feat(server): 新增文档智能识别服务，扩展OCR接口支持 Azure Document Intelligence

server/src/app/api/v1/endpoints/ocr.py

@@ -3,8 +3,9 @@ from __future__ import annotations
 from typing import Annotated
 
 from fastapi import APIRouter, Depends, File, HTTPException, UploadFile, status
+from sqlalchemy.orm import Session
 
-from app.api.deps import CurrentUserContext, get_current_user
+from app.api.deps import CurrentUserContext, get_current_user, get_db
 from app.schemas.common import ErrorResponse
 from app.schemas.ocr import OcrRecognizeBatchRead
 from app.services.ocr import OcrService
@@ -35,6 +36,7 @@ router = APIRouter(prefix="/ocr")
 async def recognize_ocr_documents(
     files: Annotated[list[UploadFile], File(description="待识别的票据图片或 PDF。")],
     _: Annotated[CurrentUserContext, Depends(get_current_user)],
+    db: Annotated[Session, Depends(get_db)],
 ) -> OcrRecognizeBatchRead:
     try:
         payload = []
@@ -46,7 +48,7 @@ async def recognize_ocr_documents(
                     upload.content_type,
                 )
             )
-        return OcrService().recognize_files(payload)
+        return OcrService(db).recognize_files(payload)
     except ValueError as exc:
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(exc)) from exc
     except RuntimeError as exc:

server/src/app/schemas/ocr.py

@@ -10,6 +10,12 @@ class OcrRecognizeLineRead(BaseModel):
     page_index: int | None = Field(default=None, description="页码，从 0 开始。")
 
 
+class OcrRecognizeFieldRead(BaseModel):
+    key: str = Field(description="结构化字段键。")
+    label: str = Field(description="结构化字段展示名。")
+    value: str = Field(default="", description="结构化字段值。")
+
+
 class OcrRecognizeDocumentRead(BaseModel):
     filename: str = Field(description="原始文件名。")
     media_type: str = Field(description="文件媒体类型。")
@@ -20,6 +26,19 @@ class OcrRecognizeDocumentRead(BaseModel):
     avg_score: float = Field(default=0.0, ge=0.0, le=1.0, description="平均识别置信度。")
     line_count: int = Field(default=0, ge=0, description="文本行数。")
     page_count: int = Field(default=1, ge=0, description="识别页数。")
+    document_type: str = Field(default="other", description="识别出的票据类型编码。")
+    document_type_label: str = Field(default="其他单据", description="识别出的票据类型名称。")
+    scene_code: str = Field(default="other", description="识别出的票据场景编码。")
+    scene_label: str = Field(default="其他票据", description="识别出的票据场景名称。")
+    classification_source: str = Field(default="rule", description="票据类型判断来源，例如 rule / llm_text / llm_vision。")
+    classification_confidence: float = Field(default=0.0, ge=0.0, le=1.0, description="票据类型判断置信度。")
+    classification_evidence: list[str] = Field(default_factory=list, description="票据类型判断依据摘要。")
+    document_fields: list[OcrRecognizeFieldRead] = Field(
+        default_factory=list,
+        description="识别出的结构化票据信息。",
+    )
+    preview_kind: str = Field(default="", description="预览类型，PDF 转图后通常为 image。")
+    preview_data_url: str = Field(default="", description="用于前端展示的图片预览 data URL。")
     warnings: list[str] = Field(default_factory=list, description="该文件的识别提示或警告。")
     lines: list[OcrRecognizeLineRead] = Field(default_factory=list, description="逐行识别结果。")
 

