feat(govdoc): 新增内部公文模块全链路（后端58+前端11文件）

fastapi_modules/fastapi_leaudit/govdoc_engine/__init__.py

@@ -0,0 +1,35 @@
+"""Govdoc 公文格式审查引擎内核。
+
+从旧 govdoc-audit 项目裁剪迁入，去除独立 API 层、SQLite 存储层、
+本地运行记录器 (RunRecorder) 和旧配置系统。
+
+导出:
+  - pipeline.run() — 异步审查入口 (bridge 层主调用)
+  - pipeline.audit_file() — 同步审查入口 (兼容)
+  - models — 核心数据模型 (Pydantic)
+  - parser — 文档解析与实体抽取
+  - dsl — YAML 规则 DSL 定义与加载
+  - engine — 规则执行引擎与结果模型
+  - reporter — 报告生成 (HTML/DOCX/JSON)
+  - llm — LLM 客户端 (OpenAI 兼容协议)
+"""
+
+from __future__ import annotations
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.pipeline import (
+    audit_file,
+    run,
+)
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.result import (
+    AuditResult,
+    AuditSummary,
+    CheckedRule,
+)
+
+__all__ = [
+    "audit_file",
+    "run",
+    "AuditResult",
+    "AuditSummary",
+    "CheckedRule",
+]

fastapi_modules/fastapi_leaudit/govdoc_engine/dsl/loader.py

@@ -0,0 +1,24 @@
+"""加载并校验 rules.yaml。"""
+
+from __future__ import annotations
+from functools import lru_cache
+from pathlib import Path
+import yaml
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.dsl.schema import RuleSet
+
+
+def _load_uncached(path: Path) -> RuleSet:
+    with open(path, encoding="utf-8") as f:
+        data = yaml.safe_load(f)
+    return RuleSet.model_validate(data)
+
+
+@lru_cache(maxsize=32)
+def _load_cached(path_str: str, mtime: float) -> RuleSet:
+    return _load_uncached(Path(path_str))
+
+
+def load_rules(path: str | Path) -> RuleSet:
+    path = Path(path)
+    return _load_cached(str(path.resolve()), path.stat().st_mtime)

