refactor(server): user_agent/steward/ocr 等服务重构并适配关联任务

- user_agent 拆分 application/locations/knowledge/response/review 四个子模块，接入申请位置语义与关联草稿分支
- steward planner/runtime/slot/plan_builder 决策链路重构，travel_reimbursement_calculator/orchestrator_expense_query 适配
- ocr/document_preview/document_intelligence/receipt_folder 复用预览与资产缓存，expense_claim_draft_flow/application_handoff 适配
- pyproject.toml 新增依赖，paddleocr bootstrap 脚本与 server_start.sh 调整
- 更新差旅/交通/通信等财务规则表，同步 document_intelligence/ocr/receipt_folder/user_agent 等测试

server/pyproject.toml

@@ -19,6 +19,7 @@ dependencies = [
   "python-dotenv>=1.0.1,<2.0.0",
   "email-validator>=2.2.0,<3.0.0",
   "python-multipart>=0.0.20,<1.0.0",
+  "jieba>=0.42.1,<0.43.0",
   "openpyxl>=3.1.5,<4.0.0",
   "lightrag-hku>=1.4.16,<1.5.0",
   "qdrant-client>=1.18.0,<2.0.0",

server/scripts/bootstrap_paddleocr_gpu.sh

@@ -14,7 +14,7 @@ if ! command -v "${PYTHON_BIN}" >/dev/null 2>&1; then
 fi
 
 apt-get update
-apt-get install -y --no-install-recommends libgl1 libglib2.0-0 poppler-utils poppler-data
+apt-get install -y --no-install-recommends libgl1 libglib2.0-0 poppler-utils poppler-data mupdf-tools
 
 rm -rf "${OCR_VENV_DIR}"
 "${PYTHON_BIN}" -m venv "${OCR_VENV_DIR}"

server/scripts/bootstrap_paddleocr_mobile.sh

@@ -13,7 +13,7 @@ if ! command -v "${PYTHON_BIN}" >/dev/null 2>&1; then
 fi
 
 apt-get update
-apt-get install -y --no-install-recommends libgl1 libglib2.0-0 poppler-utils poppler-data
+apt-get install -y --no-install-recommends libgl1 libglib2.0-0 poppler-utils poppler-data mupdf-tools
 
 "${PYTHON_BIN}" -m venv "${OCR_VENV_DIR}"
 "${OCR_VENV_DIR}/bin/pip" install --upgrade pip

server/server_start.sh

@@ -272,7 +272,7 @@ run_bootstrap_python() {
 }
 
 dependencies_ready() {
-    "$PYTHON_BIN" -c "import alembic, dotenv, email_validator, fastapi, jwt, lightrag, multipart, openpyxl, psycopg, pydantic_settings, qdrant_client, sqlalchemy, uvicorn" >/dev/null 2>&1
+    "$PYTHON_BIN" -c "import alembic, dotenv, email_validator, fastapi, jieba, jwt, lightrag, multipart, openpyxl, psycopg, pydantic_settings, qdrant_client, sqlalchemy, uvicorn" >/dev/null 2>&1
 }
 
 pip_ready() {

server/src/app/services/document_intelligence.py

@@ -562,7 +562,7 @@ def _extract_document_fields(text: str, document_type: str = "") -> list[Documen
     if date_value:
         append_field("date", "日期", date_value)
 
-    merchant = _extract_merchant(text)
+    merchant = "中国铁路" if normalized_type == "train_ticket" else _extract_merchant(text)
     if merchant:
         append_field("merchant_name", "商户", merchant)
 

server/src/app/services/document_preview.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import base64
 import binascii
 import mimetypes
+import os
 import re
 import shutil
 import subprocess
@@ -11,9 +12,19 @@ from pathlib import Path
 
 
 class DocumentPreviewAssets:
-    PDF_RENDERER_ID = "pdftoppm-png-r160-poppler-data"
+    PDF_RENDERER_ID = "pdf-raster-cjk-safe-v3"
     PDF_PREVIEW_MEDIA_TYPE = "image/png"
     PDF_PREVIEW_SUFFIX = ".png"
+    PDF_UNUSABLE_PREVIEW_ERRORS = (
+        "Missing language pack",
+        "Unknown font tag",
+        "No font in show",
+    )
+    POPPLER_DATA_DIR_CANDIDATES = (
+        "/usr/share/poppler",
+        "/usr/local/share/poppler",
+        "/opt/homebrew/share/poppler",
+    )
 
     @staticmethod
     def decode_data_url(payload: str) -> tuple[str, bytes] | None:
@@ -64,30 +75,117 @@ class DocumentPreviewAssets:
     ) -> Path:
         preview_path.parent.mkdir(parents=True, exist_ok=True)
         with tempfile.TemporaryDirectory(prefix=".pdf-preview-", dir=str(preview_path.parent)) as temp_dir:
-            prefix = Path(temp_dir) / "page"
+            pages = cls.render_pdf_pages(
+                pdf_path=pdf_path,
+                output_dir=Path(temp_dir),
+                timeout_seconds=timeout_seconds,
+            )
+            shutil.copyfile(pages[0], preview_path)
+        return preview_path
+
+    @classmethod
+    def render_pdf_pages(
+        cls,
+        *,
+        pdf_path: Path,
+        output_dir: Path,
+        timeout_seconds: int | float,
+        resolution: int = 160,
+    ) -> list[Path]:
+        output_dir.mkdir(parents=True, exist_ok=True)
+        errors: list[str] = []
+        for renderer_name, command in cls._pdf_render_commands(
+            pdf_path=pdf_path,
+            output_dir=output_dir,
+            resolution=resolution,
+        ):
+            executable = shutil.which(renderer_name)
+            if not executable:
+                errors.append(f"{renderer_name}: executable not found")
+                continue
+
+            cls._clear_rendered_pdf_pages(output_dir)
+            command[0] = executable
             completed = subprocess.run(
-                [
+                command,
-                    "pdftoppm",
-                    "-png",
-                    "-r",
-                    "160",
-                    str(pdf_path),
-                    str(prefix),
-                ],
                 capture_output=True,
                 text=True,
                 timeout=timeout_seconds,
                 check=False,
+                env=cls._pdf_render_env(),
             )
-            if completed.returncode != 0:
             detail = (completed.stderr or completed.stdout or "").strip()
-                raise RuntimeError(detail or "pdftoppm failed to render PDF preview.")
+            if completed.returncode != 0:
+                errors.append(f"{renderer_name}: {detail or 'renderer returned non-zero status'}")
+                continue
+            if cls.render_output_indicates_unusable_pdf_preview(detail):
+                errors.append(f"{renderer_name}: {detail or 'renderer produced unusable output'}")
+                continue
 
-            pages = sorted(Path(temp_dir).glob("page-*.png"), key=cls._extract_pdf_page_sort_key)
+            pages = sorted(output_dir.glob("page-*.png"), key=cls._extract_pdf_page_sort_key)
-            if not pages:
+            if pages:
-                raise RuntimeError("pdftoppm did not generate a preview image.")
+                return pages
-            shutil.copyfile(pages[0], preview_path)
+            errors.append(f"{renderer_name}: renderer did not generate PNG pages")
-        return preview_path
+
+        cls._clear_rendered_pdf_pages(output_dir)
+        detail = "；".join(errors[-3:])
+        raise RuntimeError(detail or "no PDF renderer generated usable PNG pages")
+
+    @classmethod
+    def render_output_indicates_unusable_pdf_preview(cls, output: str) -> bool:
+        return any(token in str(output or "") for token in cls.PDF_UNUSABLE_PREVIEW_ERRORS)
+
+    @classmethod
+    def _pdf_render_commands(
+        cls,
+        *,
+        pdf_path: Path,
+        output_dir: Path,
+        resolution: int,
+    ) -> list[tuple[str, list[str]]]:
+        prefix = output_dir / "page"
+        page_pattern = output_dir / "page-%d.png"
+        return [
+            (
+                "pdftoppm",
+                ["pdftoppm", "-png", "-r", str(resolution), str(pdf_path), str(prefix)],
+            ),
+            (
+                "mutool",
+                ["mutool", "draw", "-r", str(resolution), "-o", str(page_pattern), str(pdf_path)],
+            ),
+            (
+                "gs",
+                [
+                    "gs",
+                    "-dSAFER",
+                    "-dBATCH",
+                    "-dNOPAUSE",
+                    "-sDEVICE=png16m",
+                    f"-r{resolution}",
+                    f"-sOutputFile={page_pattern}",
+                    str(pdf_path),
+                ],
+            ),
+            (
+                "pdftocairo",
+                ["pdftocairo", "-png", "-r", str(resolution), str(pdf_path), str(prefix)],
+            ),
+        ]
+
+    @classmethod
+    def _pdf_render_env(cls) -> dict[str, str]:
+        env = os.environ.copy()
+        for candidate in cls.POPPLER_DATA_DIR_CANDIDATES:
+            if (Path(candidate) / "cMap").exists():
+                env.setdefault("POPPLER_DATADIR", candidate)
+                break
+        return env
+
+    @staticmethod
+    def _clear_rendered_pdf_pages(output_dir: Path) -> None:
+        for page in output_dir.glob("page-*.png"):
+            page.unlink(missing_ok=True)
 
     @staticmethod
     def _extract_pdf_page_sort_key(path: Path) -> tuple[int, str]:

server/src/app/services/expense_claim_application_handoff.py

@@ -9,7 +9,10 @@ from sqlalchemy import or_, select
 
 from app.models.financial_record import ExpenseClaim
 from app.services.expense_claim_risk_stage import with_risk_business_stage
-from app.services.expense_claim_workflow_constants import APPLICATION_ARCHIVE_STAGE
+from app.services.expense_claim_workflow_constants import (
+    APPLICATION_ARCHIVE_STAGE,
+    APPLICATION_LINK_STATUS_STAGE,
+)
 
 
 APPLICATION_REIMBURSEMENT_TYPE_MAP = {
@@ -248,3 +251,151 @@ class ExpenseClaimApplicationHandoffMixin:
             )
 