server/src/app/services/document_intelligence.py

@@ -0,0 +1,582 @@
+from __future__ import annotations
+
+import json
+import re
+from dataclasses import dataclass
+from decimal import Decimal, InvalidOperation
+from typing import Any
+
+from pydantic import BaseModel, Field, ValidationError
+from sqlalchemy.orm import Session
+
+from app.services.runtime_chat import RuntimeChatService
+
+
+@dataclass(frozen=True, slots=True)
+class DocumentField:
+    key: str
+    label: str
+    value: str
+
+
+@dataclass(frozen=True, slots=True)
+class DocumentInsight:
+    document_type: str
+    document_type_label: str
+    scene_code: str
+    scene_label: str
+    expense_type: str
+    fields: tuple[DocumentField, ...] = ()
+    classification_source: str = "rule"
+    classification_confidence: float = 0.0
+    evidence: tuple[str, ...] = ()
+    warnings: tuple[str, ...] = ()
+
+
+@dataclass(frozen=True, slots=True)
+class DocumentRule:
+    document_type: str
+    document_type_label: str
+    scene_code: str
+    scene_label: str
+    expense_type: str
+    keywords: tuple[str, ...]
+    score_bias: float = 0.0
+
+
+@dataclass(frozen=True, slots=True)
+class RuleMatch:
+    rule: DocumentRule | None
+    confidence: float
+    evidence: tuple[str, ...]
+    score: float
+
+
+class LlmDocumentClassification(BaseModel):
+    document_type: str = Field(default="other")
+    scene_code: str = Field(default="other")
+    scene_label: str = Field(default="其他票据")
+    expense_type: str = Field(default="other")
+    confidence: float = Field(default=0.0, ge=0.0, le=1.0)
+    evidence: list[str] = Field(default_factory=list)
+
+
+DEFAULT_RULE = DocumentRule(
+    document_type="other",
+    document_type_label="其他单据",
+    scene_code="other",
+    scene_label="其他票据",
+    expense_type="other",
+    keywords=(),
+    score_bias=0.0,
+)
+
+DOCUMENT_RULES: tuple[DocumentRule, ...] = (
+    DocumentRule(
+        document_type="flight_itinerary",
+        document_type_label="机票/航班行程单",
+        scene_code="travel",
+        scene_label="差旅票据",
+        expense_type="travel",
+        keywords=("电子行程单", "航班号", "航班", "机票", "登机", "航空", "客票"),
+        score_bias=0.34,
+    ),
+    DocumentRule(
+        document_type="train_ticket",
+        document_type_label="火车/高铁票",
+        scene_code="travel",
+        scene_label="差旅票据",
+        expense_type="travel",
+        keywords=("高铁", "火车", "动车", "铁路", "车次", "检票", "二等座", "一等座"),
+        score_bias=0.32,
+    ),
+    DocumentRule(
+        document_type="hotel_invoice",
+        document_type_label="酒店住宿票据",
+        scene_code="hotel",
+        scene_label="住宿票据",
+        expense_type="hotel",
+        keywords=("住宿", "房费", "客房", "入住", "离店", "酒店", "宾馆", "间夜"),
+        score_bias=0.16,
+    ),
+    DocumentRule(
+        document_type="taxi_receipt",
+        document_type_label="出租车/网约车票据",
+        scene_code="transport",
+        scene_label="交通票据",
+        expense_type="transport",
+        keywords=("滴滴出行", "滴滴", "网约车", "出租车", "打车", "快车", "专车", "订单号", "上车", "下车", "起点", "终点", "里程", "司机"),
+        score_bias=0.38,
+    ),
+    DocumentRule(
+        document_type="parking_toll_receipt",
+        document_type_label="停车/通行费票据",
+        scene_code="transport",
+        scene_label="交通票据",
+        expense_type="transport",
+        keywords=("停车费", "通行费", "过路费", "收费站", "停车场", "停车"),
+        score_bias=0.28,
+    ),
+    DocumentRule(
+        document_type="meal_receipt",
+        document_type_label="餐饮票据",
+        scene_code="meal",
+        scene_label="餐饮票据",
+        expense_type="meal",
+        keywords=("餐饮", "餐费", "用餐", "饭店", "酒楼", "餐厅", "食品", "外卖", "咖啡"),
+        score_bias=0.14,
+    ),
+    DocumentRule(
+        document_type="office_invoice",
+        document_type_label="办公用品票据",
+        scene_code="office",
+        scene_label="办公用品票据",
+        expense_type="office",
+        keywords=("办公用品", "文具", "耗材", "打印纸", "墨盒", "硒鼓", "键盘", "鼠标"),
+        score_bias=0.14,
+    ),
+    DocumentRule(
+        document_type="meeting_invoice",
+        document_type_label="会议/会务票据",
+        scene_code="meeting",
+        scene_label="会务票据",
+        expense_type="meeting",
+        keywords=("会议", "会务", "会展", "论坛", "会议室", "会场"),
+        score_bias=0.12,
+    ),
+    DocumentRule(
+        document_type="training_invoice",
+        document_type_label="培训票据",
+        scene_code="training",
+        scene_label="培训票据",
+        expense_type="training",
+        keywords=("培训", "课程", "讲师", "教材", "学费", "认证"),
+        score_bias=0.12,
+    ),
+    DocumentRule(
+        document_type="vat_invoice",
+        document_type_label="增值税发票",
+        scene_code="other",
+        scene_label="通用发票",
+        expense_type="other",
+        keywords=("发票代码", "发票号码", "价税合计", "增值税", "电子发票"),
+        score_bias=-0.08,
+    ),
+    DocumentRule(
+        document_type="receipt",
+        document_type_label="一般收据/凭证",
+        scene_code="other",
+        scene_label="其他票据",
+        expense_type="other",
+        keywords=("收据", "凭证", "票据"),
+        score_bias=-0.18,
+    ),
+)
+
+DOCUMENT_TYPE_RULE_MAP = {rule.document_type: rule for rule in DOCUMENT_RULES}
+SUPPORTED_DOCUMENT_TYPES = tuple(DOCUMENT_TYPE_RULE_MAP.keys()) + ("other",)
+
+AMOUNT_PATTERNS = (
+    re.compile(r"(?:价税合计|合计|金额|总额|票价|支付金额|实付金额|实收金额)[：:\s￥¥]*([0-9]+(?:[.,][0-9]{1,2})?)"),
+    re.compile(r"([0-9]+(?:[.,][0-9]{1,2})?)\s*元"),
+)
+DATE_PATTERN = re.compile(r"((?:20\d{2}|19\d{2})[-/年.](?:1[0-2]|0?[1-9])[-/月.](?:3[01]|[12]\d|0?[1-9])日?)")
+INVOICE_NUMBER_PATTERN = re.compile(r"(?:发票号码|票号|单号|订单号)[：:\s]*([A-Za-z0-9-]{6,24})")
+INVOICE_CODE_PATTERN = re.compile(r"(?:发票代码)[：:\s]*([A-Za-z0-9-]{6,24})")
+TRIP_NO_PATTERN = re.compile(r"(?:车次|航班(?:号)?)[：:\s]*([A-Za-z0-9]{2,12})")
+ROUTE_PATTERN = re.compile(r"([\u4e00-\u9fa5]{2,12})\s*(?:至|→|->|-)\s*([\u4e00-\u9fa5]{2,12})")
+MERCHANT_PATTERNS = (
+    re.compile(r"(?:销售方(?:名称)?|商户(?:名称)?|开票方(?:名称)?|收款方(?:名称)?)[：:\s]*([A-Za-z0-9\u4e00-\u9fa5（）()·&\\-]{2,40})"),
+    re.compile(r"([A-Za-z0-9\u4e00-\u9fa5（）()·&\\-]{2,40}(?:酒店|宾馆|饭店|酒楼|餐厅|航空|铁路|滴滴出行|停车场|服务区))"),
+)
+
+
+class DocumentIntelligenceService:
+    def __init__(self, db: Session | None = None) -> None:
+        self.runtime_chat_service = RuntimeChatService(db) if db is not None else None
+
+    def build_document_insight(
+        self,
+        *,
+        filename: str = "",
+        summary: str = "",
+        text: str = "",
+        preview_data_url: str = "",
+    ) -> DocumentInsight:
+        raw_text = " ".join(
+            [str(filename or "").strip(), str(summary or "").strip(), str(text or "").strip()]
+        ).strip()
+        compact = re.sub(r"\s+", "", raw_text).lower()
+        rule_match = _match_document_rule(compact)
