[前端+RAG] 修复PDF文字重叠；上传异步化(LLM+向量化后台执行)；摘要关键词模型改为deepseek-v3

2026-04-02 14:10:08 +08:00
parent 0e25154468
commit 5158753b94
3 changed files with 122 additions and 116 deletions
--- a/chat_web_front/src/components/PdfViewer.vue
+++ b/chat_web_front/src/components/PdfViewer.vue
@@ -1,12 +1,12 @@
 <template>
  <div class="pdf-viewer" ref="containerRef">
    <div v-for="page in pages" :key="page" class="pdf-page-wrapper">
-      <div class="pdf-page" :id="'pdf-page-' + page" :style="{ position: 'relative' }">
+      <div class="pdf-page" :style="{ width: pageWidths[page] + 'px', height: pageHeights[page] + 'px' }">
        <canvas :ref="el => setCanvasRef(el, page)"></canvas>
        <div class="text-layer" :ref="el => setTextLayerRef(el, page)"></div>
      </div>
    </div>
-    <div v-if="loading" class="pdf-loading">加载中...</div>
+    <div v-if="loading" class="pdf-loading">PDF 加载中...</div>
    <div v-if="error" class="pdf-error">{{ error }}</div>
  </div>
 </template>
@@ -14,8 +14,8 @@
 <script setup lang="ts">
 import { ref, onMounted, onBeforeUnmount, nextTick, watch } from 'vue';
 import * as pdfjsLib from 'pdfjs-dist';
+import { renderTextLayer } from 'pdfjs-dist';

-// Set worker
 pdfjsLib.GlobalWorkerOptions.workerSrc = new URL(
  'pdfjs-dist/build/pdf.worker.min.js',
  import.meta.url
@@ -28,18 +28,16 @@ const props = defineProps<{

 const containerRef = ref<HTMLElement | null>(null);
 const pages = ref<number[]>([]);
+const pageWidths = ref<Record<number, number>>({});
+const pageHeights = ref<Record<number, number>>({});
 const loading = ref(true);
 const error = ref('');
 const canvasRefs: Record<number, HTMLCanvasElement> = {};
 const textLayerRefs: Record<number, HTMLElement> = {};
 let pdfDoc: any = null;

-const setCanvasRef = (el: any, page: number) => {
-  if (el) canvasRefs[page] = el;
-};
-const setTextLayerRef = (el: any, page: number) => {
-  if (el) textLayerRefs[page] = el;
-};
+const setCanvasRef = (el: any, page: number) => { if (el) canvasRefs[page] = el; };
+const setTextLayerRef = (el: any, page: number) => { if (el) textLayerRefs[page] = el; };

 const renderPage = async (pageNum: number) => {
  if (!pdfDoc) return;
@@ -47,43 +45,28 @@ const renderPage = async (pageNum: number) => {
  const scale = props.scale || 1.5;
  const viewport = page.getViewport({ scale });

+  // Canvas rendering
  const canvas = canvasRefs[pageNum];
  if (!canvas) return;
  const context = canvas.getContext('2d');
  canvas.height = viewport.height;
  canvas.width = viewport.width;
+  pageWidths.value[pageNum] = viewport.width;
+  pageHeights.value[pageNum] = viewport.height;

  await page.render({ canvasContext: context, viewport }).promise;