         return archived_applications
+
+    @staticmethod
+    def _reference_matches_deleted_reimbursement(
+        flag: dict[str, Any],
+        *,
+        reimbursement_claim_id: str,
+        reimbursement_claim_no: str,
+    ) -> bool:
+        reference_ids = {
+            str(flag.get(key) or "").strip()
+            for key in (
+                "generated_draft_claim_id",
+                "generatedDraftClaimId",
+                "reimbursement_claim_id",
+                "reimbursementClaimId",
+            )
+        }
+        reference_nos = {
+            str(flag.get(key) or "").strip().upper()
+            for key in (
+                "generated_draft_claim_no",
+                "generatedDraftClaimNo",
+                "reimbursement_claim_no",
+                "reimbursementClaimNo",
+            )
+        }
+        return (
+            bool(reimbursement_claim_id and reimbursement_claim_id in reference_ids)
+            or bool(reimbursement_claim_no and reimbursement_claim_no.upper() in reference_nos)
+        )
+
+    @classmethod
+    def _remove_deleted_reimbursement_link_references(
+        cls,
+        risk_flags: list[Any],
+        *,
+        reimbursement_claim_id: str,
+        reimbursement_claim_no: str,
+    ) -> tuple[list[Any], bool]:
+        next_flags: list[Any] = []
+        changed = False
+        stale_link_keys = {
+            "generated_draft_claim_id",
+            "generatedDraftClaimId",
+            "generated_draft_claim_no",
+            "generatedDraftClaimNo",
+            "reimbursement_claim_id",
+            "reimbursementClaimId",
+            "reimbursement_claim_no",
+            "reimbursementClaimNo",
+            "handoff_event_type",
+            "handoffEventType",
+            "handoff_message",
+            "handoffMessage",
+        }
+        for flag in list(risk_flags or []):
+            if not isinstance(flag, dict):
+                next_flags.append(flag)
+                continue
+            if not cls._reference_matches_deleted_reimbursement(
+                flag,
+                reimbursement_claim_id=reimbursement_claim_id,
+                reimbursement_claim_no=reimbursement_claim_no,
+            ):
+                next_flags.append(flag)
+                continue
+
+            next_flag = dict(flag)
+            for key in stale_link_keys:
+                if key in next_flag:
+                    next_flag.pop(key, None)
+                    changed = True
+            next_flags.append(next_flag)
+        return next_flags, changed
+
+    def _sync_linked_applications_after_reimbursement_deleted(
+        self,
+        *,
+        reimbursement_claim: ExpenseClaim,
+        operator: str,
+        current_user: Any,
+    ) -> list[dict[str, str]]:
+        reimbursement_claim_id = str(reimbursement_claim.id or "").strip()
+        reimbursement_claim_no = str(reimbursement_claim.claim_no or "").strip()
+        synced_applications: list[dict[str, str]] = []
+
+        for application_claim in self._find_linked_application_claims(reimbursement_claim):
+            previous_status = str(application_claim.status or "").strip()
+            previous_stage = str(application_claim.approval_stage or "").strip()
+            before_json = self._serialize_claim(application_claim)
+            next_flags, removed_link_references = self._remove_deleted_reimbursement_link_references(
+                list(application_claim.risk_flags_json or []),
+                reimbursement_claim_id=reimbursement_claim_id,
+                reimbursement_claim_no=reimbursement_claim_no,
+            )
+            sync_flag = with_risk_business_stage(
+                {
+                    "source": "application_link_sync",
+                    "event_type": "expense_application_reimbursement_deleted",
+                    "sync_event_id": str(uuid.uuid4()),
+                    "severity": "info",
+                    "actionability": "system_trace",
+                    "label": "关联报销单已删除",
+                    "message": (
+                        f"关联报销单 {reimbursement_claim_no or reimbursement_claim_id} 已删除，"
+                        "申请单已回到待关联状态。"
+                    ),
+                    "operator": operator,
+                    "operator_username": getattr(current_user, "username", ""),
+                    "operator_role_codes": [
+                        str(item).strip().lower()
+                        for item in getattr(current_user, "role_codes", [])
+                        if str(item).strip()
+                    ],
+                    "application_claim_id": application_claim.id,
+                    "application_claim_no": application_claim.claim_no,
+                    "deleted_reimbursement_claim_id": reimbursement_claim_id,
+                    "deleted_reimbursement_claim_no": reimbursement_claim_no,
+                    "previous_status": previous_status,
+                    "previous_approval_stage": previous_stage,
+                    "next_status": "approved",
+                    "next_approval_stage": APPLICATION_LINK_STATUS_STAGE,
+                    "removed_link_references": removed_link_references,
+                    "created_at": datetime.now(UTC).isoformat(),
+                },
+                "expense_application",
+            )
+
+            application_claim.status = "approved"
+            application_claim.approval_stage = APPLICATION_LINK_STATUS_STAGE
+            application_claim.risk_flags_json = [*next_flags, sync_flag]
+            synced_applications.append(
+                {
+                    "application_claim_id": application_claim.id,
+                    "application_claim_no": str(application_claim.claim_no or "").strip(),
+                    "next_approval_stage": APPLICATION_LINK_STATUS_STAGE,
+                }
+            )
+            self.audit_service.log_action(
+                actor=operator,
+                action="expense_application.unlink_deleted_reimbursement",
+                resource_type="expense_claim",
+                resource_id=application_claim.id,
+                before_json=before_json,
+                after_json=self._serialize_claim(application_claim),
+            )
+
+        return synced_applications

server/src/app/services/expense_claim_attachment_operations.py

@@ -714,6 +714,17 @@ class ExpenseClaimAttachmentOperationsMixin:
                 timeout_seconds=OcrService(self.db).settings.ocr_timeout_seconds,
             )
         except Exception:
+            metadata.update(
+                {
+                    "previewable": True,
+                    "preview_kind": "pdf",
+                    "preview_storage_key": self._attachment_storage.to_storage_key(file_path),
+                    "preview_media_type": "application/pdf",
+                    "preview_file_name": file_path.name,
+                    "preview_rendered_with": "",
+                }
+            )
+            self._attachment_storage.write_meta(file_path, metadata)
             return metadata
 
         metadata.update(

server/src/app/services/expense_claim_draft_flow.py

@@ -827,8 +827,8 @@ class ExpenseClaimDraftFlowMixin(ExpenseClaimApplicationLinkMixin, ExpenseClaimD
             document_count = max(len(context_documents), len(attachment_names), self._resolve_attachment_count(context_json))
             return {
                 "message": (
-                    f"检测到你已有草稿 {association_candidate.claim_no}，"
+                    f"检测到您已有草稿 {association_candidate.claim_no}，"
-                    f"当前新上传了 {document_count} 张票据，请先选择关联到现有草稿，或单独建立新的报销单。"
+                    f"当前新上传了 {document_count} 张票据，请先选择关联到现有草稿，或单独新建一张报销单。"
                 ),
                 "draft_only": False,
                 "status": "pending_association_decision",
@@ -859,7 +859,7 @@ class ExpenseClaimDraftFlowMixin(ExpenseClaimApplicationLinkMixin, ExpenseClaimD
             if existing_draft_count >= MAX_DRAFT_CLAIMS_PER_USER:
                 return {
                     "message": (
-                        f"你当前已保存 {MAX_DRAFT_CLAIMS_PER_USER} 个草稿，请先完成已保存的草稿，"
+                        f"您当前已保存 {MAX_DRAFT_CLAIMS_PER_USER} 个草稿，请先完成已保存的草稿，"
                         "才能再次新建草稿。"
                     ),
                     "draft_limit_reached": True,

server/src/app/services/expense_claims.py

@@ -688,6 +688,13 @@ class ExpenseClaimItemActionMixin:
 
         before_json = self._serialize_claim(claim)
         resource_id = claim.id
+        operator = self._access_policy.resolve_current_user_display_name(current_user)
+        if not self._is_expense_application_claim(claim):
+            self._sync_linked_applications_after_reimbursement_deleted(
+                reimbursement_claim=claim,
+                operator=operator,
+                current_user=current_user,
+            )
 
         self._release_budget_for_delete(claim, current_user)
         self._delete_claim_analysis_records(resource_id)
@@ -1008,4 +1015,3 @@ class ExpenseClaimService(ExpenseClaimStandardAdjustmentMixin, ExpenseClaimItemA
         )
 
         return claim
-

server/src/app/services/notification_states.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 from datetime import UTC, datetime
+from threading import Lock
 
 from sqlalchemy import select
 from sqlalchemy.orm import Session
@@ -16,11 +17,23 @@ from app.schemas.notification_state import (
 
 
 class NotificationStateService:
+    _storage_ready_bind_ids: set[int] = set()
+    _storage_ready_lock = Lock()
+
     def __init__(self, db: Session) -> None:
         self.db = db
 
     def ensure_storage_ready(self) -> None:
-        Base.metadata.create_all(bind=self.db.get_bind(), tables=[NotificationState.__table__])
+        bind = self.db.get_bind()
+        bind_id = id(bind)
+        if bind_id in self._storage_ready_bind_ids:
+            return
+
+        with self._storage_ready_lock:
+            if bind_id in self._storage_ready_bind_ids:
+                return
+            Base.metadata.create_all(bind=bind, tables=[NotificationState.__table__])
+            self._storage_ready_bind_ids.add(bind_id)
 
     def list_states(self, current_user: CurrentUserContext) -> NotificationStateListRead:
         self.ensure_storage_ready()

server/src/app/services/ocr.py

@@ -16,11 +16,13 @@ from sqlalchemy.orm import Session
 
 from app.core.config import SERVER_DIR, get_settings
 from app.schemas.ocr import OcrRecognizeBatchRead, OcrRecognizeDocumentRead, OcrRecognizeFieldRead, OcrRecognizeLineRead
+from app.services.document_preview import DocumentPreviewAssets
 from app.services.document_intelligence import DocumentIntelligenceService
 
 WORKER_JSON_PREFIX = "__OCR_JSON__="
 SUPPORTED_SUFFIXES = {".png", ".jpg", ".jpeg", ".webp", ".bmp", ".pdf"}
 OCR_RESULT_CACHE_LIMIT = 32
+OCR_RESULT_CACHE_PIPELINE_VERSION = f"pdf-image-ocr:{DocumentPreviewAssets.PDF_RENDERER_ID}:no-pdf-direct-v2"
 
 
 @dataclass(slots=True)
@@ -142,16 +144,6 @@ class OcrService:
                             cleanup_paths=cleanup_paths,
                             text_layer=text_layer,
                         )
-                        if self._has_usable_pdf_text_layer(text_layer):
-                            document = self._build_text_layer_document(
-                                filename=normalized_name,
-                                media_type=resolved_media_type,
-                                text_layer=text_layer,
-                                pdf_inputs=pdf_inputs,
-                            )
-                            documents.append(document)
-                            self._write_cached_document(cache_key, document)
-                            continue
                         prepared_inputs.extend(pdf_inputs)
                         for item in pdf_inputs:
                             cache_keys_by_source.setdefault(item.source_key, cache_key)
@@ -257,6 +249,7 @@ class OcrService:
         digest = hashlib.sha256(content).hexdigest()
         return "|".join(
             [
+                OCR_RESULT_CACHE_PIPELINE_VERSION,
                 self.settings.ocr_language,
                 self.settings.ocr_device,
                 self.settings.ocr_text_detection_model,
@@ -406,11 +399,15 @@ class OcrService:
         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)
+        image_paths, preview_usable = 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")
+        preview_data_url = (
+            self._build_preview_data_url(image_paths[0], media_type="image/png")
+            if preview_usable
+            else ""
+        )
         source_key = uuid4().hex
         descriptors: list[PreparedOcrInput] = []
         for page_index, image_path in enumerate(image_paths):
@@ -421,7 +418,7 @@ class OcrService:
                     filename=filename,
                     media_type=media_type,
                     page_index=page_index,
-                    preview_kind="image" if page_index == 0 else "",
+                    preview_kind="image" if page_index == 0 and preview_data_url else "",
                     preview_data_url=preview_data_url if page_index == 0 else "",
                     text_layer=text_layer if page_index == 0 else "",
                 )
@@ -450,27 +447,17 @@ class OcrService:
 
         return self._normalize_extracted_text(completed.stdout)
 
-    def _convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> list[Path]:
+    def _convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> tuple[list[Path], bool]:
-        prefix = output_dir / "page"
+        try:
-        completed = subprocess.run(
+            pages = DocumentPreviewAssets.render_pdf_pages(
-            [
+                pdf_path=pdf_path,
-                "pdftoppm",
+                output_dir=output_dir,
-                "-png",
+                timeout_seconds=self.settings.ocr_timeout_seconds,
-                "-r",
-                "160",
-                str(pdf_path),
-                str(prefix),
-            ],
-            capture_output=True,
-            text=True,
-            timeout=self.settings.ocr_timeout_seconds,
-            check=False,
             )
-        if completed.returncode != 0:
+        except RuntimeError as exc:
-            detail = (completed.stderr or completed.stdout or "").strip()
+            raise RuntimeError(f"PDF 转图片失败：{exc}") from exc
-            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)
+        return pages, True
 