+        base_rule = rule_match.rule or DEFAULT_RULE
+        fields = tuple(_extract_document_fields(raw_text))
+        rule_insight = DocumentInsight(
+            document_type=base_rule.document_type,
+            document_type_label=base_rule.document_type_label,
+            scene_code=base_rule.scene_code,
+            scene_label=base_rule.scene_label,
+            expense_type=base_rule.expense_type,
+            fields=fields,
+            classification_source="rule",
+            classification_confidence=rule_match.confidence,
+            evidence=rule_match.evidence,
+        )
+
+        llm_result = self._classify_with_model(
+            filename=str(filename or "").strip(),
+            summary=str(summary or "").strip(),
+            text=str(text or "").strip(),
+            preview_data_url=str(preview_data_url or "").strip(),
+            rule_insight=rule_insight,
+            fields=fields,
+        )
+        if llm_result is None:
+            return rule_insight
+        return self._merge_rule_and_model(
+            rule_insight=rule_insight,
+            llm_result=llm_result,
+            fields=fields,
+            has_preview=bool(preview_data_url),
+        )
+
+    def _classify_with_model(
+        self,
+        *,
+        filename: str,
+        summary: str,
+        text: str,
+        preview_data_url: str,
+        rule_insight: DocumentInsight,
+        fields: tuple[DocumentField, ...],
+    ) -> tuple[str, LlmDocumentClassification] | None:
+        if self.runtime_chat_service is None:
+            return None
+
+        trimmed_text = text.strip()
+        if not trimmed_text and not summary.strip():
+            return None
+
+        facts = {
+            "filename": filename,
+            "summary": summary[:300],
+            "ocr_text_excerpt": trimmed_text[:2000],
+            "rule_candidate": {
+                "document_type": rule_insight.document_type,
+                "document_type_label": rule_insight.document_type_label,
+                "scene_code": rule_insight.scene_code,
+                "scene_label": rule_insight.scene_label,
+                "expense_type": rule_insight.expense_type,
+                "confidence": round(rule_insight.classification_confidence, 2),
+                "evidence": list(rule_insight.evidence),
+            },
+            "extracted_fields": [
+                {"key": field.key, "label": field.label, "value": field.value}
+                for field in fields
+            ],
+            "allowed_document_types": list(SUPPORTED_DOCUMENT_TYPES),
+        }
+
+        system_prompt = (
+            "你是企业报销票据识别复核器。"
+            "你的任务不是 OCR，而是在已有 OCR 文本和票据预览基础上判断票据类型。"
+            "只输出 JSON 对象，不要输出 Markdown、解释或代码块。"
+            "document_type 只能是："
+            f"{', '.join(SUPPORTED_DOCUMENT_TYPES)}。"
+            "如果证据不足，返回 other。"
+            "严禁编造 OCR 中不存在的商户、酒店、航司、路线或金额。"
+            "如果 OCR 出现冲突碎片，应优先依据票据主体信息，而不是单个噪声词。"
+            "例如滴滴行程单/网约车发票，即使 OCR 混入酒店名称，也不能直接判成酒店票据。"
+            "输出字段：document_type, scene_code, scene_label, expense_type, confidence, evidence。"
+        )
+        user_prompt = (
+            "请根据以下票据事实给出最终分类 JSON：\n"
+            f"{json.dumps(facts, ensure_ascii=False, indent=2)}\n\n"
+            "示例输出：\n"
+            "{\n"
+            '  "document_type": "taxi_receipt",\n'
+            '  "scene_code": "transport",\n'
+            '  "scene_label": "交通票据",\n'
+            '  "expense_type": "transport",\n'
+            '  "confidence": 0.86,\n'
+            '  "evidence": ["OCR 中出现 滴滴出行、订单号、上车/下车 等交通特征"]\n'
+            "}"
+        )
+
+        if preview_data_url:
+            response_text = self.runtime_chat_service.complete(
+                [
+                    {"role": "system", "content": system_prompt},
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "text", "text": user_prompt},
+                            {"type": "image_url", "image_url": {"url": preview_data_url}},
+                        ],
+                    },
+                ],
+                slot_priority=("vlm",),
+                max_tokens=320,
+                temperature=0.0,
+            )
+            parsed = self._parse_llm_payload(response_text)
+            if parsed is not None:
+                return "llm_vision", parsed
+
+        response_text = self.runtime_chat_service.complete(
+            [
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_prompt},
+            ],
+            slot_priority=("main", "backup"),
+            max_tokens=320,
+            temperature=0.0,
+        )
+        parsed = self._parse_llm_payload(response_text)
+        if parsed is not None:
+            return "llm_text", parsed
+        return None
+
+    @staticmethod
+    def _parse_llm_payload(response_text: str | None) -> LlmDocumentClassification | None:
+        payload_json = _extract_json_payload(response_text)
+        if payload_json is None:
+            return None
+
+        try:
+            parsed = LlmDocumentClassification.model_validate(payload_json)
+        except ValidationError:
+            return None
+
+        normalized_type = str(parsed.document_type or "other").strip().lower() or "other"
+        if normalized_type not in SUPPORTED_DOCUMENT_TYPES:
+            normalized_type = "other"
+
+        base_rule = DOCUMENT_TYPE_RULE_MAP.get(normalized_type, DEFAULT_RULE)
+        evidence = [
+            str(item or "").strip()
+            for item in parsed.evidence
+            if str(item or "").strip()
+        ][:4]
+
+        return LlmDocumentClassification(
+            document_type=normalized_type,
+            scene_code=str(parsed.scene_code or base_rule.scene_code).strip() or base_rule.scene_code,
+            scene_label=str(parsed.scene_label or base_rule.scene_label).strip() or base_rule.scene_label,
+            expense_type=str(parsed.expense_type or base_rule.expense_type).strip() or base_rule.expense_type,
+            confidence=float(parsed.confidence),
+            evidence=evidence,
+        )
+
+    @staticmethod
+    def _merge_rule_and_model(
+        *,
+        rule_insight: DocumentInsight,
+        llm_result: tuple[str, LlmDocumentClassification],
+        fields: tuple[DocumentField, ...],
+        has_preview: bool,
+    ) -> DocumentInsight:
+        source, parsed = llm_result
+        if parsed.confidence < 0.55:
+            return rule_insight
+
+        should_override = False
+        if parsed.document_type == rule_insight.document_type:
+            should_override = True
+        elif rule_insight.document_type == "other" and parsed.document_type != "other":
+            should_override = True
+        elif parsed.document_type != "other":