-  // Text layer for text selection
+  // Text layer using pdfjs built-in API
  const textLayerDiv = textLayerRefs[pageNum];
  if (textLayerDiv) {
-    textLayerDiv.style.width = viewport.width + 'px';
-    textLayerDiv.style.height = viewport.height + 'px';
    textLayerDiv.innerHTML = '';
-
    const textContent = await page.getTextContent();
-    const textItems = textContent.items;
-
-    for (const item of textItems) {
-      if (!item.str) continue;
-      const tx = pdfjsLib.Util.transform(viewport.transform, item.transform);
-      const span = document.createElement('span');
-      span.textContent = item.str;
-      span.style.position = 'absolute';
-      span.style.left = tx[4] + 'px';
-      span.style.top = (viewport.height - tx[5]) + 'px';
-      span.style.fontSize = Math.abs(tx[0]) + 'px';
-      span.style.fontFamily = item.fontName || 'sans-serif';
-      span.style.transformOrigin = '0% 0%';
-      // Width matching
-      if (item.width) {
-        const textWidth = item.width * scale;
-        span.style.width = textWidth + 'px';
-        span.style.display = 'inline-block';
-      }
-      textLayerDiv.appendChild(span);
-    }
+    renderTextLayer({
+      textContentSource: textContent,
+      container: textLayerDiv,
+      viewport: viewport,
+      textDivs: []
+    });
  }
 };

@@ -97,33 +80,20 @@ const loadPdf = async () => {
    pdfDoc = await loadingTask.promise;
    const numPages = pdfDoc.numPages;
    pages.value = Array.from({ length: numPages }, (_, i) => i + 1);
-
    await nextTick();
    for (let i = 1; i <= numPages; i++) {
      await renderPage(i);
    }
  } catch (e: any) {
    error.value = 'PDF 加载失败: ' + (e.message || e);
-    console.error('PDF load error:', e);
  } finally {
    loading.value = false;
  }
 };

-watch(() => props.src, () => {
-  if (props.src) loadPdf();
-});
-
-onMounted(() => {
-  if (props.src) loadPdf();
-});
-
-onBeforeUnmount(() => {
-  if (pdfDoc) {
-    pdfDoc.destroy();
-    pdfDoc = null;
-  }
-});
+watch(() => props.src, () => { if (props.src) loadPdf(); });
+onMounted(() => { if (props.src) loadPdf(); });
+onBeforeUnmount(() => { if (pdfDoc) { pdfDoc.destroy(); pdfDoc = null; } });
 </script>

 <style scoped>
@@ -131,54 +101,51 @@ onBeforeUnmount(() => {
  width: 100%;
  height: 100%;
  overflow-y: auto;
-  background: #f5f5f5;
+  background: #e8e8e8;
+  padding: 16px 0;
 }
-
 .pdf-page-wrapper {
  display: flex;
  justify-content: center;
-  margin-bottom: 16px;
+  margin-bottom: 12px;
 }
-
 .pdf-page {
  position: relative;
  background: white;
  box-shadow: 0 2px 8px rgba(0, 0, 0, 0.15);
 }
-
 .pdf-page canvas {
  display: block;
 }
-
 .text-layer {
  position: absolute;
  top: 0;
  left: 0;
+  right: 0;
+  bottom: 0;
  overflow: hidden;
-  opacity: 0.3;
  line-height: 1;
 }
-
-.text-layer span {
+/* pdfjs renderTextLayer creates spans with absolute positioning */
+.text-layer :deep(span) {
  color: transparent;
  position: absolute;
  white-space: pre;
  cursor: text;
+  transform-origin: 0% 0%;
 }
-
-.text-layer span::selection {
+.text-layer :deep(span::selection) {
  background: rgba(0, 78, 160, 0.3);
  color: transparent;
 }
-
+.text-layer :deep(br) {
+  display: none;
+}
 .pdf-loading, .pdf-error {
  text-align: center;
  padding: 40px;
  color: #999;
  font-size: 14px;
 }
-
-.pdf-error {
-  color: #c00;
-}
+.pdf-error { color: #c00; }
 </style>
--- a/langchain-chat/server/knowledge_base/kb_doc_api.py
+++ b/langchain-chat/server/knowledge_base/kb_doc_api.py
@@ -269,6 +269,62 @@ def upload_docs(
    return BaseResponse(code=200, msg="文件上传与向量化完成", data={"failed_files": failed_files})


