重构：移除对 debug_info 的依赖，只使用新的结构化状态字段

backend/app/main_graph/nodes/rag_nodes.py

@@ -48,8 +48,8 @@ async def _rag_retrieve_core(state: MainGraphState, pipeline) -> MainGraphState:
     """执行 RAG 检索的核心逻辑"""
     retrieval_query = state.user_query
 
-    # 优先使用推理结果中的优化查询
-    reasoning_result = state.debug_info.get("reasoning_result")
+    # 优先使用推理结果中的优化查询 - 从新的结构化字段获取
+    reasoning_result = state.react_reasoning.reasoning_result
     if reasoning_result and hasattr(reasoning_result, "retrieval_config"):
         cfg = reasoning_result.retrieval_config
         if cfg and cfg.retrieval_query:
@@ -67,8 +67,7 @@ async def _rag_retrieve_core(state: MainGraphState, pipeline) -> MainGraphState:
     state.rag_docs = documents  # 保存文档用于置信度评估
     state.rag_retrieved = bool(documents)  # 有文档才算检索成功
     state.rag_attempts = getattr(state, 'rag_attempts', 0) + 1
-    state.debug_info["rag_source"] = "pipeline"
-    state.debug_info["rag_scores"] = pipeline.last_scores  # 保存分数信息
+    # 移除对 debug_info 的依赖，不再保存 rag_scores
 
     return state
 
@@ -148,16 +147,7 @@ async def _evaluate_rag_confidence(state: MainGraphState) -> float:
 
 
 def _get_embedding_similarity(state: MainGraphState) -> float:
-    """从 rag_scores 或 rag_docs 中获取向量相似度分数"""
-    # 优先从 pipeline 提供的分数中获取
-    rag_scores = state.debug_info.get("rag_scores", [])
-    if rag_scores:
-        scores = [s.get("embedding_score", 0.0) for s in rag_scores]
-        if scores:
-            # 归一化到 0-1
-            normalized = [min(s / 10.0, 1.0) if s > 1.0 else s for s in scores]
-            return max(normalized)
-
+    """从 rag_docs 中获取向量相似度分数（不再从 debug_info 获取）"""
     # 降级：从 rag_docs 中获取
     rag_docs = getattr(state, "rag_docs", [])
     scores = []
@@ -176,13 +166,7 @@ def _get_embedding_similarity(state: MainGraphState) -> float:
 
 
 def _get_rerank_score(state: MainGraphState) -> float:
-    """从 rag_scores 或 rag_docs 中获取重排序分数"""
-    # 优先从 pipeline 提供的分数中获取
-    rag_scores = state.debug_info.get("rag_scores", [])
-    if rag_scores:
-        scores = [s.get("rerank_score", 0.0) for s in rag_scores]
-        return max(scores) if scores else 0.0
-
+    """从 rag_docs 中获取重排序分数（不再从 debug_info 获取）"""
     # 降级：从 rag_docs 中获取
     rag_docs = getattr(state, "rag_docs", [])
     scores = []