+            threshold = 0.60 if has_preview else max(0.76, rule_insight.classification_confidence + 0.12)
+            should_override = parsed.confidence >= threshold
+
+        if not should_override:
+            return rule_insight
+
+        rule = DOCUMENT_TYPE_RULE_MAP.get(parsed.document_type, DEFAULT_RULE)
+        warnings = list(rule_insight.warnings)
+        if parsed.document_type != rule_insight.document_type:
+            warnings.append("票据类型已结合大模型复核结果修正，建议人工再核对原图。")
+
+        return DocumentInsight(
+            document_type=rule.document_type,
+            document_type_label=rule.document_type_label,
+            scene_code=rule.scene_code if parsed.scene_code == "other" else parsed.scene_code,
+            scene_label=rule.scene_label if parsed.scene_label == "其他票据" else parsed.scene_label,
+            expense_type=rule.expense_type if parsed.expense_type == "other" else parsed.expense_type,
+            fields=fields,
+            classification_source=source,
+            classification_confidence=max(parsed.confidence, rule_insight.classification_confidence),
+            evidence=tuple(parsed.evidence or rule_insight.evidence),
+            warnings=tuple(warnings),
+        )
+
+
+def build_document_insight(
+    *,
+    filename: str = "",
+    summary: str = "",
+    text: str = "",
+    preview_data_url: str = "",
+) -> DocumentInsight:
+    return DocumentIntelligenceService().build_document_insight(
+        filename=filename,
+        summary=summary,
+        text=text,
+        preview_data_url=preview_data_url,
+    )
+
+
+def _match_document_rule(compact_text: str) -> RuleMatch:
+    best_rule = DEFAULT_RULE
+    best_evidence: tuple[str, ...] = ()
+    best_score = 0.0
+
+    for rule in DOCUMENT_RULES:
+        matched = tuple(keyword for keyword in rule.keywords if keyword.lower() in compact_text)
+        if not matched:
+            continue
+        score = float(rule.score_bias) + len(matched) * 0.92 + sum(min(len(keyword), 6) * 0.08 for keyword in matched)
+        if score > best_score:
+            best_rule = rule
+            best_evidence = matched
+            best_score = score
+
+    if best_score <= 0:
+        return RuleMatch(rule=None, confidence=0.0, evidence=(), score=0.0)
+
+    confidence = min(0.94, 0.30 + min(best_score, 4.8) * 0.12)
+    return RuleMatch(
+        rule=best_rule,
+        confidence=round(confidence, 2),
+        evidence=best_evidence[:4],
+        score=best_score,
+    )
+
+
+def _extract_json_payload(response_text: str | None) -> dict[str, Any] | None:
+    if not response_text:
+        return None
+
+    cleaned = re.sub(r"<think>.*?</think>", "", response_text, flags=re.DOTALL | re.IGNORECASE).strip()
+    if not cleaned:
+        return None
+
+    fenced_match = re.search(r"```(?:json)?\s*(\{.*\})\s*```", cleaned, flags=re.DOTALL)
+    candidates = [fenced_match.group(1)] if fenced_match else []
+    candidates.append(cleaned)
+
+    start = cleaned.find("{")
+    end = cleaned.rfind("}")
+    if start != -1 and end != -1 and end > start:
+        candidates.append(cleaned[start : end + 1])
+
+    for candidate in candidates:
+        try:
+            parsed = json.loads(candidate)
+        except json.JSONDecodeError:
+            continue
+        if isinstance(parsed, dict):
+            return parsed
+    return None
+
+
+def _extract_document_fields(text: str) -> list[DocumentField]:
+    fields: list[DocumentField] = []
+    amount = _extract_amount(text)
+    if amount:
+        fields.append(DocumentField(key="amount", label="金额", value=amount))
+
+    date_value = _extract_date(text)
+    if date_value:
+        fields.append(DocumentField(key="date", label="日期", value=date_value))
+
+    merchant = _extract_merchant(text)
+    if merchant:
+        fields.append(DocumentField(key="merchant_name", label="商户", value=merchant))
+
+    invoice_number = _extract_pattern(INVOICE_NUMBER_PATTERN, text)
+    if invoice_number:
+        fields.append(DocumentField(key="invoice_number", label="票据号码", value=invoice_number))
+
+    invoice_code = _extract_pattern(INVOICE_CODE_PATTERN, text)
+    if invoice_code:
+        fields.append(DocumentField(key="invoice_code", label="发票代码", value=invoice_code))
+
+    trip_no = _extract_pattern(TRIP_NO_PATTERN, text)
+    if trip_no:
+        fields.append(DocumentField(key="trip_no", label="车次/航班", value=trip_no))
+
+    route = _extract_route(text)
+    if route:
+        fields.append(DocumentField(key="route", label="行程", value=route))
+
+    return fields
+
+
+def _extract_amount(text: str) -> str:
+    best_value: Decimal | None = None
+    for pattern in AMOUNT_PATTERNS:
+        for match in pattern.finditer(text):
+            raw_value = str(match.group(1) or "").replace(",", ".").strip()
+            try:
+                candidate = Decimal(raw_value)
+            except InvalidOperation:
+                continue
+            if candidate <= Decimal("0.00"):
+                continue
+            if best_value is None or candidate > best_value:
+                best_value = candidate
+        if best_value is not None:
+            break
+
+    if best_value is None:
+        return ""
+    normalized = best_value.quantize(Decimal("0.01"))
+    text_value = format(normalized, "f").rstrip("0").rstrip(".")
+    return f"{text_value}元"
+
+
+def _extract_date(text: str) -> str:
+    match = DATE_PATTERN.search(text)
+    if not match:
+        return ""
+    raw_value = str(match.group(1) or "").strip()
+    normalized = raw_value.replace("年", "-").replace("月", "-").replace("日", "")
+    normalized = normalized.replace("/", "-").replace(".", "-")
+    parts = [part for part in normalized.split("-") if part]
+    if len(parts) != 3:
+        return raw_value
+    year, month, day = parts
+    return f"{year.zfill(4)}-{month.zfill(2)}-{day.zfill(2)}"
+
+
+def _extract_merchant(text: str) -> str:
+    for pattern in MERCHANT_PATTERNS:
+        match = pattern.search(text)
+        if not match:
+            continue
+        value = _clean_field_value(match.group(1))
+        if value:
+            return value
+    return ""
+
+
+def _extract_route(text: str) -> str:
+    match = ROUTE_PATTERN.search(text)
+    if not match:
+        return ""
+    start = _clean_field_value(match.group(1))
+    end = _clean_field_value(match.group(2))
+    if not start or not end or start == end:
+        return ""
+    return f"{start}-{end}"
+
+
+def _extract_pattern(pattern: re.Pattern[str], text: str) -> str:
+    match = pattern.search(text)
+    if not match:
+        return ""
+    return _clean_field_value(match.group(1))
+
+
+def _clean_field_value(value: str) -> str:
+    return str(value or "").strip().strip("：:，,。.；;")

