feat: capture reasoning content from thinking models, improve thinking UI

- Backend: intercept reasoning_content from DeepSeek-R1/QwQ streaming
  chunks and emit as thinking events (stage=reasoning)
- RAG chain: show retrieved document titles/previews in thinking steps
- Conversation chain: directly stream from LLM to capture reasoning
- Frontend: collapsible thinking panel with reasoning section,
  document details, and time summary
- Replace relative time with HH:mm format for message timestamps

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/src/api/chat.py

@@ -295,6 +295,7 @@ async def stream_message(
                 # 使用带思考过程的流式输出
                 full_answer = ""
                 thinking_steps = []
+                first_chunk = True
                 async for result in conversation_chain.astream_with_thinking(request.message, chat_history=chat_history):
                     if result["type"] == "thinking":
                         # 收集思考过程
@@ -305,6 +306,9 @@ async def stream_message(
                         })
                         yield f"data: {json.dumps(result, ensure_ascii=False)}\n\n"
                     elif result["type"] == "chunk":
+                        if first_chunk:
+                            # 第一个实际内容chunk前发送reasoning完成信号
+                            first_chunk = False
                         full_answer += result["content"]
                         yield f"data: {json.dumps({'type': 'chunk', 'content': result['content']}, ensure_ascii=False)}\n\n"
                     elif result["type"] == "complete":

backend/src/rag/chains.py

@@ -140,27 +140,47 @@ class RAGChain:
         docs = await self.retriever.ainvoke(question)
         retrieval_time = time.time() - retrieval_start
 
-        # 发送检索结果
+        # 发送检索结果 — 包含文档标题和摘要
+        doc_details = []
+        for i, doc in enumerate(docs[:5]):
+            metadata = doc.metadata if hasattr(doc, 'metadata') else {}
+            title = metadata.get("title", metadata.get("filename", f"文档 {i+1}"))
+            preview = doc.page_content[:100].replace('\n', ' ')
+            doc_details.append(f"**{title}**: {preview}...")
+
         yield {
             "type": "thinking",
             "stage": "retrieved",
-            "message": f"找到 {len(docs)} 条相关文档",
+            "message": f"检索到 {len(docs)} 篇相关文档",
             "doc_count": len(docs),
-            "time": round(retrieval_time, 2)
+            "time": round(retrieval_time, 2),
+            "details": doc_details
         }
 
         context = "\n\n".join(doc.page_content for doc in docs)
 
         # 2. 构建prompt
-        yield {"type": "thinking", "stage": "generating", "message": "正在生成回答..."}
+        yield {"type": "thinking", "stage": "generating", "message": f"基于 {len(docs)} 篇文档生成回答..."}
         messages = await self.prompt.ainvoke({"context": context, "question": question})
 
-        # 3. 流式生成答案
+        # 3. 流式生成答案（捕获推理内容）
         answer_chunks = []
+        reasoning_parts = []
         async for chunk in self.llm.astream(messages):
+            # 捕获推理内容（DeepSeek-R1/QwQ 等推理模型）
+            if hasattr(chunk, 'additional_kwargs') and 'reasoning_content' in chunk.additional_kwargs:
+                reasoning_text = chunk.additional_kwargs['reasoning_content']
+                if reasoning_text:
+                    reasoning_parts.append(reasoning_text)
+                    yield {"type": "thinking", "stage": "reasoning", "message": reasoning_text}
+            elif hasattr(chunk, 'reasoning_content') and chunk.reasoning_content:
+                reasoning_parts.append(chunk.reasoning_content)
+                yield {"type": "thinking", "stage": "reasoning", "message": chunk.reasoning_content}
+
             content = chunk.content if hasattr(chunk, 'content') else str(chunk)
-            answer_chunks.append(content)
-            yield {"type": "chunk", "content": content}
+            if content:
+                answer_chunks.append(content)
+                yield {"type": "chunk", "content": content}
         
         # 4. 完成，返回sources
         total_time = time.time() - start_time

backend/src/rag/conversation_chains.py

@@ -80,7 +80,7 @@ class ConversationChain:
         }
     
     async def astream_with_thinking(self, question: str, chat_history: List[Dict] = None):
-        """流式调用（包含思考过程）"""
+        """流式调用（包含思考过程，捕获推理模型的真实推理内容）"""
         import time
 
