feat(思维导图助手)：思维导图助手增加智能整理总结-分批次整理

mineru/cli/fast_api.py

@@ -388,8 +388,117 @@ DEFAULT_MINDMAP_ORGANIZE_PROMPT = """你是文档结构整理助手。请基于
 8. 节点标题尽量简短，正文说明使用短句列表。
 9. 只输出 Markdown，不要输出解释、代码块围栏或额外说明。"""
 
+MINDMAP_MERGE_PROMPT = """你是 Markdown 思维导图结构校对助手。下面是多个分块已经整理总结后的局部 Markdown 大纲。
 
-def _call_mindmap_llm(markdown: str, mode: str = "smart", custom_prompt: Optional[str] = None, task_id: Optional[str] = None) -> str:
+任务：
+1. 合并这些局部大纲为一份完整 Markdown。
+2. 只检查和调整标题层级结构、顺序、重复标题和父子关系。
+3. 不要重新总结正文，不要扩写内容，不要新增原文没有的信息。
+4. 保留各局部大纲中的源语言：英文保持英文，中文保持中文，多语言分别保留。
+5. 最大层级不超过 4 层。
+6. 每个父节点下最多 8 个子节点，必要时只合并相近标题。
+7. 只输出 Markdown，不要输出解释、代码块围栏或额外说明。"""
+
+
+def _estimate_tokens(text: str) -> int:
+    ascii_chars = 0
+    non_ascii_chars = 0
+    for ch in text:
+        if ch.isspace():
+            continue
+        if ord(ch) < 128:
+            ascii_chars += 1
+        else:
+            non_ascii_chars += 1
+    return max(1, int(ascii_chars / 4) + int(non_ascii_chars * 1.5))
+
+
+def _get_mindmap_context_budget(prompt: str, reserve_output_tokens: int = 4096) -> tuple[int, int]:
+    max_context_tokens = int(os.getenv("MINDMAP_LLM_MAX_CONTEXT_TOKENS", "32768"))
+    prompt_tokens = _estimate_tokens(prompt)
+    safety_tokens = int(os.getenv("MINDMAP_LLM_SAFETY_TOKENS", "1024"))
+    input_budget_tokens = max(2048, max_context_tokens - prompt_tokens - reserve_output_tokens - safety_tokens)
+    logger.info(
+        "Mindmap context budget max_context_tokens={} prompt_tokens={} reserve_output_tokens={} safety_tokens={} input_budget_tokens={}",
+        max_context_tokens, prompt_tokens, reserve_output_tokens, safety_tokens, input_budget_tokens
+    )
+    return max_context_tokens, input_budget_tokens
+
+
+def _split_markdown_blocks(markdown: str) -> list[str]:
+    lines = markdown.splitlines()
+    blocks: list[str] = []
+    current: list[str] = []
+
+    heading_pattern = re.compile(r"^#{1,6}\s+")
+    for line in lines:
+        if heading_pattern.match(line) and current:
+            blocks.append("\n".join(current).strip())
+            current = [line]
+        else:
+            current.append(line)
+
+    if current:
+        blocks.append("\n".join(current).strip())
+
+    return [block for block in blocks if block]
+
+
+def _split_large_block(block: str, max_tokens: int) -> list[str]:
+    paragraphs = re.split(r"\n{2,}", block)
+    chunks: list[str] = []
+    current: list[str] = []
+    current_tokens = 0
+
+    for paragraph in paragraphs:
+        paragraph = paragraph.strip()
+        if not paragraph:
+            continue
+        paragraph_tokens = _estimate_tokens(paragraph)
+        if current and current_tokens + paragraph_tokens > max_tokens:
+            chunks.append("\n\n".join(current))
+            current = [paragraph]
+            current_tokens = paragraph_tokens
+        else:
+            current.append(paragraph)
+            current_tokens += paragraph_tokens
+
+    if current:
+        chunks.append("\n\n".join(current))
+    return chunks or [block]
+
+
+def _chunk_markdown_by_headings(markdown: str, max_tokens: int) -> list[str]:
+    blocks = _split_markdown_blocks(markdown)
+    chunks: list[str] = []
+    current: list[str] = []
+    current_tokens = 0
+
+    for block in blocks:
+        block_tokens = _estimate_tokens(block)
+        if block_tokens > max_tokens:
+            if current:
+                chunks.append("\n\n".join(current))
+                current = []
+                current_tokens = 0
+            chunks.extend(_split_large_block(block, max_tokens))
+            continue
+
+        if current and current_tokens + block_tokens > max_tokens:
+            chunks.append("\n\n".join(current))
+            current = [block]
+            current_tokens = block_tokens
+        else:
+            current.append(block)
+            current_tokens += block_tokens
+
+    if current:
+        chunks.append("\n\n".join(current))
+
+    return chunks or [markdown]
+
+
+def _call_mindmap_llm(markdown: str, mode: str = "smart", custom_prompt: Optional[str] = None, task_id: Optional[str] = None, request_role: str = "organize") -> str:
     base_url = os.getenv("MINDMAP_LLM_BASE_URL", "").rstrip("/")
     model = os.getenv("MINDMAP_LLM_MODEL", "gemma-4-26B")
     api_key = os.getenv("MINDMAP_LLM_API_KEY", "")
@@ -399,10 +508,6 @@ def _call_mindmap_llm(markdown: str, mode: str = "smart", custom_prompt: Optiona
         raise RuntimeError("未配置智能整理模型服务，请设置 MINDMAP_LLM_BASE_URL")
 