+def _background_llm_and_vectorize(
+    knowledge_base_name: str,
+    file_names: List[str],
+    chunk_size: int,
+    chunk_overlap: int,
+    zh_title_enhance: bool,
+    docs: dict,
+    not_refresh_vs_cache: bool,
+):
+    """后台线程：执行 LLM 导读生成 + 向量化，不阻塞上传响应。"""
+    import time
+    start_time = time.time()
+
+    kb = KBServiceFactory.get_service_by_name(knowledge_base_name)
+
+    # 1. 生成 LLM 导读（摘要、关键词、章节速览）
+    for filename in file_names:
+        try:
+            knowledge_file = KnowledgeFile(filename=filename, knowledge_base_name=knowledge_base_name)
+            new_loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(new_loop)
+            try:
+                llm_result = new_loop.run_until_complete(knowledge_file.get_llm_result())
+            finally:
+                new_loop.close()
+
+            # 将 LLM 结果写入缓存文件，供 Java 后端轮询读取
+            import json
+            cache_dir = os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(__file__))), "knowledge_base", knowledge_base_name)
+            os.makedirs(cache_dir, exist_ok=True)
+            cache_file = os.path.join(cache_dir, f"{filename}.llm_result.json")
+            with open(cache_file, 'w', encoding='utf-8') as f:
+                json.dump(llm_result, f, ensure_ascii=False)
+            logger.info(f"[后台] LLM 导读生成完成: {filename}")
+        except Exception as e:
+            logger.error(f"[后台] LLM 导读生成失败 {filename}: {e}")
+
+    # 2. 向量化
+    try:
+        _update_docs_impl(
+            knowledge_base_name=knowledge_base_name,
+            file_names=file_names,
+            override_custom_docs=True,
+            chunk_size=chunk_size,
+            chunk_overlap=chunk_overlap,
+            zh_title_enhance=zh_title_enhance,
+            docs=docs,
+            not_refresh_vs_cache=True,
+        )
+        if kb and not not_refresh_vs_cache:
+            kb.save_vector_store()
+        logger.info(f"[后台] 向量化完成，总耗时: {time.time() - start_time:.2f}s")
+    except Exception as e:
+        logger.error(f"[后台] 向量化失败: {e}")
+
+
 def upload_docs_new(
        files: List[UploadFile] = File(..., description="上传文件，支持多文件"),
        knowledge_base_name: str = Form(..., description="知识库名称", examples=["samples"]),
@@ -282,16 +338,15 @@ def upload_docs_new(
        not_refresh_vs_cache: bool = Form(False, description="暂不保存向量库（用于FAISS）"),
 ) -> BaseResponse:
    """
-    API接口：上传文件，并/或向量化
+    API接口：上传文件，先提取全文快速返回，LLM导读+向量化后台异步执行
    """
-    import time  # 添加计时模块
+    import time
    start_time = time.time()
    if not validate_kb_name(knowledge_base_name):
        return BaseResponse(code=403, msg="Don't attack me")

    kb = KBServiceFactory.get_service_by_name(knowledge_base_name)
    if kb is None:
-        # 自动创建知识库
        kb = KBServiceFactory.get_service(knowledge_base_name, DEFAULT_VS_TYPE, EMBEDDING_MODEL)
        try:
            kb.create_kb()
@@ -303,68 +358,52 @@ def upload_docs_new(

    failed_files = {}
    file_names = list(docs.keys())
-
-    # 生成摘要、关键词、章节速览的结果存储
    llm_results = {}

-    # 先将上传的文件保存到磁盘
+    # 保存文件到磁盘 + 提取全文（快速操作）
    for result in _save_files_in_thread(files, knowledge_base_name=knowledge_base_name, override=override):
        filename = result["data"]["file_name"]
        if result["code"] != 200:
            failed_files[filename] = result["msg"]
-
        if filename not in file_names:
            file_names.append(filename)

