支持局域网 GPU OCR 服务，配置方式类似 Ollama。

.env.example

@@ -16,6 +16,9 @@ CORS_ORIGINS=http://127.0.0.1:23566,http://localhost:23566
 
 OLLAMA_BASE_URL=http://127.0.0.1:11434
 OLLAMA_MODEL=qwen2.5:7b
+# 局域网 GPU OCR 服务（deploy/ocr-worker），留空则本机 CPU 识别
+OCR_SERVICE_URL=
+OCR_API_KEY=
 OPENAI_BASE_URL=https://api.openai.com/v1
 OPENAI_MODEL=gpt-4o-mini
 FLUCTUATION_THRESHOLD=0.08

backend/.env.example

@@ -5,4 +5,6 @@ UPLOAD_DIR=uploads
 API_PORT=23568
 OLLAMA_BASE_URL=http://127.0.0.1:11434
 OLLAMA_MODEL=qwen2.5:7b
+OCR_SERVICE_URL=
+OCR_API_KEY=
 FLUCTUATION_THRESHOLD=0.08

backend/app/core/config.py

@@ -25,6 +25,9 @@ class Settings(BaseSettings):
     OCR_MAX_SIDE: int = 1280
     UPLOAD_MAX_SIDE: int = 2048
     OCR_WARMUP: bool = True
+    OCR_SERVICE_URL: str = ""
+    OCR_API_KEY: str = ""
+    OCR_USE_GPU: bool = False
 
     class Config:
         env_file = ".env"

backend/app/models/user.py

@@ -159,6 +159,7 @@ class SystemSettings(Base):
     openai_base_url: Mapped[str | None] = mapped_column(String(256), nullable=True)
     openai_model: Mapped[str | None] = mapped_column(String(128), nullable=True)
     openai_api_key: Mapped[str | None] = mapped_column(String(512), nullable=True)
+    ocr_service_url: Mapped[str | None] = mapped_column(String(256), nullable=True)
     updated_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), default=lambda: datetime.now(timezone.utc)
     )

backend/app/routers/admin.py

@@ -31,6 +31,7 @@ def settings_to_out(row: SystemSettings) -> SystemSettingsOut:
         openai_base_url=row.openai_base_url,
         openai_model=row.openai_model,
         openai_api_key_set=bool(row.openai_api_key),
+        ocr_service_url=row.ocr_service_url,
         updated_at=row.updated_at,
     )
 
@@ -74,6 +75,8 @@ def update_settings(
         row.openai_model = data.openai_model or None
     if data.openai_api_key is not None and data.openai_api_key.strip():
         row.openai_api_key = data.openai_api_key.strip()
+    if data.ocr_service_url is not None:
+        row.ocr_service_url = data.ocr_service_url.strip() or None
     row.updated_at = datetime.now(timezone.utc)
     db.commit()
     db.refresh(row)

backend/app/routers/wrong_questions.py

@@ -11,7 +11,7 @@ from sqlalchemy.orm import Session, joinedload
 from app.core.config import settings
 from app.core.database import SessionLocal, get_db
 from app.core.deps import get_current_user
-from app.models.user import Subject, User, WrongQuestion, WrongQuestionCategory, WrongQuestionStatus
+from app.models.user import Subject, SystemSettings, User, WrongQuestion, WrongQuestionCategory, WrongQuestionStatus
 from app.schemas import WrongQuestionCategoryEnum, WrongQuestionOut, WrongQuestionUpdate
 from app.services import annotation as annotation_service
 from app.services import llm as llm_service
@@ -50,6 +50,13 @@ def _expire_stale_processing(wq: WrongQuestion, db: Session) -> None:
     db.commit()
 