     compact_markdown = markdown.strip()
-    max_chars = int(os.getenv("MINDMAP_LLM_MAX_INPUT_CHARS", "30000"))
-    if len(compact_markdown) > max_chars:
-        compact_markdown = compact_markdown[:max_chars] + "\n\n...（后续内容已截断）"
-
     prompt_template = (custom_prompt or "").strip() or DEFAULT_MINDMAP_ORGANIZE_PROMPT
     prompt = f"""{prompt_template}
 
@@ -410,8 +515,8 @@ def _call_mindmap_llm(markdown: str, mode: str = "smart", custom_prompt: Optiona
 {compact_markdown}
 """
     logger.info(
-        "Mindmap LLM request start task_id={} model={} base_url={} mode={} input_chars={} prompt_chars={}",
+        "Mindmap LLM request start task_id={} role={} model={} base_url={} mode={} input_chars={} input_tokens_est={} prompt_chars={}",
-        task_id or "-", model, base_url, mode, len(compact_markdown), len(prompt_template)
+        task_id or "-", request_role, model, base_url, mode, len(compact_markdown), _estimate_tokens(compact_markdown), len(prompt_template)
     )
 
     payload = {
@@ -444,12 +549,60 @@ def _call_mindmap_llm(markdown: str, mode: str = "smart", custom_prompt: Optiona
     if not organized:
         raise RuntimeError("智能整理模型未返回有效内容")
     logger.info(
-        "Mindmap LLM request completed task_id={} output_chars={}",
+        "Mindmap LLM request completed task_id={} role={} output_chars={} output_tokens_est={}",
-        task_id or "-", len(organized)
+        task_id or "-", request_role, len(organized), _estimate_tokens(organized)
     )
     return organized
 