server/src/app/services/ocr.py

@@ -1,21 +1,55 @@
 from __future__ import annotations
 
+import base64
 import json
 import shutil
 import subprocess
+from dataclasses import dataclass, field
 from pathlib import Path
 from uuid import uuid4
 
+from sqlalchemy.orm import Session
+
 from app.core.config import SERVER_DIR, get_settings
-from app.schemas.ocr import OcrRecognizeBatchRead, OcrRecognizeDocumentRead, OcrRecognizeLineRead
+from app.schemas.ocr import OcrRecognizeBatchRead, OcrRecognizeDocumentRead, OcrRecognizeFieldRead, OcrRecognizeLineRead
+from app.services.document_intelligence import DocumentIntelligenceService
 
 WORKER_JSON_PREFIX = "__OCR_JSON__="
 SUPPORTED_SUFFIXES = {".png", ".jpg", ".jpeg", ".webp", ".bmp", ".pdf"}
 
 
+@dataclass(slots=True)
+class PreparedOcrInput:
+    input_path: Path
+    source_key: str
+    filename: str
+    media_type: str
+    page_index: int | None = None
+    preview_kind: str = ""
+    preview_data_url: str = ""
+
+
+@dataclass(slots=True)
+class AggregatedOcrDocument:
+    filename: str
+    media_type: str
+    source_key: str
+    engine: str = "paddleocr_mobile"
+    model: str = "PP-OCRv5_mobile"
+    summary_fragments: list[str] = field(default_factory=list)
+    text_fragments: list[str] = field(default_factory=list)
+    score_values: list[float] = field(default_factory=list)
+    warnings: list[str] = field(default_factory=list)
+    lines: list[OcrRecognizeLineRead] = field(default_factory=list)
+    page_count: int = 0
+    preview_kind: str = ""
+    preview_data_url: str = ""
+
+
 class OcrService:
-    def __init__(self) -> None:
+    def __init__(self, db: Session | None = None) -> None:
         self.settings = get_settings()
+        self.document_intelligence_service = DocumentIntelligenceService(db)
 
     def recognize_files(
         self,
@@ -28,10 +62,11 @@ class OcrService:
         temp_root.mkdir(parents=True, exist_ok=True)
 
         documents: list[OcrRecognizeDocumentRead] = []
-        input_paths: list[Path] = []
-        meta_by_path: dict[str, tuple[str, str]] = {}
+        prepared_inputs: list[PreparedOcrInput] = []
+        cleanup_paths: list[Path] = []
         python_bin = self._resolve_python_bin()
         worker_path = self._resolve_worker_path()
+        worker_payload: dict = {}
 
         try:
             for filename, content, media_type in files:
@@ -73,17 +108,55 @@ class OcrService:
 
                 temp_path = temp_root / f"{uuid4().hex}{suffix}"
                 temp_path.write_bytes(content)
-                input_paths.append(temp_path)
-                meta_by_path[str(temp_path)] = (normalized_name, resolved_media_type)
+                cleanup_paths.append(temp_path)
 
-            if input_paths:
+                if suffix == ".pdf":
+                    try:
+                        prepared_inputs.extend(
+                            self._prepare_pdf_inputs(
+                                pdf_path=temp_path,
+                                filename=normalized_name,
+                                media_type=resolved_media_type,
+                                cleanup_paths=cleanup_paths,
+                            )
+                        )
+                    except RuntimeError as exc:
+                        documents.append(
+                            OcrRecognizeDocumentRead(
+                                filename=normalized_name,
+                                media_type=resolved_media_type,
+                                warnings=[str(exc)],
+                            )
+                        )
+                    continue
+
+                prepared_inputs.append(
+                    PreparedOcrInput(
+                        input_path=temp_path,
+                        source_key=uuid4().hex,
+                        filename=normalized_name,
+                        media_type=resolved_media_type,
+                        preview_kind="image" if resolved_media_type.startswith("image/") else "",
+                        preview_data_url=(
+                            self._build_preview_data_url(temp_path, media_type=resolved_media_type)
+                            if resolved_media_type.startswith("image/")
+                            else ""
+                        ),
+                    )
+                )
+
+            if prepared_inputs:
                 worker_payload = self._invoke_worker(
                     python_bin=python_bin,
                     worker_path=worker_path,
-                    input_paths=input_paths,
+                    input_paths=[item.input_path for item in prepared_inputs],
+                )
+                documents.extend(
+                    self._build_documents(
+                        worker_documents=worker_payload.get("documents", []),
+                        prepared_inputs=prepared_inputs,
+                    )
                 )
-                for item in worker_payload.get("documents", []):
-                    documents.append(self._build_document(item, meta_by_path))
 
             success_count = sum(
                 1
@@ -92,12 +165,12 @@ class OcrService:
             )
             engine = (
                 str(worker_payload.get("engine", "paddleocr_mobile"))
-                if input_paths
+                if prepared_inputs
                 else "paddleocr_mobile"
             )
             model = (
                 str(worker_payload.get("model", "PP-OCRv5_mobile"))
-                if input_paths
+                if prepared_inputs
                 else "PP-OCRv5_mobile"
             )
             return OcrRecognizeBatchRead(
@@ -108,8 +181,7 @@ class OcrService:
                 documents=documents,
             )
         finally:
-            for path in input_paths:
-                path.unlink(missing_ok=True)
+            self._cleanup_temp_paths(cleanup_paths)
 
     def _resolve_python_bin(self) -> str:
         candidates = []
@@ -182,17 +254,182 @@ class OcrService:
                 return json.loads(normalized[len(WORKER_JSON_PREFIX) :])
         return None
 