     @staticmethod
     def _extract_pdf_page_sort_key(path: Path) -> tuple[int, str]:
@@ -595,30 +582,6 @@ class OcrService:
 
         return documents
 
-    def _build_text_layer_document(
-        self,
-        *,
-        filename: str,
-        media_type: str,
-        text_layer: str,
-        pdf_inputs: list[PreparedOcrInput],
-    ) -> OcrRecognizeDocumentRead:
-        first_input = pdf_inputs[0] if pdf_inputs else None
-        aggregated = AggregatedOcrDocument(
-            filename=filename,
-            media_type=media_type,
-            source_key=first_input.source_key if first_input is not None else uuid4().hex,
-            page_count=max(1, len(pdf_inputs)),
-            preview_kind=str(first_input.preview_kind if first_input is not None else ""),
-            preview_data_url=str(first_input.preview_data_url if first_input is not None else ""),
-        )
-        aggregated.text_layer_fragments.append(text_layer)
-        return self._finalize_document(aggregated)
-
-    @classmethod
-    def _has_usable_pdf_text_layer(cls, text_layer: str) -> bool:
-        return cls._meaningful_char_count(text_layer) >= 8
-
     @staticmethod
     def _collect_descriptor_text_layer(descriptors: list[PreparedOcrInput]) -> str:
         for descriptor in descriptors:
@@ -685,13 +648,6 @@ class OcrService:
             summary = self._summarize_text(full_text)
         preview_kind = aggregated.preview_kind
         preview_data_url = aggregated.preview_data_url
-        if (
-            used_text_layer
-            and aggregated.media_type == "application/pdf"
-            and self._placeholder_ratio(ocr_text) >= 0.12
-        ):
-            preview_kind = ""
-            preview_data_url = ""
         insight = self.document_intelligence_service.build_document_insight(
             filename=aggregated.filename,
             summary=summary,

server/src/app/services/ontology_validation.py

@@ -214,7 +214,7 @@ class OntologyValidationMixin:
         labels = [self._display_slot_label(item) for item in missing_slots[:4]]
         if not labels:
             return "请补充更多上下文后再继续。"
-        return f"请补充{'、'.join(labels)}，我再继续帮你解析和处理。"
+        return f"请补充{'、'.join(labels)}，我会继续帮您解析和处理。"
 
     @staticmethod
     def _compute_confidence(

server/src/app/services/orchestrator_expense_query.py

@@ -570,7 +570,7 @@ class OrchestratorDatabaseQueryBuilder:
                 scoped_to_current_user = True
             else:
                 conditions.append(ExpenseClaim.id == "__no_visible_claim__")
-                scope_label = "你的报销单"
+                scope_label = "您的报销单"
                 scoped_to_current_user = True
         elif explicit_employee_names:
             conditions.append(ExpenseClaim.employee_name.in_(explicit_employee_names))
@@ -586,7 +586,7 @@ class OrchestratorDatabaseQueryBuilder:
                 scoped_to_current_user = True
             else:
                 conditions.append(ExpenseClaim.id == "__no_visible_claim__")
-                scope_label = "你的报销单"
+                scope_label = "您的报销单"
                 scoped_to_current_user = True
         else:
             scope_label = "全部报销单"
@@ -703,7 +703,7 @@ class OrchestratorDatabaseQueryBuilder:
 
         subject_name = (employee.name if employee is not None else "") or normalized_user_id
         if subject_name:
-            return conditions, "你的报销单"
+            return conditions, "您的报销单"
         return conditions, "当前用户的报销单"
 
     def _employee_name_is_unique(self, employee: Employee) -> bool:

server/src/app/services/receipt_folder.py

@@ -37,10 +37,18 @@ TRAIN_ROUTE_PATTERN = re.compile(
     r"([\u4e00-\u9fa5]{2,12})站?\s*(?:至|到|→|->|—|–|-)\s*"
     r"([\u4e00-\u9fa5]{2,12})站?"
 )
+TRAIN_ROUTE_WITH_NO_PATTERN = re.compile(
+    r"([\u4e00-\u9fa5]{2,12})站?\s+[GCDZKTLYS]\d{1,5}\s+"
+    r"([\u4e00-\u9fa5]{2,12})站?",
+    re.IGNORECASE,
+)
 TRAIN_NO_PATTERN = re.compile(r"(?:车次|列车号)\s*[:：]?\s*([GCDZKTLYS]\d{1,5})", re.IGNORECASE)
 TRAIN_STANDALONE_NO_PATTERN = re.compile(r"(?<![A-Z0-9])([GCDZKTLYS]\d{1,5})(?![A-Z0-9])", re.IGNORECASE)
-TRAIN_PASSENGER_PATTERN = re.compile(r"(?:乘车人|旅客姓名|姓名)\s*[:：]?\s*([\u4e00-\u9fa5·]{2,20})")
+TRAIN_PASSENGER_PATTERN = re.compile(r"(?:乘车人|乘客|旅客姓名|姓名)\s*[:：]?\s*([\u4e00-\u9fa5·]{2,20})")
-TRAIN_ID_PATTERN = re.compile(r"(?:有效身份证件号码|身份证件号码|证件号码|身份证号)\s*[:：]?\s*([0-9Xx*]{6,24})")
+TRAIN_PURCHASER_NAME_PATTERN = re.compile(
+    r"购买方名称\s*[:：]?\s*([·\u4e00-\u9fa5]{2,20}?)(?=\s*(?:统一社会信用代码|纳税人识别号|$))"
+)
+TRAIN_ID_PATTERN = re.compile(r"(?:有效身份证件号码|身份证件号码|证件号码|身份证号|证件号)\s*[:：]?\s*([0-9Xx*]{6,24})")
 TRAIN_ID_FALLBACK_PATTERN = re.compile(r"(?<![0-9A-Za-z])([0-9]{6,17}[0-9Xx*]{2,8})(?![0-9A-Za-z])")
 TRAIN_ETICKET_PATTERN = re.compile(r"(?:电子客票号|客票号)\s*[:：]?\s*([A-Z0-9]{6,32})", re.IGNORECASE)
 TRAIN_SEAT_CLASS_PATTERN = re.compile(r"(商务座|特等座|一等座|二等座|一等卧|二等卧|软卧|硬卧|软座|硬座|无座)")
@@ -50,6 +58,28 @@ TRAIN_COMBINED_SEAT_PATTERN = re.compile(r"([0-9]{1,2})车\s*([0-9]{1,3}[A-F])
 TRAIN_LOOSE_SEAT_PATTERN = re.compile(r"(?<!\d)([0-9]{1,2})\s+([0-9]{1,3}[A-F])(?![A-Za-z0-9])", re.IGNORECASE)
 TRAIN_FARE_PATTERN = re.compile(r"(?:票价|金额)\s*[:：￥¥\s]*([0-9]+(?:[.,][0-9]{1,2})?)")
 TRAIN_LOOSE_FARE_PATTERN = re.compile(r"(?<!\d)([0-9]{1,6}\.\d{1,2})(?!\d)")
+TRAIN_STATION_FIELD_KEYS = {"departure_station", "arrival_station"}
+TRAIN_STATION_FIELD_LABELS = {"出发地点", "到达地点"}
+TRAIN_INVALID_STATION_TOKENS = (
+    "座",
+    "席",
+    "扫码",
+    "无效",
+    "票价",
+    "金额",
+    "车厢",
+    "座位",
+    "乘客",
+    "证件",
+    "身份证",
+    "订单",
+    "单据",
+    "日期",
+    "渠道",
+    "官方",
+    "演示",
+    "不可报销",
+)
 
 
 class ReceiptFolderStorageMixin:
@@ -119,6 +149,29 @@ class ReceiptFolderStorageMixin:
                 "preview_media_type": preview_media_type,
                 "preview_rendered_with": DocumentPreviewAssets.renderer_id_for_source(media_type),
             }
+        if str(media_type or "").strip() == "application/pdf":
+            preview_path = receipt_dir / f"preview{DocumentPreviewAssets.PDF_PREVIEW_SUFFIX}"
+            try:
+                DocumentPreviewAssets.render_pdf_first_page(
+                    pdf_path=source_path,
+                    preview_path=preview_path,
+                    timeout_seconds=get_settings().ocr_timeout_seconds,
+                )
+            except Exception:
+                return {
+                    "previewable": True,
+                    "preview_kind": "pdf",
+                    "preview_file_name": source_path.name,
+                    "preview_media_type": media_type,
+                    "preview_rendered_with": "",
+                }
+            return {
+                "previewable": True,
+                "preview_kind": "image",
+                "preview_file_name": preview_path.name,
+                "preview_media_type": DocumentPreviewAssets.PDF_PREVIEW_MEDIA_TYPE,
+                "preview_rendered_with": DocumentPreviewAssets.PDF_RENDERER_ID,
+            }
         if self._is_previewable(media_type):
             return {
                 "previewable": True,
@@ -172,6 +225,16 @@ class ReceiptFolderStorageMixin:
                 timeout_seconds=get_settings().ocr_timeout_seconds,
             )
         except Exception:
+            meta.update(
+                {
+                    "previewable": True,
+                    "preview_kind": "pdf",
+                    "preview_file_name": source_path.name,
+                    "preview_media_type": "application/pdf",
+                    "preview_rendered_with": "",
+                }
+            )
+            self._write_meta(receipt_dir, meta)
             return meta
 
         meta.update(
@@ -543,6 +606,10 @@ class ReceiptFolderTrainTicketMixin:
             value = str(field.get("value") or "").strip()
             if not label or not value:
                 continue
+            if key == "merchant_name" or label == "商户":
+                continue
+            if not cls._should_keep_train_ticket_field(key=key, label=label, value=value):
+                continue
             if key == "trip_no" and label == "车次/航班":
                 label = "车次"
             if key == "route" and label == "行程":
@@ -559,6 +626,8 @@ class ReceiptFolderTrainTicketMixin:
                 return
             normalized.append({"key": key, "label": label, "value": cleaned})
 
+        add_field("merchant_name", "商户", "中国铁路")
+
         invoice_date = cls._extract_train_invoice_date(text)
         add_field("invoice_date", "开票日期", invoice_date)
 
@@ -690,6 +759,13 @@ class ReceiptFolderTrainTicketMixin:
     @classmethod
     def _extract_train_route_points(cls, text: str) -> tuple[str, str]:
         raw_text = str(text or "")
+        split_line_match = TRAIN_ROUTE_WITH_NO_PATTERN.search(raw_text)
+        if split_line_match:
+            departure = cls._clean_train_station(split_line_match.group(1))
+            arrival = cls._clean_train_station(split_line_match.group(2))
+            if cls._is_valid_train_station_value(departure) and cls._is_valid_train_station_value(arrival) and departure != arrival:
+                return departure, arrival
+
         station_candidates: list[str] = []
         for line in raw_text.replace("\r", "\n").splitlines():
             candidate = cls._clean_train_station(line)
@@ -697,7 +773,7 @@ class ReceiptFolderTrainTicketMixin:
                 continue
             if not str(line or "").strip().endswith("站"):
                 continue
-            if any(token in candidate for token in ("发票", "客票", "铁路", "票价", "日期")):
+            if not cls._is_valid_train_station_value(candidate):
                 continue
             station_candidates.append(candidate)
             if len(station_candidates) >= 2:
@@ -707,7 +783,7 @@ class ReceiptFolderTrainTicketMixin:
         if match:
             departure = cls._clean_train_station(match.group(1))
             arrival = cls._clean_train_station(match.group(2))
-            if departure and arrival and departure != arrival:
+            if cls._is_valid_train_station_value(departure) and cls._is_valid_train_station_value(arrival) and departure != arrival:
                 return departure, arrival
         return "", ""
 
@@ -717,6 +793,25 @@ class ReceiptFolderTrainTicketMixin:
         cleaned = re.sub(r"(?:火车站|高铁站|站)$", "", cleaned)
         return cleaned.strip()
 
+    @classmethod
+    def _should_keep_train_ticket_field(cls, *, key: str, label: str, value: str) -> bool:
+        if key in TRAIN_STATION_FIELD_KEYS or label in TRAIN_STATION_FIELD_LABELS:
+            return cls._is_valid_train_station_value(value)
+        if key == "passenger_name" or label == "乘车人":
+            return bool(cls._clean_train_passenger_candidate(value))
+        return True
+
+    @classmethod
+    def _is_valid_train_station_value(cls, value: str) -> bool:
+        cleaned = cls._clean_train_station(value)
+        if not 2 <= len(cleaned) <= 12:
+            return False
+        if any(token in cleaned for token in TRAIN_INVALID_STATION_TOKENS):
+            return False
+        if re.search(r"[A-Za-z0-9]", cleaned):
+            return False
+        return True
+
     @staticmethod
     def _extract_first(pattern: re.Pattern[str], text: str) -> str:
         match = pattern.search(str(text or ""))
@@ -724,14 +819,19 @@ class ReceiptFolderTrainTicketMixin:
 
