fix(avatar): follow language changes each turn

Reviewed-on: #19

digital-avatar-app/backend/routers/chat.py

@@ -478,6 +478,24 @@ def _qa_requires_language_adaptation(question: str, answer: str) -> bool:
     )
 
 
+def _qa_requires_per_turn_rendering(
+    question: str,
+    answer: str,
+    history: list[Any],
+) -> bool:
+    """Keep the direct QA fast path only when no conversation can bias language."""
+    return bool(history) or _qa_requires_language_adaptation(question, answer)
+
+
+def _per_turn_language_instruction() -> str:
+    return (
+        "本轮语言覆盖指令：只根据紧随其后的最新用户消息判断本轮回答语言。"
+        "即使此前整段对话一直使用另一种语言，只要最新消息切换了语言，本轮就必须立即切换到相同语言；"
+        "不要沿用上一轮语言。若最新消息明确指定回答语言，以该指定为准；若混用多种语言，使用其中占主导的"
+        "自然语言。不要说明你检测、切换或翻译了语言。"
+    )
+
+
 def _canonicalize_question(value: str) -> str:
     value = _normalize_question(value)
     replacements = (
@@ -709,6 +727,8 @@ def _build_prompt(
     messages = [{"role": "system", "content": system}]
     for item in history[-MAX_HISTORY_MESSAGES:]:
         messages.append({"role": item.role, "content": item.content} if hasattr(item, "role") else item)
+    # Keep the language instruction adjacent to the current turn so long histories cannot override it.
+    messages.append({"role": "system", "content": _per_turn_language_instruction()})
     messages.append({"role": "user", "content": question.strip()})
     return messages
 
@@ -845,7 +865,7 @@ def _resolve_reply(
         qa_pairs = db.query(QAPair).filter(QAPair.avatar_id == avatar.id).all()
     matched = _match_standard_qa(question, qa_pairs)
     adapt_qa_language = bool(
-        matched and _qa_requires_language_adaptation(question, matched.answer)
+        matched and _qa_requires_per_turn_rendering(question, matched.answer, history)
     )
     if matched and not adapt_qa_language and not image_contexts:
         return {"answer": matched.answer, "source": "qa", "references": []}
@@ -940,7 +960,7 @@ def _stream_reply(
     qa_pairs = db.query(QAPair).filter(QAPair.avatar_id == avatar.id).all()
     matched = _match_standard_qa(question, qa_pairs)
     adapt_qa_language = bool(
-        matched and _qa_requires_language_adaptation(question, matched.answer)
+        matched and _qa_requires_per_turn_rendering(question, matched.answer, history)
     )
     messages, reservation = [], None
     if matched and not adapt_qa_language and not image_contexts:

digital-avatar-app/backend/routers/knowledge.py

@@ -1,4 +1,7 @@
 import os
+import json
+import shutil
+import time
 import uuid
 
 from fastapi import APIRouter, UploadFile, File, Depends, Header, HTTPException
@@ -19,6 +22,9 @@ os.makedirs(UPLOAD_DIR, exist_ok=True)
 ALLOWED_EXT = {".md", ".txt", ".pdf", ".doc", ".docx", ".xlsx"}
 MAX_UPLOAD_BYTES = 50 * 1024 * 1024
 UPLOAD_CHUNK_BYTES = 1024 * 1024
+MULTIPART_CHUNK_BYTES = 5 * 1024 * 1024
+MULTIPART_ROOT = ".multipart"
+MULTIPART_TTL_SECONDS = 24 * 60 * 60
 
 
 class QAIn(BaseModel):
@@ -31,6 +37,74 @@ class EnabledIn(BaseModel):
     enabled: bool = True
 