 
+def _ocr_service_url(db: Session) -> str | None:
+    row = db.get(SystemSettings, 1)
+    if row and row.ocr_service_url:
+        return row.ocr_service_url.strip() or None
+    return ocr_service.resolve_ocr_service_url()
+
+
 def _parse_mark_regions(raw: str | None) -> list[dict] | None:
     if not raw:
         return None
@@ -140,9 +147,12 @@ def _process_wrong_question(question_id: uuid.UUID):
 
         wq.error_message = None
         image_full = Path(settings.UPLOAD_DIR) / wq.image_path
+        ocr_url = _ocr_service_url(db)
         try:
             with ThreadPoolExecutor(max_workers=1) as pool:
-                future = pool.submit(ocr_service.run_ocr_with_regions, str(image_full))
+                future = pool.submit(
+                    ocr_service.run_ocr_with_regions, str(image_full), ocr_url
+                )
                 ocr_result = future.result(timeout=settings.OCR_TIMEOUT_SECONDS)
             ocr_text = ocr_result["text"]
             ocr_lines = ocr_result["lines"]
@@ -166,6 +176,8 @@ def _process_wrong_question(question_id: uuid.UUID):
             msg = _short_error(exc, "OCR 识别失败：")
             if "libGL" in str(exc):
                 msg += " 请在服务器执行: sudo bash deploy/install-ocr-deps.sh && systemctl restart grade-archive"
+            elif ocr_url:
+                msg += f" 请检查 OCR 服务是否可达: {ocr_url} （可浏览器访问 {ocr_url.rstrip('/')}/health）"
             wq.error_message = msg
             db.commit()
             return

backend/app/schemas/__init__.py

@@ -74,6 +74,7 @@ class SystemSettingsOut(BaseModel):
     openai_base_url: str | None = None
     openai_model: str | None = None
     openai_api_key_set: bool = False
+    ocr_service_url: str | None = None
     updated_at: datetime
 
     model_config = {"from_attributes": True}
@@ -87,6 +88,7 @@ class SystemSettingsUpdate(BaseModel):
     openai_base_url: str | None = None
     openai_model: str | None = None
     openai_api_key: str | None = None
+    ocr_service_url: str | None = None
 
 
 class AdminProfileUpdate(BaseModel):

backend/app/services/migrate.py

@@ -61,6 +61,8 @@ def run_migrations() -> None:
             alters.append("ADD COLUMN openai_model VARCHAR(128)")
         if "openai_api_key" not in ss_columns:
             alters.append("ADD COLUMN openai_api_key VARCHAR(512)")
+        if "ocr_service_url" not in ss_columns:
+            alters.append("ADD COLUMN ocr_service_url VARCHAR(256)")
         if alters:
             with engine.begin() as conn:
                 for clause in alters:

backend/app/services/ocr.py

@@ -5,6 +5,7 @@ import threading
 from io import BytesIO
 from pathlib import Path
 
+import httpx
 from PIL import Image
 
 from app.core.config import settings
@@ -23,22 +24,32 @@ def get_ocr_engine():
     if _ocr_engine is None:
         from paddleocr import PaddleOCR
 
+        use_gpu = settings.OCR_USE_GPU
         _ocr_engine = PaddleOCR(
             use_angle_cls=False,
             lang="ch",
             show_log=False,
-            use_gpu=False,
+            use_gpu=use_gpu,
-            enable_mkldnn=True,
+            enable_mkldnn=not use_gpu,
             det_limit_side_len=min(settings.OCR_MAX_SIDE, 1280),
             rec_batch_num=8,
         )
     return _ocr_engine
 
 
+def resolve_ocr_service_url(service_url: str | None = None) -> str | None:
+    url = (service_url or settings.OCR_SERVICE_URL or "").strip()
+    return url or None
+
+
+def uses_remote_ocr(service_url: str | None = None) -> bool:
+    return resolve_ocr_service_url(service_url) is not None
+
+
 def warmup_ocr_engine() -> None:
     """后台预加载 OCR 模型，避免首张图片等待数分钟。"""
     global _ocr_warmup_started
-    if _ocr_warmup_started or not settings.OCR_WARMUP:
+    if _ocr_warmup_started or not settings.OCR_WARMUP or uses_remote_ocr():
         return
     _ocr_warmup_started = True
 
@@ -110,8 +121,20 @@ def _prepare_ocr_image(image_path: str) -> tuple[str, float, float, int, int, Pa
     return str(tmp), scale_x, scale_y, orig_w, orig_h, tmp
 
 
-def run_ocr_with_regions(image_path: str) -> dict:
+def _run_remote_ocr(service_url: str, image_path: str) -> dict:
-    """Return OCR text plus line-level bounding boxes for annotation."""
+    url = f"{service_url.rstrip('/')}/api/ocr/regions"
+    headers: dict[str, str] = {}
+    if settings.OCR_API_KEY:
+        headers["X-OCR-Key"] = settings.OCR_API_KEY
+    with open(image_path, "rb") as handle:
+        files = {"file": (Path(image_path).name, handle, "image/jpeg")}
+        with httpx.Client(timeout=settings.OCR_TIMEOUT_SECONDS) as client:
+            resp = client.post(url, files=files, headers=headers)
+            resp.raise_for_status()
+            return resp.json()
+
+
+def _run_local_ocr(image_path: str) -> dict:
     engine = get_ocr_engine()
     ocr_path, scale_x, scale_y, orig_w, orig_h, tmp_path = _prepare_ocr_image(image_path)
     try:
@@ -150,6 +173,14 @@ def run_ocr_with_regions(image_path: str) -> dict:
     }
 
 
+def run_ocr_with_regions(image_path: str, service_url: str | None = None) -> dict:
+    """Return OCR text plus line-level bounding boxes for annotation."""
+    remote = resolve_ocr_service_url(service_url)
+    if remote:
+        return _run_remote_ocr(remote, image_path)
+    return _run_local_ocr(image_path)
+
+
 def run_ocr(image_path: str) -> str:
     return run_ocr_with_regions(image_path)["text"]
 

deploy/ocr-worker/app.py

@@ -0,0 +1,140 @@
+"""局域网 OCR 服务：在带 NVIDIA 显卡的机器上运行，供成绩档案系统调用。"""
+
+import os
+import tempfile
+from pathlib import Path
+
+from fastapi import FastAPI, File, Header, HTTPException, UploadFile
+from PIL import Image
+
+os.environ.setdefault("OPENCV_IO_ENABLE_OPENEXR", "0")
+
+OCR_MAX_SIDE = int(os.getenv("OCR_MAX_SIDE", "1280"))
+OCR_API_KEY = os.getenv("OCR_API_KEY", "").strip()
+OCR_USE_GPU = os.getenv("OCR_USE_GPU", "true").lower() in {"1", "true", "yes"}
+
+app = FastAPI(title="Grade Archive OCR Worker", version="1.0.0")
+_engine = None
+
+
+def _check_key(key: str | None) -> None:
+    if OCR_API_KEY and key != OCR_API_KEY:
+        raise HTTPException(status_code=401, detail="Invalid OCR API key")
+
+
+def get_engine():
+    global _engine
+    if _engine is None:
+        from paddleocr import PaddleOCR
+
+        _engine = PaddleOCR(
+            use_angle_cls=False,
+            lang="ch",
+            show_log=False,
+            use_gpu=OCR_USE_GPU,
+            enable_mkldnn=not OCR_USE_GPU,
+            det_limit_side_len=min(OCR_MAX_SIDE, 1280),
+            rec_batch_num=8,
+        )
+    return _engine
+
+
+def _bbox_from_box(box: list) -> list[float]:
+    xs = [float(p[0]) for p in box]
+    ys = [float(p[1]) for p in box]
+    return [min(xs), min(ys), max(xs), max(ys)]
+
+
+def _scale_box(box: list, scale_x: float, scale_y: float) -> list:
+    return [[float(p[0]) * scale_x, float(p[1]) * scale_y] for p in box]
+
+
+def _prepare_image_bytes(content: bytes) -> tuple[bytes, float, float, int, int]:
+    with Image.open(__import__("io").BytesIO(content)) as img:
+        img = img.convert("RGB")
+        orig_w, orig_h = img.size
+        longest = max(orig_w, orig_h)
+        if longest <= OCR_MAX_SIDE:
+            buf = __import__("io").BytesIO()
+            img.save(buf, format="JPEG", quality=88)
+            return buf.getvalue(), 1.0, 1.0, orig_w, orig_h
+
+        ratio = OCR_MAX_SIDE / longest
+        new_w = max(1, int(orig_w * ratio))
+        new_h = max(1, int(orig_h * ratio))
+        resized = img.resize((new_w, new_h), Image.Resampling.LANCZOS)
+        buf = __import__("io").BytesIO()
+        resized.save(buf, format="JPEG", quality=85)
+        scale_x = orig_w / new_w
+        scale_y = orig_h / new_h
+        return buf.getvalue(), scale_x, scale_y, orig_w, orig_h
+
+
+def run_ocr_on_bytes(content: bytes) -> dict:
+    engine = get_engine()
+    image_bytes, scale_x, scale_y, orig_w, orig_h = _prepare_image_bytes(content)
+    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as tmp:
+        tmp.write(image_bytes)
+        tmp_path = tmp.name
+    try:
+        result = engine.ocr(tmp_path, cls=False)
+    finally:
+        Path(tmp_path).unlink(missing_ok=True)
+
+    lines: list[dict] = []
+    if result and result[0]:
+        for item in result[0]:
+            if not item or len(item) < 2:
+                continue
+            box, rec = item[0], item[1]
+            text = rec[0] if rec else ""
+            conf = float(rec[1]) if rec and len(rec) > 1 else 0.0
+            if not text:
+                continue
+            if scale_x != 1.0 or scale_y != 1.0:
+                box = _scale_box(box, scale_x, scale_y)
+            lines.append(
+                {
+                    "text": text,
+                    "confidence": conf,
+                    "box": box,
+                    "bbox": _bbox_from_box(box),
+                }
+            )
+
+    return {
+        "text": "\n".join(line["text"] for line in lines),
+        "lines": lines,
+        "width": orig_w,
+        "height": orig_h,
+    }
+
+
+@app.on_event("startup")
+def warmup():
+    buf = __import__("io").BytesIO()
+    Image.new("RGB", (120, 40), color=(255, 255, 255)).save(buf, format="JPEG")
+    try:
+        run_ocr_on_bytes(buf.getvalue())
+    except Exception:
+        pass
+
+
+@app.get("/health")
+def health():
+    return {"status": "ok", "gpu": OCR_USE_GPU}
+
+
+@app.post("/api/ocr/regions")
+async def ocr_regions(
+    file: UploadFile = File(...),
+    x_ocr_key: str | None = Header(default=None, alias="X-OCR-Key"),
+):
+    _check_key(x_ocr_key)
+    content = await file.read()
+    if not content:
+        raise HTTPException(status_code=400, detail="Empty image")
+    try:
+        return run_ocr_on_bytes(content)
+    except Exception as exc:
+        raise HTTPException(status_code=500, detail=str(exc)) from exc

deploy/ocr-worker/install.sh

@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+# 在带 NVIDIA 显卡（如 RTX 3060 Ti）的 Linux 机器上安装 OCR Worker
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")" && pwd)"
+VENV="${ROOT}/.venv"
+PORT="${OCR_PORT:-23567}"
+
+echo "==> OCR Worker 安装目录: ${ROOT}"
+
+if ! command -v python3 >/dev/null; then
+  echo "请先安装 python3"
+  exit 1
+fi
+
+python3 -m venv "${VENV}"
+# shellcheck disable=SC1091
+source "${VENV}/bin/activate"
+pip install -U pip wheel
+
+# Paddle GPU（CUDA 11.8，适配多数 3060 Ti 驱动）
+pip install paddlepaddle-gpu==2.6.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/
+pip install -r "${ROOT}/requirements.txt"
+
+cat <<EOF
+
+安装完成。
+
+启动（前台测试）:
+  cd ${ROOT}
+  OCR_USE_GPU=true OCR_PORT=${PORT} bash start.sh
+
+局域网地址示例:
+  http://192.168.x.x:${PORT}
+
+在成绩档案「系统设置 → AI 模型」里填写 OCR 服务地址:
+  http://192.168.x.x:${PORT}
+
+可选环境变量:
+  OCR_USE_GPU=true          # 使用显卡（默认 true）
+  OCR_API_KEY=随机字符串     # 局域网鉴权（可选）
+  OCR_MAX_SIDE=1280           # 识别长边上限
+  OCR_PORT=${PORT}            # 监听端口
+
+EOF

deploy/ocr-worker/ocr-worker.service

@@ -0,0 +1,17 @@
+[Unit]
+Description=Grade Archive OCR Worker (PaddleOCR GPU)
+After=network.target
+
+[Service]
+Type=simple
+WorkingDirectory=/opt/ocr-worker
+Environment=OCR_USE_GPU=true
+Environment=OCR_PORT=23567
+Environment=OCR_HOST=0.0.0.0
+# Environment=OCR_API_KEY=请设置随机密钥
+ExecStart=/opt/ocr-worker/.venv/bin/uvicorn app:app --host 0.0.0.0 --port 23567
+Restart=on-failure
+RestartSec=5
+
+[Install]
+WantedBy=multi-user.target

deploy/ocr-worker/requirements.txt

@@ -0,0 +1,6 @@
+fastapi==0.115.6
+uvicorn[standard]==0.34.0
+python-multipart==0.0.20
+paddleocr==2.9.1
+Pillow==11.0.0
+# GPU 版 Paddle 请用 install.sh 安装，勿直接 pip install paddlepaddle

deploy/ocr-worker/start.sh

@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")" && pwd)"
+VENV="${ROOT}/.venv"
+HOST="${OCR_HOST:-0.0.0.0}"
+PORT="${OCR_PORT:-23567}"
+
+if [[ ! -d "${VENV}" ]]; then
+  echo "未找到虚拟环境，请先运行: bash install.sh"
+  exit 1
+fi
+
+# shellcheck disable=SC1091
+source "${VENV}/bin/activate"
+cd "${ROOT}"
+exec uvicorn app:app --host "${HOST}" --port "${PORT}"

frontend/dist/index.html

@@ -9,7 +9,7 @@
     <meta name="author" content="马建军" />
     <meta name="copyright" content="Copyright (c) 马建军. All rights reserved." />
     <title>中学成绩档案</title>
-    <script type="module" crossorigin src="/assets/index-DzzkB1zh.js"></script>
+    <script type="module" crossorigin src="/assets/index-19dlnnB9.js"></script>
     <link rel="stylesheet" crossorigin href="/assets/index-GY2etMYN.css">
   </head>
   <body>

frontend/src/api/client.ts

@@ -78,6 +78,7 @@ export const adminApi = {
     openai_base_url?: string | null
     openai_model?: string | null
     openai_api_key?: string
+    ocr_service_url?: string | null
   }) => api.patch<SystemSettings>('/admin/settings', data),
   updateProfile: (data: {
     username?: string

frontend/src/pages/SettingsPage.tsx

@@ -52,6 +52,7 @@ export default function SettingsPage() {
         openai_base_url: settingsRes.data.openai_base_url || '',
         openai_model: settingsRes.data.openai_model || '',
         openai_api_key: '',
+        ocr_service_url: settingsRes.data.ocr_service_url || '',
       })
     } finally {
       setLoading(false)
@@ -115,6 +116,7 @@ export default function SettingsPage() {
     openai_base_url?: string
     openai_model?: string
     openai_api_key?: string
+    ocr_service_url?: string
   }) => {
     const payload: Parameters<typeof adminApi.updateSettings>[0] = {
       ai_provider: values.ai_provider,
@@ -122,6 +124,7 @@ export default function SettingsPage() {
       ollama_model: values.ollama_model || null,
       openai_base_url: values.openai_base_url || null,
       openai_model: values.openai_model || null,
+      ocr_service_url: values.ocr_service_url?.trim() || null,
     }
     if (values.openai_api_key?.trim()) {
       payload.openai_api_key = values.openai_api_key.trim()
@@ -234,6 +237,13 @@ export default function SettingsPage() {
                         </Form.Item>
                       </>
                     )}
+                    <Form.Item
+                      name="ocr_service_url"
+                      label="OCR 服务地址（局域网 GPU 机器）"
+                      extra="留空则在应用服务器本机 CPU 识别。填写后类似 Ollama，例如 http://192.168.8.100:23567"
+                    >
+                      <Input placeholder="http://192.168.8.100:23567" />
+                    </Form.Item>
                     <Typography.Paragraph type="secondary">
                       错题/奥数解法将按学生学段（初中/高中）生成，并严格禁止超纲解题。
                     </Typography.Paragraph>

frontend/src/types/index.ts

@@ -23,6 +23,7 @@ export interface SystemSettings {
   openai_base_url: string | null
   openai_model: string | null
   openai_api_key_set: boolean
+  ocr_service_url: string | null
   updated_at: string
 }