fix(avatar): OCR image-only knowledge PDFs

digital-avatar-app/backend/services/knowledge_vectorizer.py

@@ -8,7 +8,8 @@ import threading
 from datetime import datetime, timezone
 
 from database import SessionLocal
-from models import KnowledgeChunk, KnowledgeDoc
+from models import Avatar, KnowledgeChunk, KnowledgeDoc
+from services.pdf_ocr_service import extract_scanned_pdf_text
 import embeddings
 
 logger = logging.getLogger(__name__)
@@ -76,7 +77,23 @@ class KnowledgeVectorizer:
 
             self._set_progress(db, doc, "extracting", 8)
             text = embeddings.extract_text(path, f".{doc.file_type}")
-            self._set_progress(db, doc, "chunking", 22)
+            if doc.file_type == "pdf" and not text.strip():
+                avatar = db.get(Avatar, doc.avatar_id)
+                if not avatar:
+                    raise ValueError("文档所属分身不存在")
+
+                def ocr_progress(done: int, total: int):
+                    percent = 8 + int((done / max(1, total)) * 20)
+                    self._set_progress(db, doc, "ocr", min(percent, 28))
+
+                self._set_progress(db, doc, "ocr", 8)
+                text = extract_scanned_pdf_text(
+                    db,
+                    avatar,
+                    path,
+                    on_progress=ocr_progress,
+                )
+            self._set_progress(db, doc, "chunking", 29)
             chunks = embeddings.chunk_text(text)
             if not chunks:
                 raise ValueError("文档没有可建立索引的文字内容")