 
+class MultipartUploadIn(BaseModel):
+    filename: str
+    fileSize: int
+    totalChunks: int
+
+
+def _validate_document(filename: str, file_size: int):
+    ext = os.path.splitext(filename or "")[1].lower()
+    if ext not in ALLOWED_EXT:
+        return None, f"不支持的文件类型：{ext or '空'}，仅支持 md/txt/pdf/doc/docx/xlsx"
+    if file_size <= 0:
+        return None, "文件内容不能为空"
+    if file_size > MAX_UPLOAD_BYTES:
+        return None, "文件不能超过 50MB"
+    return ext, ""
+
+
+def _multipart_dir(avatar_id: str, upload_id: str) -> str:
+    safe_avatar_id = os.path.basename(avatar_id)
+    safe_upload_id = os.path.basename(upload_id)
+    if (
+        safe_avatar_id != avatar_id
+        or safe_upload_id != upload_id
+        or len(upload_id) != 32
+        or any(character not in "0123456789abcdef" for character in upload_id)
+    ):
+        raise HTTPException(status_code=400, detail="上传标识无效")
+    return os.path.join(UPLOAD_DIR, MULTIPART_ROOT, safe_avatar_id, safe_upload_id)
+
+
+def _purge_stale_multipart_uploads(avatar_id: str):
+    avatar_upload_root = os.path.join(UPLOAD_DIR, MULTIPART_ROOT, os.path.basename(avatar_id))
+    if not os.path.isdir(avatar_upload_root):
+        return
+    cutoff = time.time() - MULTIPART_TTL_SECONDS
+    for entry in os.scandir(avatar_upload_root):
+        if entry.is_dir(follow_symlinks=False) and entry.stat(follow_symlinks=False).st_mtime < cutoff:
+            shutil.rmtree(entry.path, ignore_errors=True)
+
+
+def _read_multipart_metadata(avatar_id: str, upload_id: str) -> tuple[str, dict]:
+    upload_dir = _multipart_dir(avatar_id, upload_id)
+    metadata_path = os.path.join(upload_dir, "metadata.json")
+    if not os.path.isfile(metadata_path):
+        raise HTTPException(status_code=404, detail="上传任务不存在或已过期")
+    with open(metadata_path, "r", encoding="utf-8") as stream:
+        return upload_dir, json.load(stream)
+
+
+def _create_knowledge_doc(db: Session, avatar_id: str, filename: str, ext: str, file_size: int, stored: str):
+    doc = KnowledgeDoc(
+        id=uuid.uuid4().hex,
+        avatar_id=avatar_id,
+        filename=filename,
+        file_type=ext.lstrip("."),
+        file_size=file_size,
+        file_url=f"/api/files/{avatar_id}/{stored}",
+        status="parsing",
+        index_stage="queued",
+        index_progress=0,
+    )
+    db.add(doc)
+    db.commit()
+    db.refresh(doc)
+    knowledge_vectorizer.enqueue(doc.id)
+    return doc
+
+
 def _doc_payload(doc: KnowledgeDoc) -> dict:
     payload = doc.to_dict()
     stored_name = os.path.basename(doc.file_url or "")
@@ -74,9 +148,9 @@ def list_docs(avatar_id: str, authorization: str = Header(None), db: Session = D
 @router.post("/avatar/{avatar_id}/knowledge/docs")
 async def upload_doc(avatar_id: str, file: UploadFile = File(...), authorization: str = Header(None), db: Session = Depends(get_db)):
     _require_owned_avatar(db, avatar_id, authorization)
-    ext = os.path.splitext(file.filename or "")[1].lower()
-    if ext not in ALLOWED_EXT:
-        return fail(f"不支持的文件类型：{ext or '空'}，仅支持 md/txt/pdf/doc/docx/xlsx", code=400)
+    ext, validation_error = _validate_document(file.filename or "", 1)
+    if validation_error:
+        return fail(validation_error, code=400)
     avatar_dir = os.path.join(UPLOAD_DIR, avatar_id)
     os.makedirs(avatar_dir, exist_ok=True)
     stored = f"{uuid.uuid4().hex}{ext}"
@@ -94,25 +168,124 @@ async def upload_doc(avatar_id: str, file: UploadFile = File(...), authorization
         if os.path.exists(path):
             os.remove(path)
         return fail(str(exc), code=400)
-    doc = KnowledgeDoc(
-        id=uuid.uuid4().hex,
-        avatar_id=avatar_id,
-        filename=file.filename,
-        file_type=ext.lstrip("."),
-        file_size=file_size,
-        file_url=f"/api/files/{avatar_id}/{stored}",
-        status="parsing",
-        index_stage="queued",
-        index_progress=0,
+    if file_size == 0:
+        if os.path.exists(path):
+            os.remove(path)
+        return fail("文件内容不能为空", code=400)
+
+    doc = _create_knowledge_doc(db, avatar_id, file.filename or stored, ext, file_size, stored)
+    return ok(_doc_payload(doc))
+
+
+@router.post("/avatar/{avatar_id}/knowledge/uploads")
+def create_multipart_upload(
+    avatar_id: str,
+    body: MultipartUploadIn,
+    authorization: str = Header(None),
+    db: Session = Depends(get_db),
+):
+    _require_owned_avatar(db, avatar_id, authorization)
+    ext, validation_error = _validate_document(body.filename, body.fileSize)
+    if validation_error:
+        return fail(validation_error, code=400)
+    expected_chunks = (body.fileSize + MULTIPART_CHUNK_BYTES - 1) // MULTIPART_CHUNK_BYTES
+    if body.totalChunks != expected_chunks:
+        return fail("文件分片数量不正确", code=400)
+
+    _purge_stale_multipart_uploads(avatar_id)
+    upload_id = uuid.uuid4().hex
+    upload_dir = _multipart_dir(avatar_id, upload_id)
+    os.makedirs(upload_dir, exist_ok=False)
+    metadata = {
+        "filename": body.filename,
+        "fileSize": body.fileSize,
+        "totalChunks": body.totalChunks,
+        "extension": ext,
+    }
+    with open(os.path.join(upload_dir, "metadata.json"), "w", encoding="utf-8") as stream:
+        json.dump(metadata, stream, ensure_ascii=False)
+    return ok({"uploadId": upload_id, "chunkSize": MULTIPART_CHUNK_BYTES})
+
+
+@router.post("/avatar/{avatar_id}/knowledge/uploads/{upload_id}/chunks/{chunk_index}")
+async def upload_multipart_chunk(
+    avatar_id: str,
+    upload_id: str,
+    chunk_index: int,
+    file: UploadFile = File(...),
+    authorization: str = Header(None),
+    db: Session = Depends(get_db),
+):
+    _require_owned_avatar(db, avatar_id, authorization)
+    upload_dir, metadata = _read_multipart_metadata(avatar_id, upload_id)
+    total_chunks = int(metadata["totalChunks"])
+    if chunk_index < 0 or chunk_index >= total_chunks:
+        return fail("文件分片序号不正确", code=400)
+
+    expected_size = min(
+        MULTIPART_CHUNK_BYTES,
+        int(metadata["fileSize"]) - chunk_index * MULTIPART_CHUNK_BYTES,
     )
+    part_path = os.path.join(upload_dir, f"{chunk_index}.part")
+    temporary_path = f"{part_path}.uploading"
+    received = 0
+    try:
+        with open(temporary_path, "wb") as stream:
+            while chunk := await file.read(UPLOAD_CHUNK_BYTES):
+                received += len(chunk)
+                if received > expected_size:
+                    raise ValueError("文件分片大小不正确")
+                stream.write(chunk)
+        if received != expected_size:
+            raise ValueError("文件分片大小不正确")
+        os.replace(temporary_path, part_path)
+    except ValueError as exc:
+        if os.path.exists(temporary_path):
+            os.remove(temporary_path)
+        return fail(str(exc), code=400)
+    return ok({"chunkIndex": chunk_index, "uploadedBytes": received})
 