fastapi_modules/fastapi_leaudit/govdoc_engine/dsl/schema.py

@@ -0,0 +1,141 @@
+"""规则文件的 Pydantic schema。"""
+
+from __future__ import annotations
+from typing import Any, Literal
+from pydantic import BaseModel, Field, model_validator
+
+
+CheckType = Literal[
+    "required", "font", "style_match", "line_spacing",
+    "attachment_marker_style",
+    "regex_require", "regex_forbid",
+    "confused_pair", "forbid_phrase", "forbid_chars",
+    "punctuation", "wenzhong_whitelist",
+    "hierarchy", "cross_role", "ai",
+]
+
+
+class AppliesTo(BaseModel):
+    role: str | None = None
+    roles: list[str] | None = None
+    paragraph_index: int | None = None
+
+
+class RuleStage(BaseModel):
+    id: str | None = None
+    check: CheckType
+    when: str | None = None
+    field: str | None = None
+    expect: dict[str, Any] | None = None
+    pattern: str | None = None
+    chars: list[str] | None = None
+    pairs: list[dict[str, Any]] | None = None
+    phrases: list[str] | None = None
+    rules: list[dict[str, Any]] | None = None
+    expected_order: list[dict[str, Any]] | None = None
+    forbid_patterns: list[str] | None = None
+    prompt: str | None = None
+    format: str | None = None
+
+    model_config = {"extra": "allow"}
+
+
+class Messages(BaseModel):
+    pass_msg: str = Field(alias="pass", default="")
+    fail: str = ""
+
+    model_config = {"populate_by_name": True}
+
+
+class Rule(BaseModel):
+    rule_id: str
+    name: str
+    severity: Literal["error", "warning", "info"] = "warning"
+    category: str
+    score: int | None = None
+    # 二选一：target 通道（推荐，绑定语义实体）或 applies_to 通道（旧，按 role 选段）
+    applies_to: AppliesTo | None = None
+    target: str | None = None
+    on_missing: Literal["pass", "warn", "fail", "skip"] = "skip"
+    activate_if: str | None = None
+    stages: list[RuleStage]
+    messages: Messages
+
+    @model_validator(mode="after")
+    def _check_at_least_one_target(self) -> "Rule":
+        if self.applies_to is None and self.target is None:
+            raise ValueError(
+                f"Rule {self.rule_id}: 必须声明 target 或 applies_to 之一"
+            )
+        return self
+
+
+class RuleGroup(BaseModel):
+    group: str
+    rules: list[Rule]
+
+
+# 8 个内置语义实体名（与 entity_builder.BUILTIN_ENTITY_NAMES 同步）
+_BUILTIN_ENTITY_NAMES = frozenset({
+    "title", "doc_number", "recipient", "date",
+    "signature", "attachments", "wenzhong", "issuer",
+})
+
+
+class EntitySpec(BaseModel):
+    """声明用户自定义语义实体（builtin 实体由代码自动产出，无需在 yaml 出现）。"""
+
+    name: str
+    type: Literal["string", "number", "list"] = "string"
+    description: str = ""
+
+    @model_validator(mode="after")
+    def _no_clash_with_builtin(self) -> "EntitySpec":
+        if self.name in _BUILTIN_ENTITY_NAMES:
+            raise ValueError(
+                f"entity '{self.name}' 与内置实体重名，"
+                f"去掉该声明即可（内置实体自动产出）"
+            )
+        return self
+
+
+class ExtractSpec(BaseModel):
+    entities: list[EntitySpec] = Field(default_factory=list)
+
+
+class RuleSetMetadata(BaseModel):
+    type_id: str
+    name: str
+    version: str
+    source: str | None = None
+    description: str | None = None
+
+
+class RuleSet(BaseModel):
+    metadata: RuleSetMetadata
+    extract: ExtractSpec = Field(default_factory=ExtractSpec)
+    rules: list[RuleGroup]
+
+    @model_validator(mode="after")
+    def _check_unique_ids(self) -> "RuleSet":
+        seen: set[str] = set()
+        for g in self.rules:
+            for r in g.rules:
+                if r.rule_id in seen:
+                    raise ValueError(f"duplicate rule_id: {r.rule_id}")
+                seen.add(r.rule_id)
+        return self
+
+    def all_rules(self) -> list[Rule]:
+        return [r for g in self.rules for r in g.rules]
+
+
+class FontCheck(BaseModel):
+    eastasia: str | None = None
+    ascii: str | None = None
+    size_pt: float | None = None
+
+
+class RegexForbidCheck(BaseModel):
+    pattern: str
+    message: str | None = None

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/__init__.py

@@ -0,0 +1,24 @@
+"""Check 原语注册中心：通过 register 装饰器收集，runner 通过 get 查找。"""
+
+from __future__ import annotations
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckBase
+
+_REGISTRY: dict[str, type[CheckBase]] = {}
+
+
+def register(name: str):
+    def deco(cls):
+        cls.name = name
+        _REGISTRY[name] = cls
+        return cls
+    return deco
+
+
+def get_check(name: str) -> type[CheckBase]:
+    if name not in _REGISTRY:
+        raise KeyError(f"unknown check: {name}; known: {list(_REGISTRY)}")
+    return _REGISTRY[name]
+
+
+def all_checks() -> list[str]:
+    return list(_REGISTRY.keys())