-    @staticmethod
-    def _build_document(
-        payload: dict,
-        meta_by_path: dict[str, tuple[str, str]],
-    ) -> OcrRecognizeDocumentRead:
-        input_path = str(payload.get("input_path") or "")
-        filename, media_type = meta_by_path.get(
-            input_path,
-            (Path(input_path).name or "upload.bin", "application/octet-stream"),
+    def _prepare_pdf_inputs(
+        self,
+        *,
+        pdf_path: Path,
+        filename: str,
+        media_type: str,
+        cleanup_paths: list[Path],
+    ) -> list[PreparedOcrInput]:
+        output_dir = pdf_path.with_suffix("")
+        output_dir.mkdir(parents=True, exist_ok=True)
+        cleanup_paths.append(output_dir)
+
+        image_paths = self._convert_pdf_to_images(pdf_path=pdf_path, output_dir=output_dir)
+        if not image_paths:
+            raise RuntimeError("PDF 转图片后未生成可识别页面。")
+
+        preview_data_url = self._build_preview_data_url(image_paths[0], media_type="image/png")
+        source_key = uuid4().hex
+        descriptors: list[PreparedOcrInput] = []
+        for page_index, image_path in enumerate(image_paths):
+            descriptors.append(
+                PreparedOcrInput(
+                    input_path=image_path,
+                    source_key=source_key,
+                    filename=filename,
+                    media_type=media_type,
+                    page_index=page_index,
+                    preview_kind="image" if page_index == 0 else "",
+                    preview_data_url=preview_data_url if page_index == 0 else "",
                 )
-        lines = [
+            )
+        return descriptors
+
+    def _convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> list[Path]:
+        prefix = output_dir / "page"
+        completed = subprocess.run(
+            [
+                "pdftoppm",
+                "-png",
+                "-r",
+                "160",
+                str(pdf_path),
+                str(prefix),
+            ],
+            capture_output=True,
+            text=True,
+            timeout=self.settings.ocr_timeout_seconds,
+            check=False,
+        )
+        if completed.returncode != 0:
+            detail = (completed.stderr or completed.stdout or "").strip()
+            raise RuntimeError(f"PDF 转图片失败：{detail or 'pdftoppm 返回非 0 状态码。'}")
+
+        return sorted(output_dir.glob("page-*.png"), key=self._extract_pdf_page_sort_key)
+
+    @staticmethod
+    def _extract_pdf_page_sort_key(path: Path) -> tuple[int, str]:
+        suffix = path.stem.rsplit("-", 1)[-1]
+        try:
+            return int(suffix), path.name
+        except ValueError:
+            return 0, path.name
+
+    @staticmethod
+    def _build_preview_data_url(path: Path, *, media_type: str) -> str:
+        encoded = base64.b64encode(path.read_bytes()).decode("ascii")
+        return f"data:{media_type};base64,{encoded}"
+
+    def _build_documents(
+        self,
+        *,
+        worker_documents: list[dict],
+        prepared_inputs: list[PreparedOcrInput],
+    ) -> list[OcrRecognizeDocumentRead]:
+        descriptor_by_path = {str(item.input_path): item for item in prepared_inputs}
+        source_order: list[str] = []
+        seen_sources: set[str] = set()
+        for item in prepared_inputs:
+            if item.source_key in seen_sources:
+                continue
+            seen_sources.add(item.source_key)
+            source_order.append(item.source_key)
+
+        aggregated_by_source: dict[str, AggregatedOcrDocument] = {}
+        for payload in worker_documents:
+            if not isinstance(payload, dict):
+                continue
+            input_path = str(payload.get("input_path") or "")
+            descriptor = descriptor_by_path.get(input_path)
+            if descriptor is None:
+                continue
+
+            aggregated = aggregated_by_source.get(descriptor.source_key)
+            if aggregated is None:
+                aggregated = AggregatedOcrDocument(
+                    filename=descriptor.filename,
+                    media_type=descriptor.media_type,
+                    source_key=descriptor.source_key,
+                    engine=str(payload.get("engine", "paddleocr_mobile")),
+                    model=str(payload.get("model", "PP-OCRv5_mobile")),
+                )
+                aggregated_by_source[descriptor.source_key] = aggregated
+
+            aggregated.page_count = max(
+                aggregated.page_count,
+                (descriptor.page_index + 1)
+                if descriptor.page_index is not None
+                else int(payload.get("page_count", 1) or 1),
+            )
+            if descriptor.preview_kind and not aggregated.preview_kind:
+                aggregated.preview_kind = descriptor.preview_kind
+            if descriptor.preview_data_url and not aggregated.preview_data_url:
+                aggregated.preview_data_url = descriptor.preview_data_url
+
+            page_summary = str(payload.get("summary", "") or "").strip()
+            if page_summary:
+                aggregated.summary_fragments.append(page_summary)
+
+            page_text = str(payload.get("text", "") or "").strip()
+            if page_text:
+                aggregated.text_fragments.append(page_text)
+
+            lines = self._build_lines(
+                payload.get("lines", []),
+                page_index_override=descriptor.page_index,
+            )
+            aggregated.lines.extend(lines)
+            aggregated.score_values.extend(line.score for line in lines if line.score > 0)
+
+            if not lines:
+                avg_score = float(payload.get("avg_score", 0.0) or 0.0)
+                if avg_score > 0:
+                    aggregated.score_values.append(avg_score)
+
+            for warning in payload.get("warnings", []):
+                normalized_warning = str(warning or "").strip()
+                if normalized_warning and normalized_warning not in aggregated.warnings:
+                    aggregated.warnings.append(normalized_warning)
+
+        documents: list[OcrRecognizeDocumentRead] = []
+        for source_key in source_order:
+            descriptors = [item for item in prepared_inputs if item.source_key == source_key]
+            if not descriptors:
+                continue
+            aggregated = aggregated_by_source.get(source_key)
+            if aggregated is None:
+                first_descriptor = descriptors[0]
+                documents.append(
+                    OcrRecognizeDocumentRead(
+                        filename=first_descriptor.filename,
+                        media_type=first_descriptor.media_type,
+                        page_count=max(1, len(descriptors)),
+                        preview_kind=first_descriptor.preview_kind,
+                        preview_data_url=first_descriptor.preview_data_url,
+                        warnings=["OCR worker 未返回该文件的识别结果。"],
+                    )
+                )
+                continue
+            documents.append(self._finalize_document(aggregated))
+
+        return documents
+
+    @staticmethod
+    def _build_lines(
+        items: list[dict],
+        *,
+        page_index_override: int | None = None,
+    ) -> list[OcrRecognizeLineRead]:
+        lines: list[OcrRecognizeLineRead] = []
+        for item in items:
+            if not isinstance(item, dict):
+                continue
+            page_index = page_index_override
+            if page_index is None and item.get("page_index") is not None:
+                page_index = int(item["page_index"])
+            lines.append(
                 OcrRecognizeLineRead(
                     text=str(item.get("text", "")),
                     score=float(item.get("score", 0.0) or 0.0),
@@ -201,21 +438,74 @@ class OcrService:
                         for point in item.get("box", [])
                         if isinstance(point, list) and len(point) == 2
                     ],
-                page_index=int(item["page_index"]) if item.get("page_index") is not None else None,
+                    page_index=page_index,
                 )
-            for item in payload.get("lines", [])
-            if isinstance(item, dict)
-        ]
+            )
+        return lines
+
+    @staticmethod
+    def _truncate_summary(parts: list[str]) -> str:
+        summary = "；".join([part for part in parts if part][:3])
+        if len(summary) > 180:
+            return f"{summary[:177]}..."
+        return summary
+
+    def _finalize_document(self, aggregated: AggregatedOcrDocument) -> OcrRecognizeDocumentRead:
+        full_text = "\n".join(fragment for fragment in aggregated.text_fragments if fragment).strip()
+        summary = self._truncate_summary(aggregated.summary_fragments or aggregated.text_fragments)
+        insight = self.document_intelligence_service.build_document_insight(
+            filename=aggregated.filename,
+            summary=summary,
+            text=full_text,
+            preview_data_url=aggregated.preview_data_url,
+        )
+        warnings = list(aggregated.warnings)
+        for warning in insight.warnings:
+            normalized_warning = str(warning or "").strip()
+            if normalized_warning and normalized_warning not in warnings:
+                warnings.append(normalized_warning)
         return OcrRecognizeDocumentRead(
-            filename=filename,
-            media_type=media_type,
-            engine=str(payload.get("engine", "paddleocr_mobile")),
-            model=str(payload.get("model", "PP-OCRv5_mobile")),
-            text=str(payload.get("text", "")),
-            summary=str(payload.get("summary", "")),
-            avg_score=float(payload.get("avg_score", 0.0) or 0.0),
-            line_count=int(payload.get("line_count", len(lines)) or 0),
-            page_count=int(payload.get("page_count", 1) or 1),
-            warnings=[str(item) for item in payload.get("warnings", [])],
-            lines=lines,
+            filename=aggregated.filename,
+            media_type=aggregated.media_type,
+            engine=aggregated.engine,
+            model=aggregated.model,
+            text=full_text,
+            summary=summary,
+            avg_score=(
+                sum(aggregated.score_values) / len(aggregated.score_values)
+                if aggregated.score_values
+                else 0.0
+            ),
+            line_count=len(aggregated.lines),
+            page_count=max(1, aggregated.page_count),
+            document_type=insight.document_type,
+            document_type_label=insight.document_type_label,
+            scene_code=insight.scene_code,
+            scene_label=insight.scene_label,
+            classification_source=insight.classification_source,
+            classification_confidence=insight.classification_confidence,
+            classification_evidence=list(insight.evidence),
+            document_fields=[
+                OcrRecognizeFieldRead(
+                    key=field.key,
+                    label=field.label,
+                    value=field.value,
                 )
+                for field in insight.fields
+            ],
+            preview_kind=aggregated.preview_kind,
+            preview_data_url=aggregated.preview_data_url,
+            warnings=warnings,
+            lines=sorted(
+                aggregated.lines,
+                key=lambda item: item.page_index if item.page_index is not None else -1,
+            ),
+        )
+
+    @staticmethod
+    def _cleanup_temp_paths(paths: list[Path]) -> None:
+        for path in reversed(paths):
+            if path.is_dir():
+                shutil.rmtree(path, ignore_errors=True)
+                continue
+            path.unlink(missing_ok=True)

