From 05e33d1d05bdc88dcddb7ecfe8c743d766482c26 Mon Sep 17 00:00:00 2001
From: liuguancen <liuguancen@inspur.com>
Date: Thu, 2 Apr 2026 14:59:17 +0800
Subject: [PATCH] =?UTF-8?q?[=E5=89=8D=E7=AB=AF+RAG]=20=E4=BF=AE=E5=A4=8D?=
 =?UTF-8?q?=EF=BC=9A=E5=BC=82=E6=AD=A5=E4=B8=8A=E4=BC=A0+MySQL=E5=9B=9E?=
 =?UTF-8?q?=E5=86=99=E5=AF=BC=E8=AF=BB=EF=BC=9BPDF=E5=88=87=E6=8D=A2detach?=
 =?UTF-8?q?ed(Uint8Array=E5=A4=8D=E5=88=B6)=EF=BC=9B=E5=AF=B9=E8=AF=9D?=
 =?UTF-8?q?=E5=B7=A5=E5=85=B7=E6=A0=8F=E9=81=AE=E6=8C=A1=EF=BC=9B=E9=98=85?=
 =?UTF-8?q?=E8=AF=BB=E6=A8=A1=E5=BC=8F=E7=A9=BA=E7=99=BD?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 chat_web_front/src/components/PdfViewer.vue   |  12 +-
 .../src/components/ReadingCreate.vue          |  11 +-
 chat_web_front/src/views/reading/index.vue    |  17 +--
 .../server/knowledge_base/kb_doc_api.py       | 142 ++++++++++++------
 4 files changed, 120 insertions(+), 62 deletions(-)

diff --git a/chat_web_front/src/components/PdfViewer.vue b/chat_web_front/src/components/PdfViewer.vue
index 072f977..5d170eb 100644
--- a/chat_web_front/src/components/PdfViewer.vue
+++ b/chat_web_front/src/components/PdfViewer.vue
@@ -74,9 +74,17 @@ const loadPdf = async () => {
   loading.value = true;
   error.value = '';
   try {
+    // 每次加载都复制一份，避免 pdfjs worker transfer 后 detach 原始数据
     const rawSrc = toRaw(props.src);
-    const data = rawSrc instanceof ArrayBuffer ? new Uint8Array(rawSrc) : rawSrc;
-    const loadingTask = pdfjsLib.getDocument({ data });
+    let bytes: Uint8Array;
+    if (rawSrc instanceof Uint8Array) {
+      bytes = new Uint8Array(rawSrc);  // 复制
+    } else if (rawSrc instanceof ArrayBuffer) {
+      bytes = new Uint8Array(new Uint8Array(rawSrc));  // 复制
+    } else {
+      bytes = rawSrc as any;
+    }
+    const loadingTask = pdfjsLib.getDocument({ data: bytes });
     pdfDoc = await loadingTask.promise;
     const numPages = pdfDoc.numPages;
     pages.value = Array.from({ length: numPages }, (_, i) => i + 1);
diff --git a/chat_web_front/src/components/ReadingCreate.vue b/chat_web_front/src/components/ReadingCreate.vue
index 9af880a..5be0c74 100644
--- a/chat_web_front/src/components/ReadingCreate.vue
+++ b/chat_web_front/src/components/ReadingCreate.vue
@@ -434,7 +434,7 @@ const handleStop = async () => {
 
 <style lang="less" scoped>
 .message-content {
-  height: calc(100% - 255px);
+  height: calc(100% - 290px);
   overflow-y: auto;
   padding: 20px;
 