-    # Persist and acknowledge the upload first. Extraction and embeddings may take
-    # minutes for a PDF and must never consume the browser request timeout.
-    db.add(doc)
-    db.commit()
-    db.refresh(doc)
-    knowledge_vectorizer.enqueue(doc.id)
 
+@router.post("/avatar/{avatar_id}/knowledge/uploads/{upload_id}/complete")
+def complete_multipart_upload(
+    avatar_id: str,
+    upload_id: str,
+    authorization: str = Header(None),
+    db: Session = Depends(get_db),
+):
+    _require_owned_avatar(db, avatar_id, authorization)
+    upload_dir, metadata = _read_multipart_metadata(avatar_id, upload_id)
+    total_chunks = int(metadata["totalChunks"])
+    part_paths = [os.path.join(upload_dir, f"{index}.part") for index in range(total_chunks)]
+    if not all(os.path.isfile(path) for path in part_paths):
+        return fail("文件分片尚未上传完整", code=400)
+    if sum(os.path.getsize(path) for path in part_paths) != int(metadata["fileSize"]):
+        return fail("文件分片总大小不正确", code=400)
+
+    avatar_dir = os.path.join(UPLOAD_DIR, avatar_id)
+    os.makedirs(avatar_dir, exist_ok=True)
+    stored = f"{uuid.uuid4().hex}{metadata['extension']}"
+    final_path = os.path.join(avatar_dir, stored)
+    temporary_path = f"{final_path}.assembling"
+    try:
+        with open(temporary_path, "wb") as output:
+            for part_path in part_paths:
+                with open(part_path, "rb") as source:
+                    shutil.copyfileobj(source, output, UPLOAD_CHUNK_BYTES)
+        os.replace(temporary_path, final_path)
+        doc = _create_knowledge_doc(
+            db,
+            avatar_id,
+            metadata["filename"],
+            metadata["extension"],
+            int(metadata["fileSize"]),
+            stored,
+        )
+    except Exception:
+        if os.path.exists(temporary_path):
+            os.remove(temporary_path)
+        raise
+    shutil.rmtree(upload_dir, ignore_errors=True)
     return ok(_doc_payload(doc))
 