server/tests/test_document_intelligence.py

@@ -0,0 +1,66 @@
+from __future__ import annotations
+
+import json
+
+from sqlalchemy import create_engine
+from sqlalchemy.orm import sessionmaker
+from sqlalchemy.pool import StaticPool
+
+from app.services.document_intelligence import DocumentIntelligenceService, build_document_insight
+from app.services.runtime_chat import RuntimeChatService
+
+
+def test_build_document_insight_prefers_transport_for_didi_text_with_hotel_noise() -> None:
+    insight = build_document_insight(
+        filename="didi-trip.png",
+        summary="滴滴出行行程单",
+        text="滴滴出行电子发票 订单号 12345678 上车点 深圳湾 下车点 后海 全季酒店 里程 12.4 公里 金额 48 元",
+    )
+
+    assert insight.document_type == "taxi_receipt"
+    assert insight.document_type_label == "出租车/网约车票据"
+    assert insight.scene_code == "transport"
+    assert any(field.label == "金额" and field.value == "48元" for field in insight.fields)
+
+
+def test_document_intelligence_service_uses_vlm_result_when_preview_available(monkeypatch) -> None:
+    calls: list[tuple[str, ...]] = []
+
+    def fake_complete(self, messages, *, slot_priority=("main", "backup"), max_tokens=500, temperature=0.2):
+        calls.append(slot_priority)
+        if slot_priority == ("vlm",):
+            assert isinstance(messages[1]["content"], list)
+            return json.dumps(
+                {
+                    "document_type": "taxi_receipt",
+                    "scene_code": "transport",
+                    "scene_label": "交通票据",
+                    "expense_type": "transport",
+                    "confidence": 0.91,
+                    "evidence": ["图片主体为滴滴行程单，OCR 中出现订单号、上车、下车等字段"],
+                },
+                ensure_ascii=False,
+            )
+        return None
+
+    monkeypatch.setattr(RuntimeChatService, "complete", fake_complete)
+
+    engine = create_engine(
+        "sqlite+pysqlite:///:memory:",
+        connect_args={"check_same_thread": False},
+        poolclass=StaticPool,
+    )
+    session = sessionmaker(bind=engine, autoflush=False, autocommit=False)()
+    try:
+        insight = DocumentIntelligenceService(session).build_document_insight(
+            filename="mixed-noise.png",
+            summary="OCR 混入酒店名称",
+            text="全季酒店 滴滴出行 订单号 12345678 上车 下车 金额 52 元",
+            preview_data_url="data:image/png;base64,ZmFrZQ==",
+        )
+    finally:
+        session.close()
+
+    assert insight.document_type == "taxi_receipt"
+    assert insight.classification_source == "llm_vision"
+    assert calls[0] == ("vlm",)

server/tests/test_ocr_endpoints.py

@@ -10,7 +10,7 @@ from sqlalchemy.pool import StaticPool
 from app.api.deps import get_db
 from app.db.base import Base
 from app.main import create_app
-from app.schemas.ocr import OcrRecognizeBatchRead, OcrRecognizeDocumentRead, OcrRecognizeLineRead
+from app.schemas.ocr import OcrRecognizeBatchRead, OcrRecognizeDocumentRead, OcrRecognizeFieldRead, OcrRecognizeLineRead
 from app.services.ocr import OcrService
 
 
@@ -50,14 +50,23 @@ def test_ocr_recognize_endpoint_returns_structured_payload(monkeypatch) -> None:
                 OcrRecognizeDocumentRead(
                     filename="invoice.png",
                     media_type="image/png",
-                    text="发票金额 100 元",
-                    summary="发票金额 100 元",
+                    text="增值税电子发票 发票号码12345678 金额 100 元 2026-05-13",
+                    summary="增值税电子发票，金额 100 元。",
                     avg_score=0.98,
                     line_count=1,
                     page_count=1,
+                    document_type="vat_invoice",
+                    document_type_label="增值税发票",
+                    scene_code="other",
+                    scene_label="通用发票",
+                    document_fields=[
+                        OcrRecognizeFieldRead(key="amount", label="金额", value="100元"),
+                        OcrRecognizeFieldRead(key="date", label="日期", value="2026-05-13"),
+                        OcrRecognizeFieldRead(key="invoice_number", label="票据号码", value="12345678"),
+                    ],
                     lines=[
                         OcrRecognizeLineRead(
-                            text="发票金额 100 元",
+                            text="增值税电子发票 发票号码12345678 金额 100 元 2026-05-13",
                             score=0.98,
                             box=[[1, 2], [10, 2], [10, 8], [1, 8]],
                             page_index=0,
@@ -81,4 +90,7 @@ def test_ocr_recognize_endpoint_returns_structured_payload(monkeypatch) -> None:
     assert payload["engine"] == "paddleocr_mobile"
     assert payload["success_count"] == 1
     assert payload["documents"][0]["filename"] == "invoice.png"
-    assert payload["documents"][0]["summary"] == "发票金额 100 元"
+    assert payload["documents"][0]["summary"] == "增值税电子发票，金额 100 元。"
+    assert payload["documents"][0]["document_type"] == "vat_invoice"
+    assert payload["documents"][0]["document_type_label"] == "增值税发票"
+    assert payload["documents"][0]["document_fields"][0]["label"] == "金额"