     @classmethod
     def _extract_train_passenger_name(cls, text: str, *, id_number: str = "") -> str:
-        labeled = cls._extract_first(TRAIN_PASSENGER_PATTERN, text)
+        lines = [line.strip() for line in str(text or "").replace("\r", "\n").splitlines() if line.strip()]
+        for line in lines:
+            labeled = cls._clean_train_passenger_candidate(cls._extract_first(TRAIN_PASSENGER_PATTERN, line))
             if labeled:
                 return labeled
 
-        lines = [line.strip() for line in str(text or "").replace("\r", "\n").splitlines() if line.strip()]
+        if id_number:
             for index, line in enumerate(lines):
-            if id_number and id_number not in line:
+                if id_number not in line:
                     continue
+                candidate = cls._clean_train_passenger_candidate(line.replace(id_number, " "))
+                if candidate:
+                    return candidate
                 for offset in (1, -1, 2):
                     target_index = index + offset
                     if target_index < 0 or target_index >= len(lines):
@@ -740,8 +840,9 @@ class ReceiptFolderTrainTicketMixin:
                     if candidate:
                         return candidate
         for line in lines:
-            if "购买方名称" in line:
+            purchase_match = TRAIN_PURCHASER_NAME_PATTERN.search(line)
-                candidate = cls._clean_train_passenger_candidate(line.split(":", 1)[-1].split("：", 1)[-1])
+            if purchase_match:
+                candidate = cls._clean_train_passenger_candidate(purchase_match.group(1))
                 if candidate:
                     return candidate
         return ""
@@ -764,6 +865,16 @@ class ReceiptFolderTrainTicketMixin:
                 "开票",
                 "日期",
                 "车厢",
+                "席别",
+                "二等座",
+                "一等座",
+                "商务座",
+                "特等座",
+                "软座",
+                "硬座",
+                "无座",
+                "软卧",
+                "硬卧",
                 "座位",
                 "票价",
                 "金额",
@@ -771,6 +882,14 @@ class ReceiptFolderTrainTicketMixin:
                 "出发",
                 "到达",
                 "车次",
+                "公司",
+                "信用代码",
+                "纳税人",
+                "扫码",
+                "无效",
+                "二维码",
+                "座席",
+                "证件",
             )
         ):
             return ""

server/src/app/services/steward_model_plan_builder.py

@@ -266,7 +266,7 @@ class StewardModelPlanBuilder:
                 event_id="intent_agent_function_call",
                 stage="llm_function_call",
                 title="识别财务事项",
-                content="我识别到这句话包含出差事项，但还需要确认你要进入申请流程还是报销流程。",
+                content="我识别到这句话包含出差事项，但还需要确认您是要进入申请流程还是报销流程。",
             )
         ]
         raw_events = payload.get("thinking_events")
@@ -292,7 +292,7 @@ class StewardModelPlanBuilder:
                     event_id="intent_agent_pending_flow",
                     stage="flow_confirmation",
                     title="等待确认流程方向",
-                    content=f"当前输入“{request.message}”缺少明确动作词，需要先由你选择补办出差申请或发起费用报销。",
+                    content=f"当前输入“{request.message}”缺少明确的动作词，需要先由您选择是补办出差申请，还是发起费用报销。",
                 )
             )
         return events
@@ -302,7 +302,7 @@ class StewardModelPlanBuilder:
         candidate_labels = [item.label for item in pending_flow_confirmation.candidate_flows if item.label]
         if len(candidate_labels) >= 2:
             return (
-                f"我识别到这是一次财务事项，但还不能确定你要做的是"
+                f"我识别到这是一次财务事项，但还不能确定您要做的是"
                 f"**{candidate_labels[0]}**还是**{candidate_labels[1]}**。请先选择一个方向。"
             )
         return "我识别到这是一次财务事项，但还需要先确认具体流程方向。"

server/src/app/services/steward_planner_fallback.py

@@ -335,7 +335,7 @@ class StewardPlannerFallbackMixin:
                     flow_id="travel_application",
                     label="先发起出差申请",
                     confidence=0.86,
-                    reason="已先查询你名下可关联的差旅申请单，暂未查到可关联单据，因此应先申请单据。",
+                    reason="已先查询您名下可关联的差旅申请单，暂未查到可关联单据，因此应先申请单据。",
                     ontology_fields=application_fields,
                     missing_fields=self._resolve_missing_fields("expense_application", application_fields),
                 )
@@ -345,7 +345,7 @@ class StewardPlannerFallbackMixin:
         if gate.get("checked"):
             candidate_count = int(gate.get("candidate_count") or 0)
             reimbursement_label = "关联已有申请单并发起报销"
-            reimbursement_reason = f"已先查到 {candidate_count} 个可关联申请单，选择后会先请你关联具体单据。"
+            reimbursement_reason = f"已先查到 {candidate_count} 个可关联申请单，选择后会先请您关联具体单据。"
         return [
             StewardCandidateFlow(
                 flow_id="travel_application",
@@ -390,10 +390,10 @@ class StewardPlannerFallbackMixin:
     @staticmethod
     def _build_pending_flow_reason(gate: dict[str, Any]) -> str:
         if gate.get("checked") and int(gate.get("candidate_count") or 0) <= 0:
-            return "我已经先查询你名下可关联的差旅申请单，未查到可关联单据，所以当前应先申请单据。"
+            return "我已先查询您名下可关联的差旅申请单，未查到可关联单据，所以当前应先申请单据。"
         if gate.get("checked"):
             candidate_count = int(gate.get("candidate_count") or 0)
-            return f"我已经先查询你名下的差旅申请单，查到 {candidate_count} 个可关联申请单，需要你确认是否关联单据后发起报销。"
+            return f"我已先查询您名下的差旅申请单，查到 {candidate_count} 个可关联申请单，需要您确认是否关联单据后发起报销。"
         return "当前话术描述了出差事项，但没有明确说明要补办申请还是发起报销。"
 
     @staticmethod
@@ -404,10 +404,10 @@ class StewardPlannerFallbackMixin:
             candidate_count = int(gate.get("candidate_count") or 0)
             return (
                 f"我已先查询可关联申请单，查到 {candidate_count} 个可关联申请单；"
-                "你可以选择关联已有申请单发起报销，或改为补办新的出差申请。"
+                "您可以选择关联已有申请单发起报销，也可以改为补办新的出差申请。"
             )
         return (
-            "我识别到这是一次出差事项，但还不能确定你要做的是"
+            "我识别到这是一次出差事项，但还不能确定您要做的是"
             "**补办出差申请**还是**发起费用报销**。请先选择一个方向。"
         )
 

server/src/app/services/steward_runtime_decision_agent.py

@@ -90,7 +90,7 @@ class StewardRuntimeDecisionAgent:
             next_action="continue_selected_flow",
             target_task_id=selected_flow_id,
             response_text=self._build_selected_flow_response_text(selected_flow_id),
-            rationale="已按你选择的候选流程继续处理。",
+            rationale="已按您选择的候选流程继续处理。",
             steward_state=next_state,
             model_call_traces=traces,
         )
@@ -268,7 +268,7 @@ class StewardRuntimeDecisionAgent:
                 next_action="submit_current_application",
                 target_message_id=str(pending_application.get("message_id") or ""),
                 target_task_id=str(pending_application.get("task_id") or ""),
-                rationale="模型运行时决策暂不可用，我先按当前待提交申请单上下文处理你的确认。",
+                rationale="模型运行时决策暂不可用，我先按当前待提交申请单的上下文处理您的确认。",
                 model_call_traces=traces,
             )
         if confirmation_text and pending_steward_action:
@@ -295,7 +295,7 @@ class StewardRuntimeDecisionAgent:
                     target_task_id=str(current_task.get("task_id") or ""),
                     field_key=field_key,
                     field_value=request.user_message,
-                    rationale="模型运行时决策暂不可用，我先把你的补充写入当前小财管家流程字段。",
+                        rationale="模型运行时决策暂不可用，我先把您的补充写入当前小财管家流程字段。",
                     model_call_traces=traces,
                 )
             if field_key:

server/src/app/services/steward_slot_decision_agent.py

@@ -275,7 +275,7 @@ class StewardSlotDecisionAgent:
                 missing_fields=missing_fields,
                 question=self._build_fallback_question(field),
                 options=self._sanitize_options([], [field]),
-                rationale="模型字段决策暂不可用，我先按上游意图识别给出的本体缺口向你确认。",
+                rationale="模型字段决策暂不可用，我先按上游意图识别给出的本体缺口向您确认。",
                 model_call_traces=traces,
             )
         return StewardSlotDecisionResponse(
@@ -285,7 +285,7 @@ class StewardSlotDecisionAgent:
             missing_fields=[],
             question="",
             options=[],
-            rationale="当前任务没有上游标记的关键字段缺口，可以先生成核对结果供你确认。",
+            rationale="当前任务没有上游标记的关键字段缺口，可以先生成核对结果供您确认。",
             model_call_traces=traces,
         )
 
@@ -293,7 +293,7 @@ class StewardSlotDecisionAgent:
     def _build_fallback_question(field: str) -> str:
         label = FIELD_CATALOG.get(field, {}).get("label") or field
         if field == "transport_mode":
-            return "请问你这次打算怎么出行？可以选择火车、飞机或轮船。"
+            return "请问您这次打算怎么出行？可以选择火车、飞机或轮船。"
         return f"当前还缺少{label}，请先补充后我再继续处理。"
 
     @staticmethod

server/src/app/services/travel_reimbursement_calculator.py

@@ -15,14 +15,16 @@ from app.schemas.reimbursement import (
     TravelReimbursementCalculatorResponse,
 )
 from app.services.agent_assets import AgentAssetService
+from app.services.application_location_semantics import validate_application_location_text
 from app.services.expense_claims import ExpenseClaimService
-from app.services.expense_rule_runtime import RuntimeTravelPolicy, ExpenseRuleRuntimeService
+from app.services.expense_rule_runtime import ExpenseRuleRuntimeService, RuntimeTravelPolicy
 from app.services.travel_policy_grades import travel_policy_grade_key_candidates
 from app.services.travel_reimbursement_regions import (
     AMBIGUOUS_PROVINCE_CITY_NAMES,
     OTHER_REGION_LOCATION_KEYWORDS,
     OTHER_REGION_PROVINCE_KEYWORDS,
 )
+from app.services.user_agent_application_locations import normalize_application_location
 
 
 class TravelReimbursementCalculatorService:
@@ -35,9 +37,13 @@ class TravelReimbursementCalculatorService:
         current_user: CurrentUserContext,
     ) -> TravelReimbursementCalculatorResponse:
         days = max(1, int(payload.days))
-        location = str(payload.location or "").strip()
+        raw_location = str(payload.location or "").strip()
-        if not location:
+        if not raw_location:
             raise ValueError("请先填写出差地点。")
+        location = normalize_application_location(raw_location) or raw_location
+        location_error = validate_application_location_text(location)
+        if location_error:
+            raise ValueError(f"{location_error}请填写真实出差地点后再计算。")
 
         policy = self._load_travel_policy()
         grade = self._resolve_grade(payload.grade, current_user)

server/src/app/services/user_agent_application.py

@@ -8,20 +8,25 @@ from sqlalchemy import or_, select
 
 from app.api.deps import CurrentUserContext
 from app.models.financial_record import ExpenseClaim
+from app.schemas.reimbursement import TravelReimbursementCalculatorRequest
 from app.schemas.user_agent import (
     UserAgentDraftPayload,
     UserAgentRequest,
     UserAgentResponse,
     UserAgentSuggestedAction,
 )
-from app.schemas.reimbursement import TravelReimbursementCalculatorRequest
+from app.services.application_location_semantics import (
-from app.services.expense_claim_access_policy import ExpenseClaimAccessPolicy
+    strip_route_location_prefix_with_jieba,
-from app.services.expense_claim_risk_stage import with_risk_business_stage
+    validate_application_location_text,
-from app.services.travel_reimbursement_calculator import TravelReimbursementCalculatorService
+)
+from app.services.application_system_estimate import apply_application_system_estimate_to_facts
 from app.services.document_numbering import (
     build_document_number,
     generate_unique_expense_claim_no,
 )
+from app.services.expense_claim_access_policy import ExpenseClaimAccessPolicy
+from app.services.expense_claim_risk_stage import with_risk_business_stage
+from app.services.travel_reimbursement_calculator import TravelReimbursementCalculatorService
 from app.services.user_agent_application_dates import (
     expand_application_time_with_days,
     resolve_application_date_range,
@@ -33,7 +38,6 @@ from app.services.user_agent_application_summary import (
     build_application_summary_table,
     resolve_application_time_label,
 )
-from app.services.application_system_estimate import apply_application_system_estimate_to_facts
 