 
+def _organize_mindmap_markdown(markdown: str, mode: str, custom_prompt: Optional[str], task_id: str) -> str:
+    prompt_template = (custom_prompt or "").strip() or DEFAULT_MINDMAP_ORGANIZE_PROMPT
+    _, input_budget_tokens = _get_mindmap_context_budget(prompt_template)
+    source_tokens = _estimate_tokens(markdown)
+    logger.info(
+        "Mindmap organize strategy task_id={} source_chars={} source_tokens_est={} input_budget_tokens={}",
+        task_id, len(markdown), source_tokens, input_budget_tokens
+    )
+
+    if source_tokens <= input_budget_tokens:
+        _update_task_progress(task_id, 35, "调用智能整理模型")
+        return _call_mindmap_llm(markdown, mode, prompt_template, task_id, "single")
+
+    chunks = _chunk_markdown_by_headings(markdown, input_budget_tokens)
+    logger.info("Mindmap large input split task_id={} chunks={}", task_id, len(chunks))
+    partial_results: list[str] = []
+    for index, chunk in enumerate(chunks, start=1):
+        progress = 20 + int(index / max(len(chunks), 1) * 55)
+        _update_task_progress(task_id, progress, f"智能整理分块 {index}/{len(chunks)}")
+        logger.info(
+            "Mindmap chunk organize task_id={} chunk={}/{} chars={} tokens_est={}",
+            task_id, index, len(chunks), len(chunk), _estimate_tokens(chunk)
+        )
+        partial = _call_mindmap_llm(chunk, mode, prompt_template, task_id, f"chunk-{index}")
+        partial_results.append(partial)
+
+    merged_input = "\n\n".join(
+        f"<!-- chunk {index} -->\n{partial}"
+        for index, partial in enumerate(partial_results, start=1)
+    )
+    _, merge_budget_tokens = _get_mindmap_context_budget(MINDMAP_MERGE_PROMPT)
+    merge_tokens = _estimate_tokens(merged_input)
+    if merge_tokens > merge_budget_tokens:
+        logger.warning(
+            "Mindmap merged outline still exceeds context task_id={} tokens_est={} budget={} chunks={}",
+            task_id, merge_tokens, merge_budget_tokens, len(partial_results)
+        )
+        merge_chunks = _chunk_markdown_by_headings(merged_input, merge_budget_tokens)
+        merged_round: list[str] = []
+        for index, chunk in enumerate(merge_chunks, start=1):
+            _update_task_progress(task_id, 78 + int(index / max(len(merge_chunks), 1) * 10), f"合并局部大纲 {index}/{len(merge_chunks)}")
+            merged_round.append(_call_mindmap_llm(chunk, mode, MINDMAP_MERGE_PROMPT, task_id, f"merge-round-{index}"))
+        merged_input = "\n\n".join(merged_round)
+
+    _update_task_progress(task_id, 90, "全局整理标题结构")
+    return _call_mindmap_llm(merged_input, mode, MINDMAP_MERGE_PROMPT, task_id, "merge")
+
+
 async def _run_mindmap_organize_task(task_id: str, markdown: str, mode: str, prompt: Optional[str]):
     try:
         _store_task_progress(task_id, {
@@ -460,8 +613,7 @@ async def _run_mindmap_organize_task(task_id: str, markdown: str, mode: str, pro
             "file_names": "",
             "result_md": None,
         })
-        _update_task_progress(task_id, 35, "调用智能整理模型")
+        organized = await asyncio.to_thread(_organize_mindmap_markdown, markdown, mode, prompt, task_id)
-        organized = await asyncio.to_thread(_call_mindmap_llm, markdown, mode, prompt, task_id)
         state = _get_task_progress(task_id) or {}
         state.update({
             "progress": 100,

web_ui/src/components/ConfigPanel.vue

@@ -14,29 +14,6 @@
         />
       </el-form-item>
 
-      <el-form-item label="导出分辨率" class="form-item">
-        <el-select v-model="exportResolutionPreset" style="width: 100%" class="select">
-          <el-option
-            v-for="option in exportResolutionOptions"
-            :key="option.value"
-            :label="option.label"
-            :value="option.value"
-          />
-        </el-select>
-        <el-input-number
-          v-if="exportResolutionPreset === 'custom'"
-          v-model="draftConfig.exportResolution"
-          :min="720"
-          :max="16384"
-          :step="100"
-          controls-position="right"
-          class="export-resolution-input"
-        />
-        <div class="form-item-description">
-          PNG/JPEG 导出长边上限；默认 8K，导出文件会按该最大长边生效。
-        </div>
-      </el-form-item>
-      
       <!-- 解析后端 -->
       <el-form-item :label="$t('config.backend')" class="form-item">
         <el-select 
@@ -73,25 +50,6 @@
         </div>
       </el-form-item>
       
-      <div class="divider"></div>
-      
-      <div class="section-title">思维导图智能整理</div>
-
-      <el-form-item label="智能整理提示词" class="form-item">
-        <el-input
-          v-model="draftConfig.mindmapPrompt"
-          type="textarea"
-          :rows="10"
-          resize="vertical"
-          class="input"
-        />
-        <div class="form-item-description">
-          切换到“智能整理”时发送给后端大模型；默认保留标题结构、总结段落要点，并保持原文语言。
-        </div>
-      </el-form-item>
-
-      <div class="divider"></div>
-
       <!-- 识别选项 -->
       <div class="section-title">{{ $t('config.recognitionOptions') }}</div>
       