@@ -456,12 +456,9 @@ const handleStop = async () => {
 .tool-bar {
   display: flex;
   justify-content: space-between;
-  width: calc(100% - 15px);
-  padding-top: 30px;
-  height: 65px;
-  position:absolute;
-  bottom:275px;
-  background: linear-gradient( 360deg, #C0D4FD 0%, rgba(199,219,255,0) 100%);
+  width: 100%;
+  padding-top: 8px;
+  height: 40px;
   .label {
     display: flex;
     justify-content: space-around;
diff --git a/chat_web_front/src/views/reading/index.vue b/chat_web_front/src/views/reading/index.vue
index 8fff702..37920b3 100644
--- a/chat_web_front/src/views/reading/index.vue
+++ b/chat_web_front/src/views/reading/index.vue
@@ -110,7 +110,7 @@
           </div>
           <!-- PDF 原生渲染 -->
           <div v-if="fileType === 'pdf' && !readingMode" class="file-content" ref="fileContent" id="file-content">
-            <PdfViewer v-if="pdfData" :src="pdfData" :scale="1.3" />
+            <PdfViewer v-if="pdfBytes" :src="pdfBytes" :scale="1.3" />
           </div>
           <!-- HTML 阅读模式（PDF 阅读模式 + 非PDF文件） -->
           <div v-else class="file-content" ref="fileContent" id="file-content">
@@ -284,13 +284,6 @@ const docHtml = ref('');
 const fileContent = ref(null);
 const readingBox = ref(null);
 const pdfBytes = ref<Uint8Array | null>(null);  // 存原始字节，不会被 detach
-const pdfData = computed(() => {
-  // 每次访问时复制一份新的 ArrayBuffer 给 PdfViewer
-  if (!pdfBytes.value) return null;
-  const copy = new ArrayBuffer(pdfBytes.value.byteLength);
-  new Uint8Array(copy).set(pdfBytes.value);
-  return copy;
-});
 const readingMode = ref(false);
 const fileType = computed(() => {
   const name = selectedFile.value?.fileName || '';
@@ -915,8 +908,12 @@ onMounted(async () => {
       .view-md {
         padding: 20px;
         // 覆盖 PyMuPDF get_text("html") 输出的固定宽度
-        :deep(div[style*="width:"]) { width: 100% !important; max-width: 100% !important; }
-        :deep(.pdf-page > div) { width: 100% !important; }
+        :deep(div) { max-width: 100% !important; }
+        :deep(div[style*="width:"]) { width: auto !important; max-width: 100% !important; }
+        :deep(.pdf-page) { max-width: 100% !important; }
+        :deep(.pdf-page > div) { width: auto !important; max-width: 100% !important; }
+        :deep(.pdf-preview) { max-width: 100% !important; }
+        :deep(section) { max-width: 100% !important; }
         :deep(p) { font-size: 15px; line-height: 1.8rem; margin-block-start: 0; }
         :deep(.highlight) { background: #D0EAC8; }
         :deep(.note-flag) { width: 23px; height: 28px; line-height: 28px; display: inline-block; text-align: center; font-weight: bold; font-size: 10px; margin-left: 8px; cursor: pointer; background: url("@/assets/images/reading/note.png"); color: #004EA0; background-size: contain !important; background-repeat: no-repeat !important; background-position: center bottom !important; }
diff --git a/langchain-chat/server/knowledge_base/kb_doc_api.py b/langchain-chat/server/knowledge_base/kb_doc_api.py
index 7cecb55..e0d6ba9 100644
--- a/langchain-chat/server/knowledge_base/kb_doc_api.py
+++ b/langchain-chat/server/knowledge_base/kb_doc_api.py
@@ -269,6 +269,75 @@ 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,
+    embedding_ids: dict,
+):
+    """后台线程：LLM 导读 + 向量化，完成后直连 MySQL 更新结果。"""
+    import time
+    import pymysql
+    start_time = time.time()
+
+    kb = KBServiceFactory.get_service_by_name(knowledge_base_name)
+
+    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()
+
+            # 直连 MySQL 更新导读结果
+            embedding_id = embedding_ids.get(filename, filename)
+            try:
+                conn = pymysql.connect(**ck_mysql_config)
+                with conn.cursor() as cursor:
+                    cursor.execute(
+                        "UPDATE gpt_upload_file SET article_abstract=%s, article_keywords=%s, article_paragraph=%s WHERE embedding_id=%s",
+                        (
+                            str(llm_result.get("article_abstract", "生成摘要失败")),
+                            str(llm_result.get("article_keywords", "生成关键词失败")),
+                            str(llm_result.get("article_paragraph", "生成章节速览失败")),
+                            embedding_id
+                        )
+                    )
+                    conn.commit()
+                conn.close()
+                logger.info(f"[后台] LLM 导读已更新到数据库: {filename}")
+            except Exception as db_e:
+                logger.error(f"[后台] MySQL 更新失败 {filename}: {db_e}")
+
+        except Exception as e:
+            logger.error(f"[后台] LLM 导读生成失败 {filename}: {e}")
+
+    # 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()
+    except Exception as e:
+        logger.error(f"[后台] 向量化失败: {e}")
+    logger.info(f"[后台] 总耗时: {time.time() - start_time:.2f}s")
+
+
 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"]),
@@ -282,7 +351,7 @@ 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，提取全文后快速返回，LLM导读+向量化后台异步执行并直连MySQL更新结果
     """
     import time
     start_time = time.time()
@@ -303,64 +372,51 @@ def upload_docs_new(
     failed_files = {}
     file_names = list(docs.keys())
     llm_results = {}
+    embedding_ids = {}
 
-    # 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）
     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)
+        embedding_ids[filename] = filename
 
-        # 生成摘要、关键词、章节速览（模型已优化为 deepseek-v3）
         try:
             knowledge_file = KnowledgeFile(filename=filename, knowledge_base_name=knowledge_base_name)
-            import concurrent.futures
-            def run_async_in_thread():
-                new_loop = asyncio.new_event_loop()
-                asyncio.set_event_loop(new_loop)
-                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_data = knowledge_file.get_full_text()
+            import json as _json
+            try:
+                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，请稍后刷新查看...",
+                "article_keywords": "导读生成中，请稍后刷新查看...",
+                "article_paragraph": "导读生成中，请稍后刷新查看..."
             }
         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，请稍后刷新查看...",
+                "article_keywords": "导读生成中，请稍后刷新查看...",
+                "article_paragraph": "导读生成中，请稍后刷新查看..."
             }
 
-    # 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,
-        )
-        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={
+    # 后台异步：LLM 导读 + 向量化，完成后直连 MySQL 更新
+    import threading
+    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, embedding_ids),
+        daemon=True
+    ).start()
+
+    logger.info(f"文件上传+全文提取: {time.time() - start_time:.2f}s，LLM+向量化转后台")
+    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
     })