 APPLICATION_CONTEXT_VALUES = {
     "application",
@@ -182,6 +186,17 @@ class UserAgentApplicationSlotMixin:
             if not str(facts.get(field) or "").strip()
         ]
 
+    @staticmethod
+    def _resolve_application_validation_issues(facts: dict[str, str]) -> list[dict[str, str]]:
+        issues: list[dict[str, str]] = []
+        location_error = validate_application_location_text(facts.get("location", ""))
+        if location_error:
+            issues.append({
+                "field": "location",
+                "message": location_error,
+            })
+        return issues
+
     def _resolve_application_missing_fields(self, facts: dict[str, str]) -> list[str]:
         return [
             *self._resolve_application_missing_base_fields(facts),
@@ -391,6 +406,10 @@ class UserAgentApplicationSlotMixin:
         if re.fullmatch(r"(?:去|到|前往)?[\u4e00-\u9fa5]{1,8}出差(?P<days>\d+|[一二两三四五六七八九十]{1,3})?天?", text):
             return ""
 
+        tokenized = strip_route_location_prefix_with_jieba(text)
+        if tokenized != text:
+            text = tokenized
+        else:
             text = re.sub(r"^.*?(?:出差|前往|去|到|赴)[\u4e00-\u9fa5]{1,8}(?:出差)?(?P<days>\d+|[一二两三四五六七八九十]{1,3})?天?[，,\s]*", "", text)
         text = re.sub(r"^(?:出差|申请|费用申请|业务|本次|去|到|前往)\s*", "", text)
         text = text.strip(" ：:，,。；;")
@@ -537,8 +556,16 @@ class UserAgentApplicationSlotMixin:
         step: str,
         facts: dict[str, str],
     ) -> list[UserAgentSuggestedAction]:
-        if step == "ask_missing":
+        if step in {"ask_missing", "ask_invalid"}:
-            missing_fields = self._resolve_application_missing_fields(facts)
+            missing_fields = (
+                self._resolve_application_missing_fields(facts)
+                if step == "ask_missing"
+                else [
+                    issue.get("field", "")
+                    for issue in self._resolve_application_validation_issues(facts)
+                    if issue.get("field")
+                ]
+            )
             return [
                 UserAgentSuggestedAction(
                     label="一次性补充申请信息",
@@ -1209,7 +1236,22 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
                     "我已按「费用申请 / 事前审批」来处理这条内容。",
                     "已识别信息：\n" + recognized_table,
                     f"当前还需要补充：{missing_text}。",
-                    "请一次性补齐上述字段，我会继续生成申请核对结果并让你确认是否提交。",
+                    "请一次性补齐上述字段，我会继续生成申请核对结果，并请您确认是否提交。",
+                ]
+            )
+
+        if step == "ask_invalid":
+            issue_messages = [
+                item["message"]
+                for item in self._resolve_application_validation_issues(facts)
+                if str(item.get("message") or "").strip()
+            ]
+            return "\n\n".join(
+                [
+                    "我已识别到申请信息里有需要先修正的字段。",
+                    "已识别信息：\n" + recognized_table,
+                    *issue_messages,
+                    "请把地点改为真实出差地点，业务事项放在事由中；修正后我再帮您提交申请。",
                 ]
             )
 
@@ -1473,7 +1515,7 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
                 pick("applicationType", "application_type")
             ),
             "time": pick("time", "timeRange", "time_range"),
-            "location": pick("location"),
+            "location": normalize_application_location(pick("location")),
             "reason": reason,
             "days": pick("days"),
             "transport_mode": pick("transportMode", "transport_mode"),
@@ -1507,6 +1549,8 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
         payload: UserAgentRequest,
         facts: dict[str, str],
     ) -> str:
+        if self._resolve_application_validation_issues(facts):
+            return "ask_invalid"
         if self._is_application_save_draft_action(payload):
             return "draft"
         if self._resolve_application_missing_base_fields(facts):
@@ -1516,4 +1560,3 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
         if self._is_application_submit_confirmation(payload):
             return "submitted"
         return "preview"
-

server/src/app/services/user_agent_application_locations.py

@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import re
 
-
 DIRECT_MUNICIPALITY_DISPLAY = {
     "北京": "北京市",
     "北京市": "北京市",
@@ -79,7 +78,7 @@ CITY_TO_PROVINCE = {
 }
 
 LOCATION_NOISE_PATTERN = re.compile(
-    r"(?:出差|驻场|现场|支撑|支持|部署|上线|实施|拜访|验收|会议|采购|培训|协助|处理|办理|参加|进行).*$"
+    r"(?:出差|驻场|现场|支撑|支持|辅助|部署|上线|实施|拜访|验收|会议|采购|培训|协助|处理|办理|参加|进行).*$"
 )
 
 

server/src/app/services/user_agent_knowledge.py

@@ -716,7 +716,7 @@ class UserAgentKnowledgeMixin(UserAgentKnowledgeHelpersMixin):
             self._append_markdown_section(
                 answer_lines,
                 "说明",
-                ["- 请补充费用类型、适用地区、职级或具体业务场景，我再继续帮你缩小范围。"],
+                ["- 请补充费用类型、适用地区、职级或具体业务场景，我会继续帮您缩小范围。"],
             )
             return "\n".join(answer_lines).strip()
 
@@ -729,7 +729,7 @@ class UserAgentKnowledgeMixin(UserAgentKnowledgeHelpersMixin):
         self._append_markdown_section(
             answer_lines,
             "说明",
-            ["- 以上只使用当前命中的知识库证据；没有在证据中出现的适用条件或金额，我不会替你默认补齐。"],
+            ["- 以上只使用当前命中的知识库证据；没有在证据中出现的适用条件或金额，我不会替您默认补齐。"],
         )
         return "\n".join(answer_lines).strip()
 

server/src/app/services/user_agent_response.py

@@ -61,7 +61,7 @@ class UserAgentResponseMixin:
         if payload.ontology.intent == "draft":
             tool_message = str(payload.tool_payload.get("message") or "").strip()
             if payload.tool_payload.get("draft_limit_reached"):
-                return tool_message or "你当前已保存 3 个草稿，请先完成已保存的草稿，才能再次新建草稿。"
+                return tool_message or "您当前已保存 3 个草稿，请先完成已保存的草稿，才能再次新建草稿。"
             if tool_message and (
                 str(payload.tool_payload.get("claim_id") or "").strip()
                 or str(payload.tool_payload.get("claim_no") or "").strip()
@@ -88,12 +88,12 @@ class UserAgentResponseMixin:
         elif attachment_names:
             attachment_hint = (
                 f" 我已带入 {len(attachment_names)} 份附件名称，但目前还不能直接读取附件内容，"
-                "仍需要你补充关键信息。"
+                "还需要您补充关键信息。"
             )
 
         return (
-            "可以帮你发起报销。请补充费用类型、发生时间、金额、事由和相关对象，"
+            "可以帮您发起报销。请补充费用类型、发生时间、金额、事由和相关对象，"
-            "或者直接上传票据附件，我再继续帮你判断能否报、缺什么材料，并整理待核对信息。"
+            "或者直接上传票据附件，我会继续帮您判断能否报销、还缺哪些材料，并整理出待核对信息。"
             f"{attachment_hint}"
         )
 
@@ -122,8 +122,8 @@ class UserAgentResponseMixin:
 
         return (
             f"已识别到一笔{time_text}的{expense_type}支出{amount_hint}。"
-            "如果要继续整理报销核对信息，还需要补充客户单位、参与人员、费用明细和票据附件。"
+            "如果需要继续整理报销核对信息，还需要补充客户单位、参与人员、费用明细和票据附件。"
-            "你也可以继续上传发票或图片，我会把这些信息带入后续对话。"
+            "您也可以继续上传发票或图片，我会把这些信息带入后续对话。"
         )
 
 
@@ -347,7 +347,7 @@ class UserAgentResponseMixin:
             query_payload = self._build_query_payload(payload)
             scope_label = str(data.get("scope_label") or subject).strip() or subject
             if query_payload is None:
-                return f"当前没有查到{scope_label}。你可以补充时间范围、单号或状态继续筛选。"
+                return f"当前没有查到{scope_label}。您可以补充时间范围、单号或状态继续筛选。"
 
             window_prefix = (
                 f"{query_payload.window_start_date} 至 {query_payload.window_end_date}"
@@ -367,10 +367,10 @@ class UserAgentResponseMixin:
                         f"另有 {query_payload.older_record_count} 笔超过 {query_payload.window_days} 日的单据，"
                         "请前往个人报销中心查看。"
                     )
-                return f"{window_prefix}没有查到{query_payload.scope_label}。你可以补充时间范围、单号或状态继续筛选。"
+                return f"{window_prefix}没有查到{query_payload.scope_label}。您可以补充时间范围、单号或状态继续筛选。"
 
             answer_parts = [
-                f"已按你的筛选条件查询{query_payload.scope_label}。",
+                f"已按您的筛选条件查询{query_payload.scope_label}。",
                 f"下面先列出最近 {query_payload.preview_count} 条记录，点击任一单据即可查看详情。",
                 f"本次共命中 {query_payload.record_count} 笔，金额合计 {query_payload.total_amount:.2f} 元。",
             ]

server/src/app/services/user_agent_review_core.py

@@ -68,8 +68,8 @@ class UserAgentReviewCoreMixin:
         if has_time:
             context_hint += "，并看到了业务发生时间"
         return (
-            f"{context_hint}。但你还没有明确这笔单据属于哪类报销。"
+            f"{context_hint}。但您还没有明确这笔单据属于哪类报销。"
-            "请先在下面选择报销场景，我会按你选择的场景再继续识别时间、地点、事由、金额和所需票据，"
+            "请先在下面选择报销场景，我会按您选择的场景继续识别时间、地点、事由、金额和所需票据，"
             "避免系统先入为主把项目支持、部署等描述误判成差旅。"
         )
 

server/src/app/services/user_agent_review_messages.py

@@ -164,7 +164,7 @@ class UserAgentReviewMessageMixin:
         if payload.tool_payload.get("draft_limit_reached"):
             return (
                 str(payload.tool_payload.get("message") or "").strip()
-                or "你当前已保存 3 个草稿，请先完成已保存的草稿，才能再次新建草稿。"
+                or "您当前已保存 3 个草稿，请先完成已保存的草稿，才能再次新建草稿。"
             )
 
         review_action = str(payload.context_json.get("review_action") or "").strip()
@@ -254,11 +254,11 @@ class UserAgentReviewMessageMixin:
             if claim_no:
                 return (
                     f"已识别出本次上传的 {document_count} 张票据。"
-                    f"系统检测到你已有草稿 {claim_no}，请选择关联到该草稿，或单独建立一张新的报销单。"
+                    f"系统检测到您已有草稿 {claim_no}，请选择关联到该草稿，或单独新建一张报销单。"
                 )
             return (
                 f"已识别出本次上传的 {document_count} 张票据。"
-                "系统检测到你已有可用草稿，请先选择关联到现有草稿，或单独建立一张新的报销单。"
+                "系统检测到您已有可用草稿，请先选择关联到现有草稿，或单独新建一张报销单。"
             )
 
         blocked_reasons = self._resolve_submission_blocked_reasons(payload)

server/src/app/services/user_agent_review_slots.py

@@ -668,7 +668,7 @@ class UserAgentReviewSlotMixin:
             status="missing" if is_missing else "identified" if source in {"user_text", "user_form"} else "inferred",
             hint=f"建议补充 {SLOT_LABELS.get(key, key)}。"
             if is_missing and required
-            else ("该字段来自系统辅助上下文，建议你再核对一次。" if source in {"detail_context", "ocr"} else ""),
+            else ("该字段来自系统辅助上下文，建议您再核对一次。" if source in {"detail_context", "ocr"} else ""),
             evidence=evidence,
         )
 

server/tests/test_agent_asset_service.py

@@ -888,6 +888,34 @@ def test_travel_reimbursement_calculator_rejects_unrecognized_location() -> None
             )
 