 

digital-avatar-app/backend/tests/test_chat_orchestration.py

@@ -11,6 +11,7 @@ from routers.chat import (
     _match_standard_qa,
     _public_avatar_payload,
     _qa_requires_language_adaptation,
+    _qa_requires_per_turn_rendering,
     _require_owned_avatar,
     _resolve_reply,
 )
@@ -86,6 +87,41 @@ class ChatOrchestrationTests(unittest.TestCase):
         self.assertTrue(_qa_requires_language_adaptation("안녕하세요", "你好"))
         self.assertFalse(_qa_requires_language_adaptation("你好", "您好"))
 
+    def test_conversation_qa_is_rendered_for_the_current_turn_language(self):
+        history = [SimpleNamespace(role="user", content="Please answer in English.")]
+        self.assertTrue(_qa_requires_per_turn_rendering("Quelle est votre adresse ?", "Our address is Test Road 1.", history))
+
+        fake_model = Mock(return_value="Notre adresse est Test Road 1.")
+        result = _resolve_reply(
+            None,
+            self.avatar,
+            "Quelle est votre adresse ?",
+            history,
+            qa_pairs=[SimpleNamespace(question="Quelle est votre adresse ?", answer="Our address is Test Road 1.", enabled=True)],
+            search_fn=Mock(),
+            model_client=fake_model,
+        )
+
+        self.assertEqual(result["source"], "qa")
+        self.assertEqual(result["answer"], "Notre adresse est Test Road 1.")
+        messages = fake_model.call_args.kwargs["messages"]
+        self.assertEqual(messages[-1], {"role": "user", "content": "Quelle est votre adresse ?"})
+        self.assertEqual(messages[-2]["role"], "system")
+        self.assertIn("本轮语言覆盖指令", messages[-2]["content"])
+        self.assertIn("不要沿用上一轮语言", messages[-2]["content"])
+
+    def test_latest_user_message_has_an_adjacent_language_override(self):
+        history = [
+            SimpleNamespace(role="user", content="请用中文回答"),
+            SimpleNamespace(role="assistant", content="好的，请问有什么可以帮你？"),
+        ]
+        messages = _build_prompt(self.avatar, history, "What can you help me with?", [])
+
+        self.assertEqual(messages[-1], {"role": "user", "content": "What can you help me with?"})
+        self.assertEqual(messages[-2]["role"], "system")
+        self.assertIn("最新用户消息", messages[-2]["content"])
+        self.assertIn("立即切换到相同语言", messages[-2]["content"])
+
     def test_conversational_paraphrase_matches_standard_qa(self):
         for question in ("请问一下，你们公司在哪里呀？", "请问去你们那边怎么走"):
             with self.subTest(question=question):

digital-avatar-app/backend/tests/test_knowledge_storage.py

@@ -87,6 +87,84 @@ def test_upload_rejects_oversize_file_before_queuing_indexing(
     assert not list((tmp_path / context["avatar"].id).glob("*"))
 