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/ai_check.py

@@ -0,0 +1,151 @@
+"""LLM 语义检查。三级输出：pass / warn / fail。"""
+
+import logging
+import re
+from typing import Any
+from pydantic import BaseModel
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import (
+    CheckBase, CheckContext, CheckHit, CheckResult,
+)
+from fastapi_modules.fastapi_leaudit.govdoc_engine.llm.client import LlmClient, LlmJsonError, _format_exc
+
+_log = logging.getLogger(__name__)
+
+
+_OUT_FORMAT = """
+请以 JSON 输出：
+{"result": "pass|warn|fail", "reason": "<简短理由>", "suggestion": "<改进建议；pass 时填空>"}
+"""
+
+_VAR_RE = re.compile(r"\{\{\s*([^}]+?)\s*\}\}")
+
+
+def _resolve_dot_path(root: Any, path: str) -> str:
+    """点语法属性访问：title.style.font_eastasia → entities['title'].style.font_eastasia。"""
+    cur: Any = root
+    for seg in path.split("."):
+        if cur is None:
+            return ""
+        if isinstance(cur, dict):
+            cur = cur.get(seg)
+        elif isinstance(cur, BaseModel):
+            cur = getattr(cur, seg, None)
+        else:
+            cur = getattr(cur, seg, None)
+    if cur is None:
+        return ""
+    if isinstance(cur, (dict, list)):
+        return str(cur)
+    return str(cur)
+
+
+def _interpolate(
+    template: str,
+    paragraphs: list,
+    entities: dict | None = None,
+    target: Any = None,
+) -> str:
+    """渲染顺序：① paragraphs[N] ② entities 点语法
+    ③ target 隐式（无前缀时视为 target.<key>）。"""
+    entities = entities or {}
+
+    def repl(m):
+        key = m.group(1).strip()
+        # ① paragraphs[N] 索引
+        if key.startswith("paragraphs["):
+            try:
+                idx = int(key[len("paragraphs["):].rstrip("]"))
+                return paragraphs[idx].text
+            except (ValueError, IndexError):
+                return ""
+        # ② entities 点语法：title.text / title.style.font_eastasia
+        head, _, rest = key.partition(".")
+        if head in entities:
+            entity = entities[head]
+            if entity is None:
+                return ""
+            return _resolve_dot_path(entity, rest) if rest else entity.text
+        # ③ target 隐式：未带前缀且 target 存在
+        if target is not None:
+            v = _resolve_dot_path(target, key)
+            if v:
+                return v
+        return ""
+
+    return _VAR_RE.sub(repl, template)
+
+
+@register("ai")
+class AiCheck(CheckBase):
+    def __init__(self, llm_client: LlmClient | None = None):
+        self.client = llm_client or LlmClient()
+
+    def _build_prompt(self, ctx: CheckContext) -> str:
+        prompt = _interpolate(
+            ctx.stage.prompt or "",
+            ctx.paragraphs,
+            ctx.entities,
+            ctx.target,
+        )
+        return prompt + "\n\n" + _OUT_FORMAT
+
+    def _interpret(self, ctx: CheckContext, resp: dict) -> CheckResult:
+        result = resp.get("result", "fail")
+        reason = resp.get("reason", "")
+        suggestion = resp.get("suggestion", "")
+        if result == "pass":
+            return CheckResult(passed=True, hits=[])
+        target_p = ctx.paragraphs[0] if ctx.paragraphs else None
+        confidence = 0.95 if result == "fail" else 0.7
+        return CheckResult(passed=False, hits=[CheckHit(
+            paragraph=target_p,
+            char_start=0,
+            char_end=len(target_p.text) if target_p else 0,
+            actual={"llm_reason": reason, "llm_suggestion": suggestion},
+            expected={},
+            message=reason or "LLM 判定不通过",
+            confidence=confidence,
+        )])
+
+    def run(self, ctx: CheckContext) -> CheckResult:
+        label = f"ai_{ctx.rule_id or 'unknown'}"
+        try:
+            resp = self.client.chat_json(
+                [{"role": "user", "content": self._build_prompt(ctx)}],
+                label=label,
+            )
+        except LlmJsonError as e:
+            _log.warning("AI check skipped (LLM JSON error): %s", _format_exc(e))
+            return CheckResult(
+                passed=True, hits=[], skipped=True,
+                skip_reason=f"LLM 返回内容无法解析为 JSON：{e}",
+            )
+        except Exception as e:
+            _log.warning("AI check skipped (LLM error): %s", _format_exc(e))
+            return CheckResult(
+                passed=True, hits=[], skipped=True,
+                skip_reason=f"LLM 调用失败：{e}",
+            )
+        return self._interpret(ctx, resp)
+
+    async def run_async(self, ctx: CheckContext) -> CheckResult:
+        label = f"ai_{ctx.rule_id or 'unknown'}"
+        try:
+            resp = await self.client.chat_json_async(
+                [{"role": "user", "content": self._build_prompt(ctx)}],
+                label=label,
+            )
+        except LlmJsonError as e:
+            _log.warning("AI check skipped (LLM JSON error): %s", _format_exc(e))
+            return CheckResult(
+                passed=True, hits=[], skipped=True,
+                skip_reason=f"LLM 返回内容无法解析为 JSON：{e}",
+            )
+        except Exception as e:
+            _log.warning("AI check skipped (LLM error): %s", _format_exc(e))
+            return CheckResult(
+                passed=True, hits=[], skipped=True,
+                skip_reason=f"LLM 调用失败：{e}",
+            )
+        return self._interpret(ctx, resp)