 
+def test_travel_reimbursement_calculator_normalizes_location_mixed_with_business_content() -> None:
+    with build_session() as db:
+        db.add(
+            Employee(
+                employee_no="E9004",
+                name="混合地点员工",
+                email="mixed-location@example.com",
+                position="产品经理",
+                grade="P4",
+            )
+        )
+        db.commit()
+
+        result = TravelReimbursementCalculatorService(db).calculate(
+            TravelReimbursementCalculatorRequest(days=4, location="上海辅助国网仿生产服务器"),
+            CurrentUserContext(
+                username="mixed-location@example.com",
+                name="混合地点员工",
+                role_codes=[],
+                is_admin=False,
+            ),
+        )
+
+        assert result.location == "上海市"
+        assert result.matched_city == "上海"
+        assert result.hotel_amount > 0
+
+
 def test_agent_run_service_lists_seeded_trace_data() -> None:
     with build_session() as db:
         service = AgentRunService(db)

server/tests/test_document_intelligence.py

@@ -84,6 +84,33 @@ def test_document_intelligence_prefers_train_ticket_for_railway_e_ticket_invoice
     assert any(field.label == "金额" and field.value == "354元" for field in insight.fields)
 
 
+def test_document_intelligence_train_ticket_uses_railway_merchant_not_invoice_title() -> None:
+    insight = build_document_insight(
+        filename="2月20_武汉-上海.pdf",
+        summary="电子发票（铁路电子客票）；发票监；统一 制",
+        text=(
+            "电子发票（铁路电子客票）\n"
+            "发票号码:26429165800002785705 湖北\n"
+            "开票日期:2026年05月18日\n"
+            "武汉站 G458 上海虹桥站\n"
+            "Wuhan Shanghaihongqiao\n"
+            "2026年02月20日 07:55开 06车01B号 二等座\n"
+            "票价:￥354.00\n"
+            "4201061987****1615 曹笑竹\n"
+            "电子客票号:6580061086021391007342026\n"
+            "购买方名称:曹笑竹 统一社会信用代码:\n"
+            "买票请到12306 发货请到95306\n"
+            "中国铁路祝您旅途愉快"
+        ),
+    )
+
+    assert insight.document_type == "train_ticket"
+    fields = {field.label: field.value for field in insight.fields}
+    assert fields["商户"] == "中国铁路"
+    assert fields["金额"] == "354元"
+    assert fields["列车出发时间"] == "2026-02-20 07:55"
+
+
 def test_document_intelligence_recovers_train_ticket_from_english_station_ocr_text() -> None:
     insight = build_document_insight(
         filename="2月20_武汉-上海.pdf",

server/tests/test_expense_claim_service.py

@@ -28,6 +28,7 @@ from app.schemas.reimbursement import (
 )
 from app.services.agent_conversations import AgentConversationService
 from app.services.budget import BudgetService
+from app.services.document_preview import DocumentPreviewAssets
 from app.services.expense_claim_attachment_storage import ExpenseClaimAttachmentStorage
 from app.services.expense_claim_budget_flow import ExpenseClaimBudgetFlowMixin
 from app.services.expense_claim_workflow_constants import (
@@ -3314,6 +3315,68 @@ def test_attachment_preview_resolves_legacy_filename_in_claim_item_directory(mon
         assert filename == "legacy-ticket.pdf"
 
 
+def test_attachment_pdf_preview_falls_back_to_source_when_render_fonts_missing(monkeypatch, tmp_path) -> None:
+    current_user = CurrentUserContext(
+        username="emp-1",
+        name="张三",
+        role_codes=[],
+        is_admin=False,
+    )
+
+    monkeypatch.setattr(ExpenseClaimAttachmentStorage, "root", lambda self: tmp_path)
+
+    with build_session() as db:
+        claim = build_claim(expense_type="train", location="上海")
+        db.add(claim)
+        db.commit()
+
+        attachment_dir = tmp_path / claim.id / claim.items[0].id
+        attachment_dir.mkdir(parents=True)
+        file_path = attachment_dir / "2月20_武汉-上海.pdf"
+        preview_path = attachment_dir / "2月20_武汉-上海.preview.png"
+        file_path.write_bytes(b"%PDF-1.7 fake")
+        preview_path.write_bytes(b"broken-preview")
+        claim.items[0].invoice_id = f"{claim.id}/{claim.items[0].id}/{file_path.name}"
+        db.commit()
+
+        storage = ExpenseClaimAttachmentStorage()
+        storage.write_meta(
+            file_path,
+            {
+                "file_name": file_path.name,
+                "storage_key": storage.to_storage_key(file_path),
+                "media_type": "application/pdf",
+                "previewable": True,
+                "preview_kind": "image",
+                "preview_storage_key": storage.to_storage_key(preview_path),
+                "preview_media_type": "image/png",
+                "preview_file_name": preview_path.name,
+                "preview_rendered_with": "pdftoppm-png-r160-poppler-data",
+            },
+        )
+
+        def fake_render_pdf_first_page(*, pdf_path, preview_path, timeout_seconds):
+            raise RuntimeError("Missing language pack for 'Adobe-GB1' mapping")
+
+        monkeypatch.setattr(DocumentPreviewAssets, "render_pdf_first_page", fake_render_pdf_first_page)
+
+        resolved_path, media_type, filename = ExpenseClaimService(db).get_claim_item_attachment_preview_content(
+            claim_id=claim.id,
+            item_id=claim.items[0].id,
+            current_user=current_user,
+        )
+
+        assert resolved_path == file_path
+        assert media_type == "application/pdf"
+        assert filename == file_path.name
+        refreshed_meta = storage.read_meta(file_path)
+        assert refreshed_meta["preview_kind"] == "pdf"
+        assert refreshed_meta["preview_storage_key"] == storage.to_storage_key(file_path)
+        assert refreshed_meta["preview_media_type"] == "application/pdf"
+        assert refreshed_meta["preview_file_name"] == file_path.name
+        assert refreshed_meta["preview_rendered_with"] == ""
+
+
 def test_submit_claim_runs_ai_review_and_routes_to_direct_manager() -> None:
     current_user = CurrentUserContext(
         username="emp-submit@example.com",
@@ -5199,6 +5262,103 @@ def test_admin_delete_claim_unlinks_receipt_folder_items(monkeypatch, tmp_path)
         get_settings.cache_clear()
 
 
+def test_admin_delete_linked_reimbursement_resets_application_link_status() -> None:
+    admin_user = CurrentUserContext(
+        username="superadmin",
+        name="系统管理员",
+        role_codes=["admin"],
+        is_admin=True,
+    )
+
+    with build_session() as db:
+        application_claim = ExpenseClaim(
+            id="application-delete-linked-reimbursement",
+            claim_no="APP-DEL-LINKED-APPLICATION",
+            employee_name="张三",
+            department_name="交付部",
+            project_code="PRJ-A",
+            expense_type="travel_application",
+            reason="支撑国网仿生产环境部署",
+            location="上海",
+            amount=Decimal("3000.00"),
+            currency="CNY",
+            invoice_count=0,
+            occurred_at=datetime(2026, 6, 21, 22, 30, tzinfo=UTC),
+            submitted_at=datetime(2026, 6, 21, 22, 35, tzinfo=UTC),
+            status="approved",
+            approval_stage=APPLICATION_LINK_STATUS_STAGE,
+            risk_flags_json=[
+                {
+                    "source": "manual_approval",
+                    "event_type": "expense_application_approval",
+                    "operator": "向万红",
+                    "previous_approval_stage": DIRECT_MANAGER_APPROVAL_STAGE,
+                    "next_status": "approved",
+                    "next_approval_stage": APPLICATION_LINK_STATUS_STAGE,
+                    "generated_draft_claim_id": "reimbursement-delete-linked-application",
+                    "generated_draft_claim_no": "RDELETE01",
+                    "created_at": "2026-06-21T22:45:00+00:00",
+                }
+            ],
+        )
+        reimbursement_claim = ExpenseClaim(
+            id="reimbursement-delete-linked-application",
+            claim_no="RDELETE01",
+            employee_name="张三",
+            department_name="交付部",
+            project_code="PRJ-A",
+            expense_type="travel",
+            reason="支撑国网仿生产环境部署报销",
+            location="上海",
+            amount=Decimal("3000.00"),
+            currency="CNY",
+            invoice_count=1,
+            occurred_at=datetime(2026, 6, 21, 22, 30, tzinfo=UTC),
+            submitted_at=None,
+            status="draft",
+            approval_stage="待提交",
+            risk_flags_json=[
+                {
+                    "source": "application_handoff",
+                    "event_type": "expense_application_to_reimbursement_draft",
+                    "application_claim_id": application_claim.id,
+                    "application_claim_no": application_claim.claim_no,
+                }
+            ],
+        )
+        db.add_all([application_claim, reimbursement_claim])
+        db.commit()
+
+        deleted = ExpenseClaimService(db).delete_claim(reimbursement_claim.id, admin_user)
+
+        assert deleted is not None
+        assert deleted.claim_no == "RDELETE01"
+        assert db.get(ExpenseClaim, reimbursement_claim.id) is None
+        db.refresh(application_claim)
+        assert application_claim.status == "approved"
+        assert application_claim.approval_stage == APPLICATION_LINK_STATUS_STAGE
+
+        approval_flag = next(
+            flag
+            for flag in application_claim.risk_flags_json
+            if isinstance(flag, dict) and flag.get("event_type") == "expense_application_approval"
+        )
+        assert "generated_draft_claim_id" not in approval_flag
+        assert "generated_draft_claim_no" not in approval_flag
+
+        sync_flag = next(
+            flag
+            for flag in application_claim.risk_flags_json
+            if isinstance(flag, dict) and flag.get("event_type") == "expense_application_reimbursement_deleted"
+        )
+        assert sync_flag["source"] == "application_link_sync"
+        assert sync_flag["severity"] == "info"
+        assert sync_flag["actionability"] == "system_trace"
+        assert sync_flag["deleted_reimbursement_claim_id"] == "reimbursement-delete-linked-application"
+        assert sync_flag["deleted_reimbursement_claim_no"] == "RDELETE01"
+        assert sync_flag["next_approval_stage"] == APPLICATION_LINK_STATUS_STAGE
+
+
 def test_direct_manager_can_return_subordinate_claim_to_pending_submission() -> None:
     current_user = CurrentUserContext(
         username="manager-return@example.com",

server/tests/test_notification_states.py

@@ -85,6 +85,31 @@ def test_notification_state_service_persists_user_scoped_read_and_hidden_state()
         assert other_saved.states[0].hidden_at is None
 
 
+def test_notification_state_storage_ready_runs_once_per_database_bind(monkeypatch) -> None:
+    with build_session() as db:
+        service = NotificationStateService(db)
+        user = CurrentUserContext(username="alice", name="Alice", role_codes=[], is_admin=False)
+        calls: list[object] = []
+        original_create_all = Base.metadata.create_all
+
+        def track_create_all(*args, **kwargs):
+            calls.append(kwargs.get("bind"))
+            return original_create_all(*args, **kwargs)
+
+        monkeypatch.setattr(Base.metadata, "create_all", track_create_all)
+
+        service.list_states(user)
+        service.list_states(user)
+        service.patch_states(
+            NotificationStateBatchPatch(
+                states=[NotificationStatePatch(notification_id="workbench:todo:EXP-002", read=True)]
+            ),
+            user,
+        )
+
+        assert len(calls) == 1
+
+
 def test_notification_state_endpoint_reads_and_updates_current_user_state() -> None:
     client = build_client()
     headers = {"x-auth-username": "alice", "x-auth-name": "Alice"}

server/tests/test_ocr_service.py

@@ -5,19 +5,23 @@ import subprocess
 from pathlib import Path
 
 from app.core.config import get_settings
+from app.services import document_preview
 from app.services.ocr import OcrService
 