 
+def test_multipart_upload_reassembles_file_before_queuing_indexing(
+    tmp_path: Path,
+    authorization_context,
+):
+    context = authorization_context
+    avatar_id = context["avatar"].id
+    content = b"0123456789"
+    with (
+        patch("routers.knowledge.UPLOAD_DIR", str(tmp_path)),
+        patch("routers.knowledge.MULTIPART_CHUNK_BYTES", 4),
+        patch("routers.knowledge.knowledge_vectorizer.enqueue") as enqueue,
+    ):
+        created = client.post(
+            f"/api/avatar/{avatar_id}/knowledge/uploads",
+            headers=context["owner_headers"],
+            json={"filename": "large.pdf", "fileSize": len(content), "totalChunks": 3},
+        ).json()["data"]
+
+        for index, chunk in enumerate((content[:4], content[4:8], content[8:])):
+            response = client.post(
+                f"/api/avatar/{avatar_id}/knowledge/uploads/{created['uploadId']}/chunks/{index}",
+                headers=context["owner_headers"],
+                files={"file": (f"chunk-{index}", chunk, "application/octet-stream")},
+            )
+            assert response.json()["code"] == 200
+
+        completed = client.post(
+            f"/api/avatar/{avatar_id}/knowledge/uploads/{created['uploadId']}/complete",
+            headers=context["owner_headers"],
+        ).json()["data"]
+
+    assert completed["status"] == "parsing"
+    assert completed["fileSize"] == len(content)
+    enqueue.assert_called_once_with(completed["id"])
+    stored_path = tmp_path / avatar_id / Path(completed["fileUrl"]).name
+    assert stored_path.read_bytes() == content
+    assert not (tmp_path / ".multipart" / avatar_id / created["uploadId"]).exists()
+
+    db = SessionLocal()
+    try:
+        stored = db.query(KnowledgeDoc).filter(KnowledgeDoc.id == completed["id"]).one()
+        db.delete(stored)
+        db.commit()
+    finally:
+        db.close()
+
+
+def test_multipart_upload_rejects_incomplete_parts(
+    tmp_path: Path,
+    authorization_context,
+):
+    context = authorization_context
+    avatar_id = context["avatar"].id
+    with (
+        patch("routers.knowledge.UPLOAD_DIR", str(tmp_path)),
+        patch("routers.knowledge.MULTIPART_CHUNK_BYTES", 4),
+        patch("routers.knowledge.knowledge_vectorizer.enqueue") as enqueue,
+    ):
+        created = client.post(
+            f"/api/avatar/{avatar_id}/knowledge/uploads",
+            headers=context["owner_headers"],
+            json={"filename": "large.pdf", "fileSize": 6, "totalChunks": 2},
+        ).json()["data"]
+        client.post(
+            f"/api/avatar/{avatar_id}/knowledge/uploads/{created['uploadId']}/chunks/0",
+            headers=context["owner_headers"],
+            files={"file": ("chunk-0", b"0123", "application/octet-stream")},
+        )
+        response = client.post(
+            f"/api/avatar/{avatar_id}/knowledge/uploads/{created['uploadId']}/complete",
+            headers=context["owner_headers"],
+        )
+
+    assert response.json()["code"] == 400
+    assert response.json()["message"] == "文件分片尚未上传完整"
+    enqueue.assert_not_called()
+
+
 def test_background_vectorizer_commits_ready_document_and_chunks_together(
     tmp_path: Path,
     authorization_context,

digital-avatar-app/src/api/index.ts

@@ -333,12 +333,76 @@ export interface SearchResult {
 export const getKnowledgeDocs = (avatarId: string) =>
   request.get<KnowledgeDoc[]>(`/avatar/${avatarId}/knowledge/docs`)
 