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/base.py

@@ -0,0 +1,48 @@
+"""Check 原语基类与上下文。"""
+
+from __future__ import annotations
+from dataclasses import dataclass, field
+from typing import Any, TYPE_CHECKING
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document, Paragraph
+from fastapi_modules.fastapi_leaudit.govdoc_engine.dsl.schema import RuleStage
+
+if TYPE_CHECKING:
+    from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.entities import SemanticEntity
+
+
+@dataclass
+class CheckContext:
+    document: Document
+    paragraphs: list[Paragraph]
+    stage: RuleStage
+    entities: dict[str, "SemanticEntity | None"] = field(default_factory=dict)
+    target: "SemanticEntity | None" = None
+    rule_id: str = ""
+
+
+@dataclass
+class CheckHit:
+    paragraph: Paragraph | None
+    char_start: int = 0
+    char_end: int = 0
+    actual: dict[str, Any] | None = None
+    expected: dict[str, Any] | None = None
+    message: str | None = None
+    confidence: float = 1.0
+
+
+@dataclass
+class CheckResult:
+    passed: bool
+    hits: list[CheckHit] = field(default_factory=list)
+    skipped: bool = False
+    skip_reason: str = ""
+
+
+class CheckBase:
+    """所有 check 原语的抽象基类。"""
+
+    name: str = ""
+
+    def run(self, ctx: CheckContext) -> CheckResult:
+        raise NotImplementedError

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/confused_pair.py

@@ -0,0 +1,34 @@
+"""易混淆词对（字面 + 正则）。"""
+
+import re
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckBase, CheckContext, CheckHit, CheckResult
+
+
+@register("confused_pair")
+class ConfusedPairCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        pairs = ctx.stage.pairs or []
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            for pair in pairs:
+                wrong = pair.get("wrong")
+                wrong_pat = pair.get("wrong_pattern")
+                correct = pair.get("correct") or pair.get("suggest", "")
+                reason = pair.get("reason", "")
+                if wrong and wrong in p.text:
+                    start = p.text.find(wrong)
+                    hits.append(CheckHit(
+                        paragraph=p, char_start=start, char_end=start + len(wrong),
+                        actual={"text": wrong}, expected={"text": correct},
+                        message=f"\"{wrong}\" 应为 \"{correct}\"。{reason}",
+                    ))
+                elif wrong_pat:
+                    for m in re.finditer(wrong_pat, p.text):
+                        hits.append(CheckHit(
+                            paragraph=p, char_start=m.start(), char_end=m.end(),
+                            actual={"text": m.group(0)},
+                            expected={"text": correct},
+                            message=f"\"{m.group(0)}\" 应为 \"{correct}\"。{reason}",
+                        ))
+        return CheckResult(passed=not hits, hits=hits)