@@ -144,6 +102,46 @@
           </div>
         </el-form-item>
       </template>
+
+      <div class="divider"></div>
+
+      <div class="section-title">导出与智能整理</div>
+
+      <el-form-item label="导出分辨率" class="form-item">
+        <el-select v-model="exportResolutionPreset" style="width: 100%" class="select">
+          <el-option
+            v-for="option in exportResolutionOptions"
+            :key="option.value"
+            :label="option.label"
+            :value="option.value"
+          />
+        </el-select>
+        <el-input-number
+          v-if="exportResolutionPreset === 'custom'"
+          v-model="draftConfig.exportResolution"
+          :min="720"
+          :max="16384"
+          :step="100"
+          controls-position="right"
+          class="export-resolution-input"
+        />
+        <div class="form-item-description">
+          PNG/JPEG 导出长边上限；默认 8K，导出文件会按该最大长边生效。
+        </div>
+      </el-form-item>
+
+      <el-form-item label="智能整理提示词" class="form-item">
+        <el-input
+          v-model="draftConfig.mindmapPrompt"
+          type="textarea"
+          :rows="10"
+          resize="vertical"
+          class="input"
+        />
+        <div class="form-item-description">
+          切换到“智能整理”时发送给后端大模型；默认保留标题结构、总结段落要点，并保持原文语言。
+        </div>
+      </el-form-item>
       <div class="config-actions">
         <el-button @click="resetDraft">重置</el-button>
         <el-button type="primary" @click="confirmConfig">确认</el-button>

web_ui/src/components/MindMapRenderer.vue

@@ -10,6 +10,7 @@
             type="primary" 
             size="small"
             class="action-button primary"
+            :disabled="!content"
           >
             <template #icon>
               <el-icon><Download /></el-icon>
@@ -456,6 +457,10 @@ const downloadMindMap = async (format: ExportFormat = 'svg') => {
 
 // 处理下载命令
 const handleDownload = (command: string) => {
+  if (!props.content) {
+    ElMessage.warning('暂无思维导图内容可下载')
+    return
+  }
   if (command === 'svg' || command === 'png' || command === 'jpeg') {
     downloadMindMap(command)
   }

web_ui/src/composables/useDocumentProcessor.ts

@@ -129,6 +129,8 @@ export function useDocumentProcessor() {
     const validFiles: File[] = []
     const file = files[0]
     uploadedFiles.value = []
+    results.value = null
+    error.value = null
     if (files.length > 1) {
       error.value = '一次只能选择一个文件，已使用第一个文件'
     }
@@ -138,16 +140,17 @@ export function useDocumentProcessor() {
     
     console.log('Processing file:', fileName, 'type:', fileType)
     
-    // 验证文件类型 - 支持 PDF、Word 和后端可转换为 PDF 的图片格式
+    // 验证文件类型 - 支持 Markdown、PDF、Word 和后端可转换为 PDF 的图片格式
     const supportedImageExtensions = ['.png', '.jpg', '.jpeg', '.webp', '.bmp', '.tiff', '.gif', '.jp2']
     const isImage = fileType.startsWith('image/') && supportedImageExtensions.some((ext) => fileName.endsWith(ext))
     const isPdf = fileType === 'application/pdf'
     const isWord = fileName.endsWith('.docx') || fileName.endsWith('.doc') || fileType === 'application/vnd.openxmlformats-officedocument.wordprocessingml.document'
+    const isMarkdown = fileName.endsWith('.md') || fileType === 'text/markdown'
     
-    console.log('File type checks - isImage:', isImage, 'isPdf:', isPdf, 'isWord:', isWord)
+    console.log('File type checks - isImage:', isImage, 'isPdf:', isPdf, 'isWord:', isWord, 'isMarkdown:', isMarkdown)
     
-    if (!isImage && !isPdf && !isWord) {
+    if (!isImage && !isPdf && !isWord && !isMarkdown) {
-      error.value = '不支持的文件类型，请选择 PDF、Word、JPG、PNG、WEBP、BMP、TIFF、GIF 或 JP2 文件'
+      error.value = '不支持的文件类型，请选择 MD、PDF、Word、JPG、PNG、WEBP、BMP、TIFF、GIF 或 JP2 文件'
       console.log('Unsupported file type:', fileName, fileType)
       return
     }
@@ -159,8 +162,26 @@ export function useDocumentProcessor() {
       return
     }
     