-// 上传文档（支持 md/txt/pdf/doc/docx/xlsx）
-export const uploadKnowledgeDoc = (
+const KNOWLEDGE_UPLOAD_CHUNK_SIZE = 5 * 1024 * 1024
+
+const uploadKnowledgeChunk = async (
+  avatarId: string,
+  uploadId: string,
+  chunkIndex: number,
+  chunk: Blob,
+  onProgress?: (loaded: number) => void
+) => {
+  const form = new FormData()
+  form.append('file', chunk, `chunk-${chunkIndex}`)
+  let reportedLoaded = 0
+  for (let attempt = 1; attempt <= 3; attempt += 1) {
+    try {
+      await request.post(
+        `/avatar/${avatarId}/knowledge/uploads/${uploadId}/chunks/${chunkIndex}`,
+        form,
+        {
+          headers: { 'Content-Type': 'multipart/form-data' },
+          timeout: 2 * 60 * 1000,
+          onUploadProgress: (event) => {
+            reportedLoaded = Math.max(reportedLoaded, Math.min(event.loaded, chunk.size))
+            onProgress?.(reportedLoaded)
+          }
+        }
+      )
+      return
+    } catch (error: any) {
+      const status = Number(error?.response?.status || 0)
+      const retryable = !status || status === 408 || status === 429 || status >= 500
+      if (!retryable || attempt === 3) throw error
+      await new Promise((resolve) => window.setTimeout(resolve, attempt * 800))
+    }
+  }
+}
+
+// 大文件拆成 5MB 分片，避免生产代理的请求体限制拦截整个文件。
+export const uploadKnowledgeDoc = async (
   avatarId: string,
   file: File,
   onUploadProgress?: (loaded: number, total: number) => void
 ) => {
+  if (file.size > KNOWLEDGE_UPLOAD_CHUNK_SIZE) {
+    const totalChunks = Math.ceil(file.size / KNOWLEDGE_UPLOAD_CHUNK_SIZE)
+    const upload: any = await request.post(`/avatar/${avatarId}/knowledge/uploads`, {
+      filename: file.name,
+      fileSize: file.size,
+      totalChunks
+    })
+    let uploadedBytes = 0
+    for (let index = 0; index < totalChunks; index += 1) {
+      const start = index * KNOWLEDGE_UPLOAD_CHUNK_SIZE
+      const chunk = file.slice(start, Math.min(start + KNOWLEDGE_UPLOAD_CHUNK_SIZE, file.size))
+      await uploadKnowledgeChunk(
+        avatarId,
+        upload.uploadId,
+        index,
+        chunk,
+        (chunkLoaded) => onUploadProgress?.(uploadedBytes + chunkLoaded, file.size)
+      )
+      uploadedBytes += chunk.size
+      onUploadProgress?.(uploadedBytes, file.size)
+    }
+    return request.post<KnowledgeDoc>(
+      `/avatar/${avatarId}/knowledge/uploads/${upload.uploadId}/complete`,
+      undefined,
+      { timeout: 2 * 60 * 1000 }
+    )
+  }
+
   const form = new FormData()
   form.append('file', file)
   return request.post<KnowledgeDoc>(`/avatar/${avatarId}/knowledge/docs`, form, {

digital-avatar-app/src/views/KnowledgeManage.vue

@@ -32,7 +32,7 @@
         </div>
 
         <div v-if="displayDocs.length" class="mobile-card-list">
-          <article v-for="doc in displayDocs" :key="doc.id" class="knowledge-card">
+          <article v-for="doc in displayDocs" :key="doc.id" class="knowledge-card document-card">
             <div class="card-icon">{{ fileEmoji(doc.fileType) }}</div>
             <div class="card-content">
               <div class="card-title-row">
@@ -46,9 +46,14 @@
               </div>
             </div>
             <div class="card-actions">
-              <button v-if="documentState(doc).tone === 'failed'" class="card-retry" @click="retryDoc(doc.id)">重新索引</button>
               <button v-if="!doc.localUploading" class="card-delete" @click="removeDoc(doc.id)">{{ doc.localOnly ? '移除' : '删除' }}</button>
             </div>
+            <div v-if="canRetryDoc(doc)" class="card-retry-area">
+              <span v-if="retryErrors[doc.id]" class="card-retry-error">{{ retryErrors[doc.id] }}</span>
+              <button class="card-retry" :disabled="retryingDocs[doc.id]" @click="retryDoc(doc)">
+                {{ retryingDocs[doc.id] ? '重新索引中…' : '重新索引' }}
+              </button>
+            </div>
           </article>
         </div>
         <div v-else class="card-empty">📂 暂无文档，先上传一个知识文件</div>
@@ -115,6 +120,8 @@ const uploading = computed(() => pendingUploads.value.some((doc) => doc.localUpl
 const uploadError = ref('')
 const dragOver = ref(false)
 const fileInput = ref<HTMLInputElement | null>(null)
+const retryingDocs = ref<Record<string, boolean>>({})
+const retryErrors = ref<Record<string, string>>({})
 let documentPollingTimer: ReturnType<typeof setInterval> | undefined
 
 const query = ref('')
@@ -250,14 +257,24 @@ const uploadOne = async (file: File, ext: string) => {
   }
 }
 