 
-def test_ocr_runtime_installers_include_poppler_cjk_data() -> None:
+def test_ocr_runtime_installers_include_cjk_safe_pdf_rendering_tools() -> None:
     repo_root = Path(__file__).resolve().parents[2]
     dependency_sources = [
         repo_root / "docker-compose.yml",
+        repo_root / "docker-compose.full.yml",
         repo_root / "server" / "scripts" / "bootstrap_paddleocr_mobile.sh",
         repo_root / "server" / "scripts" / "bootstrap_paddleocr_gpu.sh",
     ]
 
     for path in dependency_sources:
-        assert "poppler-data" in path.read_text(encoding="utf-8")
+        content = path.read_text(encoding="utf-8")
+        assert "poppler-data" in content
+        assert "mupdf-tools" in content
 
 
 def test_ocr_service_uses_worker_runtime_and_keeps_unsupported_files_as_warnings(
@@ -163,6 +167,7 @@ def test_ocr_service_passes_configured_device_to_worker(
         text: bool,
         timeout: int,
         check: bool,
+        env: dict[str, str] | None = None,
     ) -> subprocess.CompletedProcess[str]:
         captured_commands.append(command)
         return subprocess.CompletedProcess(
@@ -194,12 +199,12 @@ 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]:
+    def fake_convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> tuple[list[Path], bool]:
         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]
+        return [first, second], True
 
     def fake_invoke_worker(
         self,
@@ -281,26 +286,143 @@ def test_ocr_service_converts_pdf_to_images_and_returns_image_preview(
     assert recognized.lines[1].page_index == 1
 
 
-def test_ocr_service_uses_pdf_text_layer_without_worker_runtime(
+def test_ocr_service_rejects_pdf_ocr_when_rendered_image_fonts_are_broken(
     monkeypatch,
     tmp_path: Path,
 ) -> None:
-    def fake_convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> list[Path]:
+    def fake_convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> tuple[list[Path], bool]:
-        page = output_dir / "page-1.png"
+        raise RuntimeError("PDF 转图片失败：检测到中文字体映射缺失，未生成可 OCR 的图片。")
-        page.write_bytes(b"fake-rendered-page")
-        return [page]
 
-    def fail_resolve_python(self) -> str:
+    def fake_invoke_worker(
-        raise AssertionError("PDF 文本层可用时不应强制解析 OCR worker。")
+        self,
-
+        *,
-    def fail_invoke_worker(self, **kwargs) -> dict:
+        python_bin: str,
-        raise AssertionError("PDF 文本层可用时不应调用 OCR worker。")
+        worker_path: str,
+        input_paths: list[Path],
+    ) -> dict:
+        raise AssertionError("PDF 转图片已确认丢中文时，不应继续调用 OCR worker。")
 
     monkeypatch.setenv("STORAGE_ROOT_DIR", str(tmp_path / "storage"))
-    monkeypatch.setattr(OcrService, "_resolve_python_bin", fail_resolve_python)
+    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", fail_invoke_worker)
+    monkeypatch.setattr(OcrService, "_invoke_worker", fake_invoke_worker)
+    get_settings.cache_clear()
+    try:
+        result = OcrService().recognize_files(
+            [
+                ("2月20_武汉-上海.pdf", b"%PDF-1.7 fake", "application/pdf"),
+            ]
+        )
+    finally:
+        get_settings.cache_clear()
+
+    failed = result.documents[0]
+    assert failed.line_count == 0
+    assert failed.preview_kind == ""
+    assert failed.preview_data_url == ""
+    assert failed.warnings == ["PDF 转图片失败：检测到中文字体映射缺失，未生成可 OCR 的图片。"]
+
+
+def test_ocr_pdf_conversion_tries_next_renderer_when_poppler_font_mapping_fails(
+    monkeypatch,
+    tmp_path: Path,
+) -> None:
+    output_dir = tmp_path / "pages"
+    output_dir.mkdir()
+    calls: list[str] = []
+
+    def fake_run(
+        command: list[str],
+        *,
+        capture_output: bool,
+        text: bool,
+        timeout: int,
+        check: bool,
+    ) -> subprocess.CompletedProcess[str]:
+        calls.append(Path(command[0]).name)
+        if Path(command[0]).name == "pdftoppm":
+            (output_dir / "page-1.png").write_bytes(b"broken-preview")
+            return subprocess.CompletedProcess(
+                args=command,
+                returncode=0,
+                stdout="",
+                stderr="Syntax Error: Missing language pack for 'Adobe-GB1' mapping",
+            )
+        (output_dir / "page-1.png").write_bytes(b"rendered-with-chinese")
+        return subprocess.CompletedProcess(
+            args=command,
+            returncode=0,
+            stdout="",
+            stderr="",
+        )
+
+    monkeypatch.setattr(
+        document_preview.shutil,
+        "which",
+        lambda name: f"/usr/bin/{name}" if name in {"pdftoppm", "mutool"} else None,
+    )
+    monkeypatch.setattr(subprocess, "run", fake_run)
+
+    pages, preview_usable = OcrService()._convert_pdf_to_images(
+        pdf_path=tmp_path / "ticket.pdf",
+        output_dir=output_dir,
+    )
+
+    assert pages == [output_dir / "page-1.png"]
+    assert preview_usable is True
+    assert calls == ["pdftoppm", "mutool"]
+
+
+def test_ocr_service_invokes_worker_even_when_pdf_text_layer_is_usable(
+    monkeypatch,
+    tmp_path: Path,
+) -> None:
+    calls = {"worker": 0}
+
+    def fake_convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> tuple[list[Path], bool]:
+        page = output_dir / "page-1.png"
+        page.write_bytes(b"fake-rendered-page")
+        return [page], True
+
+    def fake_invoke_worker(
+        self,
+        *,
+        python_bin: str,
+        worker_path: str,
+        input_paths: list[Path],
+    ) -> dict:
+        calls["worker"] += 1
+        return {
+            "engine": "paddleocr_mobile",
+            "model": "PP-OCRv5_mobile",
+            "documents": [
+                {
+                    "input_path": str(input_paths[0]),
+                    "engine": "paddleocr_mobile",
+                    "model": "PP-OCRv5_mobile",
+                    "text": "电子发票（铁路电子客票） 武汉站 上海虹桥站 G458 票价 ￥354.00",
+                    "summary": "铁路电子客票",
+                    "avg_score": 0.95,
+                    "line_count": 1,
+                    "page_count": 1,
+                    "warnings": [],
+                    "lines": [
+                        {
+                            "text": "电子发票（铁路电子客票） 武汉站 上海虹桥站 G458 票价 ￥354.00",
+                            "score": 0.95,
+                            "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)
     monkeypatch.setattr(
         OcrService,
         "_extract_pdf_text_layer",
@@ -326,9 +448,9 @@ def test_ocr_service_uses_pdf_text_layer_without_worker_runtime(
 
     recognized = result.documents[0]
     assert result.success_count == 1
+    assert calls["worker"] == 1
     assert recognized.document_type == "train_ticket"
     assert "电子发票（铁路电子客票）" in recognized.text
-    assert "电子客票号:6580061086021391007342026" in recognized.text
     assert any(field.label == "金额" and field.value == "354元" for field in recognized.document_fields)
     assert recognized.preview_kind == "image"
     assert recognized.preview_data_url.startswith("data:image/png;base64,")
@@ -392,14 +514,22 @@ def test_ocr_service_reuses_cached_document_for_same_content(
     assert second.documents[0].summary == first.documents[0].summary
 
 
+def test_ocr_cache_key_includes_pdf_render_pipeline_version() -> None:
+    cache_key = OcrService()._build_cache_key(b"same-pdf-content")
+
+    assert "pdf-image-ocr:" in cache_key
+    assert document_preview.DocumentPreviewAssets.PDF_RENDERER_ID in cache_key
+    assert "no-pdf-direct" in cache_key
+
+
 def test_ocr_service_prefers_pdf_text_layer_when_rendered_ocr_is_placeholder_heavy(
     monkeypatch,
     tmp_path: Path,
 ) -> None:
-    def fake_convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> list[Path]:
+    def fake_convert_pdf_to_images(self, *, pdf_path: Path, output_dir: Path) -> tuple[list[Path], bool]:
         page = output_dir / "page-1.png"
         page.write_bytes(b"fake-page")
-        return [page]
+        return [page], True
 
     def fake_invoke_worker(
         self,

server/tests/test_receipt_folder_service.py

@@ -4,7 +4,7 @@ import base64
 
 from app.api.deps import CurrentUserContext
 from app.core.config import get_settings
-from app.schemas.ocr import OcrRecognizeDocumentRead
+from app.schemas.ocr import OcrRecognizeDocumentRead, OcrRecognizeFieldRead
 from app.services.document_preview import DocumentPreviewAssets
 from app.services.receipt_folder import ReceiptFolderService
 
@@ -72,6 +72,55 @@ def test_receipt_folder_train_ticket_uses_invoice_date_and_enriches_fields(monke
         get_settings.cache_clear()
 
 
+def test_receipt_folder_pdf_save_eagerly_renders_image_preview(monkeypatch, tmp_path) -> None:
+    monkeypatch.setenv("STORAGE_ROOT_DIR", str(tmp_path / "storage"))
+    get_settings.cache_clear()
+    try:
+        current_user = CurrentUserContext(
+            username="pytest",
+            name="Py Test",
+            role_codes=[],
+            is_admin=False,
+        )
+
+        def fake_render_pdf_first_page(*, pdf_path, preview_path, timeout_seconds):
+            preview_path.write_bytes(b"rendered-preview")
+            return preview_path
+
+        monkeypatch.setattr(DocumentPreviewAssets, "render_pdf_first_page", fake_render_pdf_first_page)
+
+        service = ReceiptFolderService()
+        receipt = service.save_receipt(
+            filename="2月20_武汉-上海.pdf",
+            content=b"%PDF-1.4 fake",
+            media_type="application/pdf",
+            current_user=current_user,
+            document=OcrRecognizeDocumentRead(
+                filename="2月20_武汉-上海.pdf",
+                media_type="application/pdf",
+                text="铁路电子客票 武汉 上海虹桥 354.00",
+                summary="铁路电子客票，武汉至上海虹桥。",
+            ),
+        )
+
+        receipt_dir = next(service.root.glob("pytest/*"))
+        preview_path = receipt_dir / "preview.png"
+        meta = service._read_meta(receipt_dir)
+
+        assert receipt.preview_kind == "image"
+        assert preview_path.read_bytes() == b"rendered-preview"
+        assert meta["preview_file_name"] == "preview.png"
+        assert meta["preview_media_type"] == "image/png"
+        assert meta["preview_rendered_with"] == DocumentPreviewAssets.PDF_RENDERER_ID
+
+        resolved_path, media_type, file_name = service.resolve_preview(receipt.id, current_user)
+        assert resolved_path == preview_path
+        assert media_type == "image/png"
+        assert file_name == "preview.png"
+    finally:
+        get_settings.cache_clear()
+
+
 def test_receipt_folder_pdf_preview_regenerates_stale_cached_image(monkeypatch, tmp_path) -> None:
     monkeypatch.setenv("STORAGE_ROOT_DIR", str(tmp_path / "storage"))
     get_settings.cache_clear()
@@ -123,6 +172,213 @@ def test_receipt_folder_pdf_preview_regenerates_stale_cached_image(monkeypatch,
         get_settings.cache_clear()
 