-    // 如果是 Word 文档，转换为 PDF
+    if (isMarkdown) {
-    if (isWord) {
+      try {
+        isUploading.value = true
+        const markdown = await file.text()
+        validFiles.push(file)
+        results.value = {
+          markdown,
+          source: markdown,
+          mindmap: buildMindmapMarkdown(markdown)
+        }
+        error.value = null
+        console.log('Markdown file loaded directly:', fileName, markdown.length)
+      } catch (err) {
+        error.value = 'Markdown 文件读取失败: ' + (err as Error).message
+        console.error('Markdown read failed:', (err as Error).message)
+        return
+      } finally {
+        isUploading.value = false
+      }
+    } else if (isWord) {
       try {
         isUploading.value = true
         error.value = '正在将 Word 文档转换为 PDF...'
@@ -242,6 +263,14 @@ export function useDocumentProcessor() {
       return
     }
 
+    const firstFileName = uploadedFiles.value[0]?.name.toLowerCase() || ''
+    if (firstFileName.endsWith('.md')) {
+      if (!results.value) {
+        error.value = 'Markdown 文件内容为空，请重新上传'
+      }
+      return
+    }
+    
     isProcessing.value = true
     error.value = null
     processingStage.value = '准备提交解析任务'

web_ui/src/views/DocumentProcessor.vue

@@ -21,7 +21,7 @@
             <el-icon class="upload-icon"><Folder /></el-icon>
             <div class="upload-copy">
               <div class="upload-text">文件导入</div>
-              <div class="upload-hint">支持 PDF、Word、JPG、PNG、WEBP、BMP、TIFF、GIF、JP2 单个文件</div>
+              <div class="upload-hint">支持 MD、PDF、Word、JPG、PNG、BMP、TIFF 等格式文件</div>
             </div>
           </div>
 
@@ -201,7 +201,7 @@
     <input
       ref="fileInput"
       type="file"
-      accept=".pdf,.doc,.docx,.png,.jpg,.jpeg,.webp,.bmp,.tiff,.gif,.jp2"
+      accept=".md,.pdf,.doc,.docx,.png,.jpg,.jpeg,.webp,.bmp,.tiff,.gif,.jp2"
       style="position: absolute; width: 0; height: 0; overflow: hidden;"
       @change="handleFileInputChange"
     />
@@ -618,7 +618,9 @@ const handleFileInputChange = async (event: Event) => {
     if (uploadedFiles.value.length > 0) {
       resetSmartOrganizeState()
       isUploadAreaCollapsed.value = true
-      initializeManualResult()
+      if (!results.value) {
+        initializeManualResult()
+      }
       markdownRenderMode.value = 'markdown'
       sourceViewMode.value = 'markdown'
       activeTab.value = 'source'
@@ -635,7 +637,9 @@ const handleDrop = async (event: DragEvent) => {
     if (uploadedFiles.value.length > 0) {
       resetSmartOrganizeState()
       isUploadAreaCollapsed.value = true
-      initializeManualResult()
+      if (!results.value) {
+        initializeManualResult()
+      }
       markdownRenderMode.value = 'markdown'
       sourceViewMode.value = 'markdown'
       activeTab.value = 'source'
@@ -733,14 +737,16 @@ onUnmounted(() => {
   background-color: #FFFFFF;
   margin-bottom: 0;
   overflow: hidden;
-  min-height: 36px;
+  height: 52px;
+  min-height: 52px;
   display: flex;
   align-items: center;
+  box-sizing: border-box;
 }
 
 .drag-upload-area.collapsed {
   padding: 8px 16px;
-  min-height: 36px;
+  min-height: 52px;
 }
 
 .drag-upload-area:hover,
@@ -782,6 +788,7 @@ onUnmounted(() => {
 
 .uploaded-files {
   width: 100%;
+  height: 100%;
   text-align: left;
 }
 
@@ -795,6 +802,7 @@ onUnmounted(() => {
   margin-bottom: 0;
   transition: all 0.3s ease;
   width: 100%;
+  height: 100%;
   box-sizing: border-box;
   font-size: 13px;
 }