         # 思考阶段
@@ -89,15 +89,32 @@ class ConversationChain:
 
         # 准备历史
         history_messages = self._format_history(chat_history or [])
+        history_count = len([m for m in (chat_history or []) if m["role"] == "user"])
+        yield {"type": "thinking", "stage": "preparing", "message": f"加载对话上下文（{history_count} 轮历史）..." if history_count > 0 else "准备生成回答..."}
 
         yield {"type": "thinking", "stage": "generating", "message": "正在生成回答..."}
 
-        # 流式生成
-        async for chunk in self.chain.astream({
+        # 直接流式调用 LLM 以捕获推理内容
+        prompt_messages = await self.prompt.ainvoke({
             "question": question,
             "chat_history": history_messages
-        }):
-            yield {"type": "chunk", "content": chunk}
+        })
+
+        content_started = False
+        async for chunk in self.llm.astream(prompt_messages):
+            # 捕获推理内容（DeepSeek-R1/QwQ 等推理模型）
+            if hasattr(chunk, 'additional_kwargs') and 'reasoning_content' in chunk.additional_kwargs:
+                reasoning_text = chunk.additional_kwargs['reasoning_content']
+                if reasoning_text:
+                    yield {"type": "thinking", "stage": "reasoning", "message": reasoning_text}
+            elif hasattr(chunk, 'reasoning_content') and chunk.reasoning_content:
+                yield {"type": "thinking", "stage": "reasoning", "message": chunk.reasoning_content}
+
+            content = chunk.content if hasattr(chunk, 'content') else str(chunk)
+            if content:
+                if not content_started:
+                    content_started = True
+                yield {"type": "chunk", "content": content}
         
         # 完成
         total_time = time.time() - start_time

web/src/components/chat/message-item.tsx

@@ -7,8 +7,7 @@ import ReactMarkdown from "react-markdown";
 import remarkGfm from "remark-gfm";
 import { Prism as SyntaxHighlighter } from "react-syntax-highlighter";
 import { tomorrow } from "react-syntax-highlighter/dist/esm/styles/prism";
-import { formatDistanceToNow } from "date-fns";
-import { zhCN } from "date-fns/locale";
+import { format } from "date-fns";
 import { Button } from "@/components/ui/button";
 import { useChatStore } from "@/store/chat";
 import SourceReferences from "./source-references";
@@ -20,28 +19,100 @@ interface MessageItemProps {
   selectedModel?: string;
 }
 
-const ThinkingProcess = ({ thinking }: { thinking: ThinkingStep[] }) => {
+const ThinkingProcess = ({ thinking, isStreaming }: { thinking: ThinkingStep[]; isStreaming?: boolean }) => {
+  const [expanded, setExpanded] = useState(false);
   if (!thinking || thinking.length === 0) return null;
 
-  const getStageIcon = (stage: string) => {
-    switch (stage) {
-      case 'understanding': return <Brain className="h-3.5 w-3.5" />;
-      case 'retrieving': return <FileSearch className="h-3.5 w-3.5 animate-spin" />;
-      case 'retrieved': return <CheckCircle2 className="h-3.5 w-3.5 text-green-500" />;
-      case 'generating': return <Sparkles className="h-3.5 w-3.5 animate-pulse" />;
-      default: return <Loader2 className="h-3.5 w-3.5" />;
-    }
-  };
+  // 分离状态步骤和推理内容
+  const statusSteps = thinking.filter(s => s.stage !== 'reasoning');
+  const reasoningSteps = thinking.filter(s => s.stage === 'reasoning');
+  const hasReasoning = reasoningSteps.length > 0;
+
+  // 合并推理文本
+  const reasoningText = reasoningSteps.map(s => s.message).join('');
+
+  // 汇总信息
+  const retrievedStep = statusSteps.find(s => s.stage === 'retrieved');
+  const totalTime = statusSteps.reduce((sum, s) => sum + (s.time || 0), 0);
+
+  // 流式时显示最后状态
+  const lastStatusStep = statusSteps[statusSteps.length - 1];
+  const isReasoningNow = isStreaming && thinking[thinking.length - 1]?.stage === 'reasoning';
+
+  // 折叠标题
+  const collapsedTitle = isStreaming
+    ? isReasoningNow
+      ? '深度思考中...'
+      : lastStatusStep?.message || '思考中...'
+    : hasReasoning
+      ? `思考过程 (${reasoningText.length} 字)`
+      : `思考过程${totalTime > 0 ? ` (${totalTime.toFixed(1)}s)` : ''}`;
 
