = "\n".join(parts) + logger.info(f"[Pipeline] format_context 结束: 结果长度={len(result)} 字符") + return result def create_rag_pipeline(**kwargs) -> RAGPipeline: