作业帮式错题标注：OCR 定位错误红框 + 解题思路。

- PaddleOCR 行级坐标 + AI 识别错答区域，生成标注图 - 解法拆分为「解题思路」与「详细解答」 - 详情页标注图/原图切换，列表显示标注缩略图
2026-06-28 13:50:20 +08:00
parent c30e21b51e
commit a2a6d59f7c
16 changed files with 852 additions and 507 deletions
@@ -1,3 +1,4 @@
+import json
 import uuid
 from pathlib import Path

@@ -10,6 +11,7 @@ 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.schemas import WrongQuestionCategoryEnum, WrongQuestionOut, WrongQuestionUpdate
+from app.services import annotation as annotation_service
 from app.services import llm as llm_service
 from app.services import ocr as ocr_service
 from app.services.student_access import get_student_for_user
@@ -17,6 +19,16 @@ from app.services.student_access import get_student_for_user
 router = APIRouter(tags=["wrong_questions"])


+def _parse_mark_regions(raw: str | None) -> list[dict] | None:
+    if not raw:
+        return None
+    try:
+        data = json.loads(raw)
+        return data if isinstance(data, list) else None
+    except json.JSONDecodeError:
+        return None
+
+
 def _wq_to_out(wq: WrongQuestion) -> WrongQuestionOut:
    return WrongQuestionOut(
        id=wq.id,
@@ -27,12 +39,43 @@ def _wq_to_out(wq: WrongQuestion) -> WrongQuestionOut:
        image_path=wq.image_path,
        ocr_raw_text=wq.ocr_raw_text,
        question_text=wq.question_text,
+        solution_approach=wq.solution_approach,
        solution_text=wq.solution_text,
+        mark_regions=_parse_mark_regions(wq.mark_regions_json),
+        has_annotated_image=bool(wq.annotated_image_path),
        status=wq.status,
        created_at=wq.created_at,
    )


+async def _run_ai_pipeline(wq: WrongQuestion, db: Session, ocr_lines: list[dict], ocr_text: str):
+    subject_name = wq.subject.name if wq.subject else "综合"
+    school_level = wq.student.school_level if wq.student else None
+    olympiad = wq.category == WrongQuestionCategory.olympiad
+    ai_cfg = llm_service.load_ai_config(db)
+    image_full = str(Path(settings.UPLOAD_DIR) / wq.image_path)
+
+    detect_resp = await llm_service.detect_wrong_line_ids(ai_cfg, subject_name, ocr_lines, school_level)
+    wrong_ids = annotation_service.parse_wrong_line_ids(detect_resp, ocr_lines)
+    regions = annotation_service.regions_from_lines(ocr_lines, wrong_ids)
+    wq.mark_regions_json = json.dumps(regions, ensure_ascii=False)
+
+    ann_rel = ocr_service.annotated_rel_path(wq.image_path)
+    wq.annotated_image_path = annotation_service.draw_annotated_image(
+        image_full, ocr_lines, wrong_ids, ann_rel
+    )
+
+    question_text = await llm_service.format_question(ai_cfg, subject_name, ocr_text, school_level)
+    solution_full = await llm_service.generate_solution(
+        ai_cfg, subject_name, question_text, school_level, olympiad=olympiad
+    )
+    approach, solution_body = annotation_service.split_solution_sections(solution_full)
+    wq.question_text = question_text
+    wq.solution_approach = approach
+    wq.solution_text = solution_body if approach else solution_full
+    wq.status = WrongQuestionStatus.solved
+
+
 def _process_wrong_question(question_id: uuid.UUID):
    db = SessionLocal()
    try:
@@ -47,7 +90,9 @@ def _process_wrong_question(question_id: uuid.UUID):

        image_full = Path(settings.UPLOAD_DIR) / wq.image_path
        try:
-            ocr_text = ocr_service.run_ocr(str(image_full))
+            ocr_result = ocr_service.run_ocr_with_regions(str(image_full))
+            ocr_text = ocr_result["text"]
+            ocr_lines = ocr_result["lines"]
            wq.ocr_raw_text = ocr_text or None
            wq.status = WrongQuestionStatus.ocr_done if ocr_text else WrongQuestionStatus.failed
            db.commit()
@@ -59,31 +104,12 @@ def _process_wrong_question(question_id: uuid.UUID):
        if not ocr_text:
            return

-        subject_name = wq.subject.name if wq.subject else "综合"
-        school_level = wq.student.school_level if wq.student else None
-        olympiad = wq.category == WrongQuestionCategory.olympiad
-        ai_cfg = llm_service.load_ai_config(db)
-
        import asyncio

        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)
        try:
-            question_text = loop.run_until_complete(
-                llm_service.format_question(ai_cfg, subject_name, ocr_text, school_level)
-            )
-            solution_text = loop.run_until_complete(
-                llm_service.generate_solution(
-                    ai_cfg,
-                    subject_name,
-                    question_text,
-                    school_level,
-                    olympiad=olympiad,
-                )
-            )
-            wq.question_text = question_text
-            wq.solution_text = solution_text
-            wq.status = WrongQuestionStatus.solved
+            loop.run_until_complete(_run_ai_pipeline(wq, db, ocr_lines, ocr_text))
            db.commit()
        except Exception:
            wq.status = WrongQuestionStatus.ocr_done
@@ -217,6 +243,8 @@ def update_wrong_question(
        wq.question_text = data.question_text
    if data.solution_text is not None:
        wq.solution_text = data.solution_text
+    if data.solution_approach is not None:
+        wq.solution_approach = data.solution_approach

    db.commit()
    db.refresh(wq)
@@ -239,10 +267,13 @@ def delete_wrong_question(
        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="错题不存在")

    image_full = Path(settings.UPLOAD_DIR) / wq.image_path
+    ann_full = Path(settings.UPLOAD_DIR) / wq.annotated_image_path if wq.annotated_image_path else None
    db.delete(wq)
    db.commit()
    if image_full.exists():
        image_full.unlink()
+    if ann_full and ann_full.exists():
+        ann_full.unlink()


@router.post("/wrong-questions/{question_id}/retry-ocr", response_model=WrongQuestionOut)
@@ -297,13 +328,16 @@ async def regenerate_solution(
                ai_cfg, subject_name, wq.ocr_raw_text, school_level
            )
            question_text = wq.question_text
-        wq.solution_text = await llm_service.generate_solution(
+        solution_full = await llm_service.generate_solution(
            ai_cfg,
            subject_name,
            question_text,
            school_level,
            olympiad=olympiad,
        )
+        approach, solution_body = annotation_service.split_solution_sections(solution_full)
+        wq.solution_approach = approach
+        wq.solution_text = solution_body if approach else solution_full
        wq.status = WrongQuestionStatus.solved
    except Exception as exc:
        raise HTTPException(
@@ -334,3 +368,26 @@ def get_wrong_question_image(
    if not image_full.exists():
        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="图片不存在")
    return FileResponse(image_full)
+
+
+@router.get("/wrong-questions/{question_id}/annotated-image")
+def get_wrong_question_annotated_image(
+    question_id: uuid.UUID,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    wq = (
+        db.query(WrongQuestion)
+        .options(joinedload(WrongQuestion.student))
+        .filter(WrongQuestion.id == question_id)
+        .first()
+    )
+    if wq is None or wq.student.user_id != current_user.id:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="错题不存在")
+    if not wq.annotated_image_path:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="标注图尚未生成")
+
+    image_full = Path(settings.UPLOAD_DIR) / wq.annotated_image_path
+    if not image_full.exists():
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="标注图不存在")
+    return FileResponse(image_full)