server/tests/test_ocr_service.py

@@ -26,15 +26,15 @@ for index, arg in enumerate(sys.argv):
                 "input_path": input_path,
                 "engine": "paddleocr_mobile",
                 "model": "PP-OCRv5_mobile",
-                "text": "发票金额 100 元",
-                "summary": "发票金额 100 元",
+                "text": "增值税电子发票 发票号码12345678 金额 100 元 2026-05-13",
+                "summary": "增值税电子发票，金额 100 元。",
                 "avg_score": 0.98,
                 "line_count": 1,
                 "page_count": 1,
                 "warnings": [],
                 "lines": [
                     {
-                        "text": "发票金额 100 元",
+                        "text": "增值税电子发票 发票号码12345678 金额 100 元 2026-05-13",
                         "score": 0.98,
                         "box": [[1, 2], [10, 2], [10, 8], [1, 8]],
                         "page_index": 0,
@@ -74,10 +74,106 @@ print("__OCR_JSON__=" + json.dumps(payload, ensure_ascii=False))
     assert len(result.documents) == 2
 
     recognized = next(item for item in result.documents if item.filename == "invoice.png")
-    assert recognized.summary == "发票金额 100 元"
+    assert recognized.summary == "增值税电子发票，金额 100 元。"
     assert recognized.line_count == 1
-    assert recognized.lines[0].text == "发票金额 100 元"
+    assert recognized.document_type == "vat_invoice"
+    assert recognized.document_type_label == "增值税发票"
+    assert any(field.label == "金额" and field.value == "100元" for field in recognized.document_fields)
+    assert any(field.label == "票据号码" and field.value == "12345678" for field in recognized.document_fields)
+    assert any(field.label == "日期" and field.value == "2026-05-13" for field in recognized.document_fields)
+    assert recognized.lines[0].text == "增值税电子发票 发票号码12345678 金额 100 元 2026-05-13"
 
     skipped = next(item for item in result.documents if item.filename == "notes.txt")
     assert skipped.line_count == 0
     assert skipped.warnings == ["当前仅支持图片和 PDF 文件进行 OCR。"]
+
+
+def test_ocr_service_converts_pdf_to_images_and_returns_image_preview(
+    monkeypatch,
+    tmp_path: Path,
+) -> None:
+    def fake_convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> list[Path]:
+        first = output_dir / "page-1.png"
+        second = output_dir / "page-2.png"
+        first.write_bytes(b"fake-page-1")
+        second.write_bytes(b"fake-page-2")
+        return [first, second]
+
+    def fake_invoke_worker(
+        self,
+        *,
+        python_bin: str,
+        worker_path: str,
+        input_paths: list[Path],
+    ) -> dict:
+        assert [path.name for path in input_paths] == ["page-1.png", "page-2.png"]
+        return {
+            "engine": "paddleocr_mobile",
+            "model": "PP-OCRv5_mobile",
+            "documents": [
+                {
+                    "input_path": str(input_paths[0]),
+                    "engine": "paddleocr_mobile",
+                    "model": "PP-OCRv5_mobile",
+                    "text": "高铁票 深圳北-广州南 车次 G1234 2026-05-13 金额 188 元",
+                    "summary": "高铁票第一页",
+                    "avg_score": 0.97,
+                    "line_count": 1,
+                    "page_count": 1,
+                    "warnings": [],
+                    "lines": [
+                        {
+                            "text": "高铁票 深圳北-广州南 车次 G1234 2026-05-13 金额 188 元",
+                            "score": 0.97,
+                            "box": [[1, 2], [10, 2], [10, 8], [1, 8]],
+                        }
+                    ],
+                },
+                {
+                    "input_path": str(input_paths[1]),
+                    "engine": "paddleocr_mobile",
+                    "model": "PP-OCRv5_mobile",
+                    "text": "乘车人 张三",
+                    "summary": "高铁票第二页",
+                    "avg_score": 0.94,
+                    "line_count": 1,
+                    "page_count": 1,
+                    "warnings": [],
+                    "lines": [
+                        {
+                            "text": "乘车人 张三",
+                            "score": 0.94,
+                            "box": [[1, 2], [10, 2], [10, 8], [1, 8]],
+                        }
+                    ],
+                },
+            ],
+        }
+
+    monkeypatch.setenv("STORAGE_ROOT_DIR", str(tmp_path / "storage"))
+    monkeypatch.setattr(OcrService, "_resolve_python_bin", lambda self: "python")
+    monkeypatch.setattr(OcrService, "_resolve_worker_path", lambda self: "worker.py")
+    monkeypatch.setattr(OcrService, "_convert_pdf_to_images", fake_convert_pdf_to_images)
+    monkeypatch.setattr(OcrService, "_invoke_worker", fake_invoke_worker)
+    get_settings.cache_clear()
+    try:
+        result = OcrService().recognize_files(
+            [
+                ("train-ticket.pdf", b"%PDF-1.4 fake", "application/pdf"),
+            ]
+        )
+    finally:
+        get_settings.cache_clear()
+
+    assert result.success_count == 1
+    assert len(result.documents) == 1
+    recognized = result.documents[0]
+    assert recognized.filename == "train-ticket.pdf"
+    assert recognized.page_count == 2
+    assert recognized.preview_kind == "image"
+    assert recognized.preview_data_url.startswith("data:image/png;base64,")
+    assert recognized.document_type == "train_ticket"
+    assert any(field.label == "金额" and field.value == "188元" for field in recognized.document_fields)
+    assert any(field.label == "车次/航班" and field.value == "G1234" for field in recognized.document_fields)
+    assert recognized.lines[0].page_index == 0
+    assert recognized.lines[1].page_index == 1