-const retryDoc = async (id: string) => {
-  if (!avatarId.value) return
-  uploadError.value = ''
+const canRetryDoc = (doc: any) =>
+  !doc.localOnly && doc.filePresent !== false && documentState(doc).tone === 'failed'
+
+const retryDoc = async (doc: any) => {
+  if (!avatarId.value || !canRetryDoc(doc) || retryingDocs.value[doc.id]) return
+  retryingDocs.value = { ...retryingDocs.value, [doc.id]: true }
+  retryErrors.value = { ...retryErrors.value, [doc.id]: '' }
   try {
-    await retryKnowledgeDoc(avatarId.value, id)
-    await loadDocs()
+    const updated: any = await retryKnowledgeDoc(avatarId.value, doc.id)
+    Object.assign(doc, updated)
+    startDocumentPolling()
   } catch (e: any) {
-    uploadError.value = e?.message || '重新索引失败'
+    retryErrors.value = {
+      ...retryErrors.value,
+      [doc.id]: e?.response?.data?.message || e?.response?.data?.detail || e?.message || '重新索引失败'
+    }
+  } finally {
+    retryingDocs.value = { ...retryingDocs.value, [doc.id]: false }
   }
 }
 
@@ -378,6 +395,7 @@ onUnmounted(stopDocumentPolling)
 .panel-heading p { margin: -5px 0 0; color: #9398AE; font-size: 12px; }
 .mobile-card-list { display: grid; grid-template-columns: minmax(0, 1fr); width: 100%; min-width: 0; gap: 10px; }
 .knowledge-card { display: flex; align-items: center; width: 100%; min-width: 0; box-sizing: border-box; gap: 11px; padding: 14px; background: #fff; border: 1px solid #F1E1D3; border-radius: 16px; box-shadow: 0 5px 16px rgba(112, 62, 22, .04); }
+.document-card { display: grid; grid-template-columns: 42px minmax(0, 1fr) auto; align-items: center; }
 .card-icon { flex: 0 0 auto; width: 42px; height: 42px; display: grid; place-items: center; border-radius: 13px; background: #FFF3E6; font-size: 22px; }
 .card-content { min-width: 0; flex: 1; overflow: hidden; }
 .card-title-row { display: flex; align-items: center; gap: 8px; min-width: 0; }
@@ -388,10 +406,13 @@ onUnmounted(stopDocumentPolling)
 .card-meta, .card-detail { margin: 5px 0 0; color: #9398AE; font-size: 11px; line-height: 1.4; }.card-detail { color: #8B6B58; }
 .progress-track { width: 100%; height: 4px; margin-top: 8px; overflow: hidden; border-radius: 999px; background: #FDE7D1; }
 .progress-fill { display: block; height: 100%; border-radius: inherit; background: linear-gradient(90deg, #FB923C, #F97316); transition: width .25s ease; }
-.card-actions { flex: 0 0 auto; display: flex; flex-direction: column; align-items: stretch; gap: 6px; }
+.card-actions { flex: 0 0 auto; display: flex; align-items: center; }
 .card-delete, .card-retry { align-self: center; border: 0; border-radius: 8px; padding: 7px 9px; font-size: 12px; cursor: pointer; white-space: nowrap; }
 .card-delete { color: #EF4444; background: #FEF2F2; }
 .card-retry { color: #C15F18; background: #FFF3E6; }
+.card-retry:disabled { cursor: wait; opacity: .65; }
+.card-retry-area { grid-column: 1 / -1; display: flex; align-items: center; justify-content: flex-end; gap: 10px; min-width: 0; }
+.card-retry-error { min-width: 0; overflow: hidden; color: #DC2626; font-size: 11px; line-height: 1.35; text-overflow: ellipsis; white-space: nowrap; }
 .card-empty { padding: 42px 16px; border: 1px dashed #F1D9C3; border-radius: 16px; color: #9398AE; background: #fff; font-size: 14px; text-align: center; }
 .qa-card { align-items: stretch; text-align: left; }.qa-card.qa-disabled { opacity: .58; }
 .qa-card .card-content,
@@ -617,8 +638,11 @@ onUnmounted(stopDocumentPolling)
 @media (max-width: 520px) {
   .knowledge-panel { padding: 0 12px; }
   .knowledge-card { display: grid; grid-template-columns: 42px minmax(0, 1fr); align-items: start; gap: 10px; padding: 13px; }
+  .document-card { grid-template-columns: 42px minmax(0, 1fr) auto; }
   .card-content { grid-column: 2; }
-  .card-delete { grid-column: 2; justify-self: end; margin-top: -2px; }
+  .card-actions { grid-column: 3; grid-row: 1; }
+  .card-delete { justify-self: end; margin-top: -2px; }
+  .card-retry-area { grid-column: 1 / -1; }
   .qa-card { display: block; }
   .qa-card .card-content { width: 100%; grid-column: 1; }
   .card-title-row { align-items: flex-start; flex-wrap: wrap; gap: 5px 7px; }