 
+def test_receipt_folder_pdf_preview_falls_back_to_source_when_render_fonts_missing(
+    monkeypatch,
+    tmp_path,
+) -> None:
+    monkeypatch.setenv("STORAGE_ROOT_DIR", str(tmp_path / "storage"))
+    get_settings.cache_clear()
+    try:
+        current_user = CurrentUserContext(
+            username="pytest",
+            name="Py Test",
+            role_codes=[],
+            is_admin=False,
+        )
+        stale_preview = b"broken-preview"
+        preview_data_url = f"data:image/png;base64,{base64.b64encode(stale_preview).decode('ascii')}"
+        service = ReceiptFolderService()
+        receipt = service.save_receipt(
+            filename="2月20_武汉-上海.pdf",
+            content=b"%PDF-1.7 fake",
+            media_type="application/pdf",
+            current_user=current_user,
+            document=OcrRecognizeDocumentRead(
+                filename="2月20_武汉-上海.pdf",
+                media_type="application/pdf",
+                preview_kind="image",
+                preview_data_url=preview_data_url,
+            ),
+        )
+
+        receipt_dir = next(service.root.glob("pytest/*"))
+        meta = service._read_meta(receipt_dir)
+        meta["preview_rendered_with"] = "pdftoppm-png-r160-poppler-data"
+        service._write_meta(receipt_dir, meta)
+
+        def fake_render_pdf_first_page(*, pdf_path, preview_path, timeout_seconds):
+            raise RuntimeError("Missing language pack for 'Adobe-GB1' mapping")
+
+        monkeypatch.setattr(DocumentPreviewAssets, "render_pdf_first_page", fake_render_pdf_first_page)
+
+        resolved_path, media_type, file_name = service.resolve_preview(receipt.id, current_user)
+
+        assert resolved_path == receipt_dir / "2月20_武汉-上海.pdf"
+        assert media_type == "application/pdf"
+        assert file_name == "2月20_武汉-上海.pdf"
+        refreshed_meta = service._read_meta(receipt_dir)
+        assert refreshed_meta["preview_kind"] == "pdf"
+        assert refreshed_meta["preview_file_name"] == "2月20_武汉-上海.pdf"
+        assert refreshed_meta["preview_media_type"] == "application/pdf"
+        assert refreshed_meta["preview_rendered_with"] == ""
+    finally:
+        get_settings.cache_clear()
+
+
+def test_receipt_folder_train_ticket_extracts_passenger_from_id_line_and_purchase_name(
+    monkeypatch,
+    tmp_path,
+) -> None:
+    monkeypatch.setenv("STORAGE_ROOT_DIR", str(tmp_path / "storage"))
+    get_settings.cache_clear()
+    try:
+        current_user = CurrentUserContext(
+            username="pytest",
+            name="Py Test",
+            role_codes=[],
+            is_admin=False,
+        )
+        service = ReceiptFolderService()
+        receipt = service.save_receipt(
+            filename="2月20_武汉-上海.pdf",
+            content=b"%PDF-1.4 fake",
+            media_type="application/pdf",
+            current_user=current_user,
+            document=OcrRecognizeDocumentRead(
+                filename="2月20_武汉-上海.pdf",
+                media_type="application/pdf",
+                text=(
+                    "电子发票（铁路电子客票）\n"
+                    "发票号码:26429165800002785705 湖北\n"
+                    "开票日期:2026年05月18日\n"
+                    "武汉站 G458 上海虹桥站\n"
+                    "Wuhan Shanghaihongqiao\n"
+                    "2026年02月20日 07:55开 06车01B号 二等座\n"
+                    "票价:￥354.00\n"
+                    "4201061987****1615 曹笑竹\n"
+                    "电子客票号:6580061086021391007342026\n"
+                    "购买方名称:曹笑竹 统一社会信用代码:\n"
+                    "买票请到12306 发货请到95306\n"
+                    "中国铁路祝您旅途愉快"
+                ),
+                summary="电子发票（铁路电子客票）；发票监；统一 制",
+                document_type="train_ticket",
+                document_type_label="火车/高铁票",
+                scene_code="travel",
+                scene_label="差旅票据",
+                document_fields=[
+                    OcrRecognizeFieldRead(key="merchant_name", label="商户", value="电子发票（铁路"),
+                    OcrRecognizeFieldRead(key="amount", label="金额", value="354元"),
+                    OcrRecognizeFieldRead(key="date", label="列车出发时间", value="2026-02-20 07:55"),
+                    OcrRecognizeFieldRead(key="trip_no", label="车次", value="G458"),
+                    OcrRecognizeFieldRead(key="route", label="行程", value="武汉-上海"),
+                ],
+            ),
+        )
+
+        assert receipt.merchant_name == "中国铁路"
+
+        detail = service.get_receipt(receipt.id, current_user)
+        fields = {field.label: field.value for field in detail.fields}
+        assert fields["商户"] == "中国铁路"
+        assert fields["乘车人"] == "曹笑竹"
+        assert fields["出发地点"] == "武汉"
+        assert fields["到达地点"] == "上海虹桥"
+        assert fields["身份证号"] == "4201061987****1615"
+        assert fields["电子客票号"] == "6580061086021391007342026"
+        assert fields["开票日期"] == "2026-05-18"
+        assert fields["列车出发时间"] == "2026-02-20 07:55"
+        assert fields["车厢"] == "06车"
+        assert fields["座位号"] == "01B"
+    finally:
+        get_settings.cache_clear()
+
+
+def test_receipt_folder_train_ticket_repairs_invalid_generated_fields_from_ocr_text(
+    monkeypatch,
+    tmp_path,
+) -> None:
+    monkeypatch.setenv("STORAGE_ROOT_DIR", str(tmp_path / "storage"))
+    get_settings.cache_clear()
+    try:
+        current_user = CurrentUserContext(
+            username="pytest",
+            name="Py Test",
+            role_codes=[],
+            is_admin=False,
+        )
+        service = ReceiptFolderService()
+        receipt = service.save_receipt(
+            filename="2月21日_上海-深圳.png",
+            content=b"fake image",
+            media_type="image/png",
+            current_user=current_user,
+            document=OcrRecognizeDocumentRead(
+                filename="2月21日_上海-深圳.png",
+                media_type="image/png",
+                text=(
+                    "行程单示意\n"
+                    "出票渠道：示例平台\n"
+                    "非官方车票\n"
+                    "不可报销\n"
+                    "仅供演示\n"
+                    "创建日期：2026年02月15日\n"
+                    "订单号：DEMO202602210001\n"
+                    "单据编号：DEMO-IT-000001\n"
+                    "上海虹桥\n"
+                    "G999\n"
+                    "深圳北\n"
+                    "站\n"
+                    "站\n"
+                    "Shanghaihongqiao\n"
+                    "Shenzhenbei\n"
+                    "2026年02月21日\n"
+                    "08:30出发\n"
+                    "全程约7小时30分\n"
+                    "15:00到达\n"
+                    "DEMO\n"
+                    "乘客：示例旅客\n"
+                    "车厢：05车\n"
+                    "席别：二等座\n"
+                    "-\n"
+                    "扫码无效\n"
+                    "证件号：310101199001010000\n"
+                    "座位：08A\n"
+                    "票价：￥438.00\n"
+                    "仅为演示"
+                ),
+                summary="行程单示意；出票渠道：示例平台；非官方车票",
+                document_type="train_ticket",
+                document_type_label="火车/高铁票",
+                scene_code="travel",
+                scene_label="差旅票据",
+                document_fields=[
+                    OcrRecognizeFieldRead(key="amount", label="金额", value="438元"),
+                    OcrRecognizeFieldRead(key="date", label="列车出发时间", value="2026-02-21 08:30"),
+                    OcrRecognizeFieldRead(key="invoice_number", label="票据号码", value="DEMO202602210001"),
+                    OcrRecognizeFieldRead(key="trip_no", label="车次", value="G999"),
+                    OcrRecognizeFieldRead(key="route", label="行程", value="上海-深圳"),
+                    OcrRecognizeFieldRead(key="departure_station", label="出发地点", value="二等座"),
+                    OcrRecognizeFieldRead(key="arrival_station", label="到达地点", value="扫码无效"),
+                    OcrRecognizeFieldRead(key="passenger_name", label="乘车人", value="席别二等座"),
+                ],
+            ),
+        )
+
+        detail = service.get_receipt(receipt.id, current_user)
+        fields = {field.label: field.value for field in detail.fields}
+        assert fields["出发地点"] == "上海虹桥"
+        assert fields["到达地点"] == "深圳北"
+        assert fields["乘车人"] == "示例旅客"
+        assert fields["身份证号"] == "310101199001010000"
+        assert fields["席别"] == "二等座"
+        assert fields["车厢"] == "05车"
+        assert fields["座位号"] == "08A"
+        assert fields["票价"] == "438.00元"
+    finally:
+        get_settings.cache_clear()
+
+
 def test_receipt_folder_delete_removes_duplicate_marker(monkeypatch, tmp_path) -> None:
     monkeypatch.setenv("STORAGE_ROOT_DIR", str(tmp_path / "storage"))
     get_settings.cache_clear()

server/tests/test_user_agent_service.py

@@ -15,6 +15,7 @@ from app.models.financial_record import ExpenseClaim
 from app.schemas.ontology import OntologyParseRequest
 from app.schemas.user_agent import UserAgentCitation, UserAgentRequest, UserAgentReviewRiskBrief
 from app.services.agent_assets import AgentAssetService
+from app.services.application_location_semantics import resolve_jieba_tokens
 from app.services.ontology import SemanticOntologyService
 from app.services.user_agent import UserAgentService
 from app.services.user_agent_documents import UserAgentDocumentService
@@ -763,6 +764,67 @@ def test_user_agent_application_submit_blocks_overlapping_travel_dates() -> None
         assert response.draft_payload is None
 
 
+def test_user_agent_application_submit_normalizes_location_mixed_with_business_content() -> None:
+    session_factory = build_session_factory()
+    with session_factory() as db:
+        response = build_application_user_agent_response(
+            db,
+            "确认提交",
+            context_overrides={
+                "manager_name": "向万红",
+                "application_preview": {
+                    "fields": {
+                        "applicationType": "差旅费用申请",
+                        "time": "2026-02-20 至 2026-02-23",
+                        "location": "上海辅助国网仿生产服务器",
+                        "reason": "辅助国网仿生产服务器部署",
+                        "days": "4天",
+                        "transportMode": "火车",
+                        "amount": "2120元",
+                    }
+                },
+            },
+        )
+
+        claim = application_claim_query(db).one()
+        assert claim.location == "上海市"
+        assert claim.reason == "辅助国网仿生产服务器部署"
+        assert "申请单据已生成" in response.answer
+        assert response.draft_payload is not None
+
+
+def test_user_agent_application_submit_splits_location_and_reason_from_raw_sentence() -> None:
+    session_factory = build_session_factory()
+    with session_factory() as db:
+        response = build_application_user_agent_response(
+            db,
+            "确认提交",
+            history=[
+                {
+                    "role": "user",
+                    "content": "2026-02-20 至 2026-02-23，去上海辅助国网仿生产服务器部署，火车",
+                }
+            ],
+            context_overrides={
+                "manager_name": "向万红",
+                "grade": "P5",
+                "department_name": "技术部",
+            },
+        )
+
+        claim = application_claim_query(db).one()
+        assert claim.location == "上海市"
+        assert claim.reason == "辅助国网仿生产服务器部署"
+        assert "申请单据已生成" in response.answer
+
+
+def test_application_sentence_jieba_tokenizer_recognizes_location_boundary() -> None:
+    tokens = resolve_jieba_tokens("上海辅助国网仿生产服务器部署")
+
+    assert ("上海", "ns") in tokens
+    assert [word for word, _ in tokens] == ["上海", "辅助", "国网", "仿生产", "服务器", "部署"]
+
+
 def test_user_agent_application_maps_preview_travel_type_label() -> None:
     session_factory = build_session_factory()
     with session_factory() as db:
@@ -2155,7 +2217,7 @@ def test_user_agent_returns_draft_limit_message_when_save_is_blocked() -> None:
                 context_json={"review_action": "save_draft"},
                 tool_payload={
                     "draft_limit_reached": True,
-                    "message": "你当前已保存 3 个草稿，请先完成已保存的草稿，才能再次新建草稿。",
+                    "message": "您当前已保存 3 个草稿，请先完成已保存的草稿，才能再次新建草稿。",
                     "status": "blocked",
                 },
             )
@@ -2163,7 +2225,7 @@ def test_user_agent_returns_draft_limit_message_when_save_is_blocked() -> None:
 
         assert (
             response.answer
-            == "你当前已保存 3 个草稿，请先完成已保存的草稿，才能再次新建草稿。"
+            == "您当前已保存 3 个草稿，请先完成已保存的草稿，才能再次新建草稿。"
         )