   return (
-    <div className="mb-2.5 space-y-1.5 text-xs text-muted-foreground bg-muted/30 rounded-lg p-2.5">
-      {thinking.map((step, index) => (
-        <div key={index} className="flex items-center gap-2">
-          {getStageIcon(step.stage)}
-          <span>{step.message}</span>
-          {step.time && <span className="opacity-60">({step.time}s)</span>}
+    <div className="mb-2.5">
+      <button
+        onClick={() => setExpanded(!expanded)}
+        className="flex items-center gap-1.5 text-xs text-muted-foreground hover:text-foreground transition-colors"
+      >
+        <svg
+          className={cn("h-3 w-3 transition-transform", expanded && "rotate-90")}
+          fill="none" viewBox="0 0 24 24" stroke="currentColor" strokeWidth={2}
+        >
+          <path strokeLinecap="round" strokeLinejoin="round" d="M9 5l7 7-7 7" />
+        </svg>
+        <Brain className={cn("h-3 w-3", isReasoningNow && "animate-pulse text-primary")} />
+        <span>{collapsedTitle}</span>
+        {retrievedStep?.doc_count != null && (
+          <span className="opacity-60 ml-1">· {retrievedStep.doc_count} 篇文档</span>
+        )}
+        {totalTime > 0 && !isStreaming && (
+          <span className="opacity-50 ml-1">· {totalTime.toFixed(1)}s</span>
+        )}
+      </button>
+      {expanded && (
+        <div className="mt-1.5 ml-4 space-y-2 text-xs border-l-2 border-border pl-3">
+          {/* 状态步骤 */}
+          {statusSteps.map((step, index) => (
+            <div key={`s-${index}`} className="flex items-center gap-1.5 text-muted-foreground">
+              {step.stage === 'retrieving' ? (
+                <FileSearch className="h-3 w-3" />
+              ) : step.stage === 'retrieved' ? (
+                <CheckCircle2 className="h-3 w-3 text-green-500" />
+              ) : step.stage === 'generating' ? (
+                <Sparkles className="h-3 w-3" />
+              ) : (
+                <div className="h-1.5 w-1.5 rounded-full bg-current" />
+              )}
+              <span>{step.message}</span>
+              {step.time != null && <span className="opacity-50">{step.time.toFixed(1)}s</span>}
+            </div>
+          ))}
+
+          {/* 检索到的文档详情 */}
+          {retrievedStep?.details && retrievedStep.details.length > 0 && (
+            <div className="mt-1 space-y-1">
+              <div className="text-muted-foreground font-medium">参考文档：</div>
+              {retrievedStep.details.map((detail: string, i: number) => (
+                <div key={i} className="text-muted-foreground/80 pl-2 border-l border-border/50">
+                  {detail}
+                </div>
+              ))}
+            </div>
+          )}
+
+          {/* 推理内容 */}
+          {hasReasoning && (
+            <div className="mt-1">
+              <div className="text-muted-foreground font-medium mb-1">
+                {isReasoningNow ? '推理进行中...' : '推理过程：'}
+              </div>
+              <div className="text-foreground/80 whitespace-pre-wrap leading-relaxed bg-muted/30 rounded-lg p-2.5 max-h-80 overflow-y-auto">
+                {reasoningText}
+              </div>
+            </div>
+          )}
         </div>
-      ))}
+      )}
     </div>
   );
 };
@@ -131,7 +202,7 @@ export default function MessageItem({ message, selectedModel }: MessageItemProps
                 : "rounded-2xl rounded-tl-sm"
             )}>
               {isAssistant && message.thinking && (
-                <ThinkingProcess thinking={message.thinking} />
+                <ThinkingProcess thinking={message.thinking} isStreaming={message.content === ""} />
               )}
 
               <div className={cn(
@@ -208,10 +279,9 @@ export default function MessageItem({ message, selectedModel }: MessageItemProps
                 </>
               )}
               <span className="text-[10px] text-muted-foreground/50 mx-1">
-                {message.created_at ? formatDistanceToNow(
-                  new Date(new Date(message.created_at).getTime() + 8 * 60 * 60 * 1000),
-                  { addSuffix: true, locale: zhCN }
-                ) : ""}
+                {message.created_at
+                  ? format(new Date(new Date(message.created_at).getTime() + 8 * 60 * 60 * 1000), "HH:mm")
+                  : ""}
               </span>
             </div>
           </>

web/src/types/index.ts

@@ -69,10 +69,11 @@ export interface ChatSession {
 }
 
 export interface ThinkingStep {
-  stage: 'understanding' | 'retrieving' | 'retrieved' | 'generating';
+  stage: 'understanding' | 'retrieving' | 'retrieved' | 'generating' | 'reasoning' | 'preparing';
   message: string;
   doc_count?: number;
   time?: number;
+  details?: string[];
 }
 
 export interface ChatMessage {