-        # 针对成功上传的文件，生成摘要、关键词、章节速览
+        # 仅提取全文（快速），不调用 LLM
        try:
            knowledge_file = KnowledgeFile(filename=filename, knowledge_base_name=knowledge_base_name)
-            # 使用线程池运行异步函数，避免事件循环冲突
-            import concurrent.futures
-            def run_async_in_thread():
-                new_loop = asyncio.new_event_loop()
-                asyncio.set_event_loop(new_loop)
+            full_text_data = knowledge_file.get_full_text()
+            import json as _json
            try:
-                    return new_loop.run_until_complete(knowledge_file.get_llm_result())
-                finally:
-                    new_loop.close()
-
-            with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
-                future = executor.submit(run_async_in_thread)
-                llm_result = future.result()
+                full_text = _json.loads(full_text_data).get("full_text", "")
+            except:
+                full_text = ""
            llm_results[filename] = {
-                "full_text": llm_result.get("full_text", "获取全文失败"),
-                "article_abstract": llm_result.get("article_abstract", "生成摘要失败"),
-                "article_keywords": llm_result.get("article_keywords", "生成关键词失败"),
-                "article_paragraph": llm_result.get("article_paragraph", "生成章节速览失败")
+                "full_text": full_text,
+                "article_abstract": "导读生成中...",
+                "article_keywords": "导读生成中...",
+                "article_paragraph": "导读生成中..."
            }
        except Exception as e:
-            logger.error(f"生成LLM结果时出错：{e}", exc_info=e if log_verbose else None)
+            logger.error(f"提取全文失败 {filename}: {e}")
            llm_results[filename] = {
-                "article_abstract": "生成摘要失败",
-                "article_keywords": "生成关键词失败",
-                "article_paragraph": "生成章节速览失败"
+                "full_text": "",
+                "article_abstract": "导读生成中...",
+                "article_keywords": "导读生成中...",
+                "article_paragraph": "导读生成中..."
            }

-    # 对保存的文件进行向量化
-    if to_vector_store:
-        update_st = time.time()
-        result = _update_docs_impl(
-            knowledge_base_name=knowledge_base_name,
-            file_names=file_names,
-            override_custom_docs=True,
-            chunk_size=chunk_size,
-            chunk_overlap=chunk_overlap,
-            zh_title_enhance=zh_title_enhance,
-            docs=docs,
-            not_refresh_vs_cache=True,
+    # 后台异步执行 LLM 导读 + 向量化（不阻塞响应）
+    import threading
+    bg_thread = threading.Thread(
+        target=_background_llm_and_vectorize,
+        args=(knowledge_base_name, file_names, chunk_size, chunk_overlap,
+              zh_title_enhance, docs, not_refresh_vs_cache),
+        daemon=True
    )
-        failed_files.update(result.data["failed_files"])
-        if not not_refresh_vs_cache:
-            kb.save_vector_store()
-        logger.info(f'向量化用时：{time.time() - update_st}')
-    logger.info(f"总执行时间: {time.time() - start_time:.2f}s")
-    return BaseResponse(code=200, msg="文件上传与向量化完成", data={
+    bg_thread.start()
+
+    logger.info(f"文件上传+全文提取用时: {time.time() - start_time:.2f}s，LLM+向量化已转后台")
+    return BaseResponse(code=200, msg="文件上传完成，导读生成中", data={
        "failed_files": failed_files,
        "llm_results": llm_results
    })
--- a/langchain-chat/server/knowledge_base/utils.py
+++ b/langchain-chat/server/knowledge_base/utils.py
@@ -390,7 +390,7 @@ class KnowledgeFile:
            llm_time = time.time()
            abstract_task = get_llm_model_response_async(
                strategy_name="gen_abstract",
-                llm_model_name=LLM_MODELS[1],
+                llm_model_name=LLM_MODELS[0],
                template_prompt_name="gen_abstract",
                prompt_param_dict={"context": full_text},
                temperature=0.7,
@@ -399,7 +399,7 @@ class KnowledgeFile:

            keywords_task = get_llm_model_response_async(
                strategy_name="gen_keywords",
-                llm_model_name=LLM_MODELS[1],
+                llm_model_name=LLM_MODELS[0],
                template_prompt_name="gen_keywords",
                prompt_param_dict={"context": full_text},
                temperature=0.7,