feat: add concurrency safety, caption detection, admin enhancements, and performance improvements

app/services/summarizer.py

@@ -31,6 +31,7 @@ from app.services.summary_persister import (
     _cleanup_old_images,
     _handle_summary_failure,
     _persist_summary,
+    _run_post_processing,
 )
 from app.utils import TMP_DIR, release_lock, truncate_error, utc_now
 
@@ -115,12 +116,31 @@ async def _do_summarize_one(db: Session, paper: Paper, pdf_mode: str = "auto") -
         _t3 = _time.monotonic()
         logger.info("  [%s] pi生成: %.2fs", arxiv_id, _t3 - _t2)
 
-        quality = _persist_summary(db, paper, json_data, raw_output)
+        quality, schema = _persist_summary(db, paper, json_data, raw_output)
         _t4 = _time.monotonic()
         logger.info("  [%s] 持久化: %.2fs", arxiv_id, _t4 - _t3)
 
+        # 后处理（图片提取 + ChromaDB 索引）搬到线程池跑，避免 CPU 密集推理冻结
+        # 事件循环。paper 字段在此（事件循环线程）提取成纯值再传入，规避 worker
+        # 线程跨线程访问 ORM 的 DetachedInstanceError。DocLayout 推理由单例的
+        # threading.Lock 串行化，并发 worker 不会同时压模型。
+        paper_meta = {
+            "title_en": paper.title_en or "",
+            "tags": " ".join(t.tag for t in paper.tags) if paper.tags else "",
+            "paper_date": paper.paper_date.isoformat() if paper.paper_date else "",
+        }
+        _t5 = _time.monotonic()
+        try:
+            await asyncio.to_thread(_run_post_processing, arxiv_id, schema, paper_meta)
+        except Exception:
+            # 双保险：_run_post_processing 内部已 try/except，此处兜底，
+            # 确保后处理失败绝不影响已 DONE 的总结。
+            logger.warning("Post-processing error for %s", arxiv_id, exc_info=True)
+        _t6 = _time.monotonic()
+        logger.info("  [%s] 后处理(线程池): %.2fs", arxiv_id, _t6 - _t5)
+
         logger.info(
-            "✅ [%s] 完成: quality=%s  总耗时: %.2fs", arxiv_id, quality, _t4 - _t0
+            "✅ [%s] 完成: quality=%s  总耗时: %.2fs", arxiv_id, quality, _t6 - _t0
         )
         return {"arxiv_id": arxiv_id, "status": "done", "quality": quality}