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/cross_role.py

@@ -0,0 +1,69 @@
+"""跨段关系 check：例如二级标题以句号结尾后又新起一段。"""
+
+import re
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckBase, CheckContext, CheckHit, CheckResult
+
+
+# 单个附件项末尾的标点：编号(数字+点) + 内容 + 末尾标点
+_ATTACH_ITEM_TRAIL_PUNCT = re.compile(r"\d+[\.．][^\d；。，;,.]+?[；。，;,.]")
+# 整段是一个附件项
+_ATTACH_ITEM_LINE = re.compile(r"^\d+[\.．].+[；。，;,.]\s*$")
+
+
+@register("cross_role")
+class CrossRoleCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        rules = ctx.stage.rules or []
+        paras = ctx.document.paragraphs
+        hits: list[CheckHit] = []
+        for r in rules:
+            t = r.get("type")
+            if t == "h2_no_period_then_break":
+                for i, p in enumerate(paras):
+                    if p.role == "heading_2" and p.text.rstrip().endswith(("。", ".")):
+                        if i + 1 < len(paras) and paras[i + 1].text.strip():
+                            hits.append(CheckHit(
+                                paragraph=p,
+                                char_start=len(p.text) - 1, char_end=len(p.text),
+                                actual={"text": p.text},
+                                message="二级标题在换行分段时不应使用句号；如使用句号则应紧接正文",
+                            ))
+            elif t == "attachment_item_no_trailing_punct":
+                hits.extend(_attachment_item_hits(paras))
+        return CheckResult(passed=not hits, hits=hits)
+
+
+def _attachment_item_hits(paras):
+    """从 attachment_marker 起扫描附件区块，找末尾带标点的附件项。"""
+    hits: list[CheckHit] = []
+    in_attachment = False
+    for p in paras:
+        text = p.text.strip()
+        if not text:
+            continue
+
+        if p.role == "attachment_marker":
+            in_attachment = True
+            # 同段内可能出现 "附件：1.xxx；2.yyy。" 多项一行
+            for m in _ATTACH_ITEM_TRAIL_PUNCT.finditer(text):
+                hits.append(CheckHit(
+                    paragraph=p,
+                    char_start=m.start(), char_end=m.end(),
+                    actual={"snippet": m.group(0)},
+                    message=f'附件项末尾不应有标点："{m.group(0)}"',
+                ))
+            continue
+
+        if p.role in ("signature", "date", "heading_1"):
+            in_attachment = False
+            continue
+
+        if in_attachment and _ATTACH_ITEM_LINE.match(text):
+            hits.append(CheckHit(
+                paragraph=p,
+                char_start=len(p.text) - 1, char_end=len(p.text),
+                actual={"text": p.text},
+                message=f'附件项末尾不应有标点："{text}"',
+            ))
+    return hits

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/font.py

@@ -0,0 +1,162 @@
+"""字体/字号/复合样式/行距 check。"""
+
+import re
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckBase, CheckContext, CheckHit, CheckResult
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Paragraph, ParagraphStyle
+
+
+def _font_match(actual: str | None, expect: str) -> bool:
+    if not actual:
+        return False
+    return expect in actual or actual in expect
+
+
+def _size_match(actual: float | None, expect: float, tol: float = 0.5) -> bool:
+    if actual is None:
+        return False
+    return abs(actual - expect) <= tol
+
+
+def _style_matches(style: ParagraphStyle, expect: dict) -> bool:
+    if "eastasia" in expect and not _font_match(style.font_eastasia, expect["eastasia"]):
+        return False
+    if "size_pt" in expect and not _size_match(
+        style.font_size_pt, float(expect["size_pt"])
+    ):
+        return False
+    if "bold" in expect and bool(style.bold) != bool(expect["bold"]):
+        return False
+    return True
+
+
+@register("font")
+class FontCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        expect = ctx.stage.expect or {}
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            ok = True
+            actual = {
+                "font": p.style.font_eastasia,
+                "size": p.style.font_size_pt,
+            }
+            if "eastasia" in expect and not _font_match(p.style.font_eastasia, expect["eastasia"]):
+                ok = False
+            if "size_pt" in expect and not _size_match(
+                p.style.font_size_pt, float(expect["size_pt"])
+            ):
+                ok = False
+            if not ok:
+                hits.append(CheckHit(
+                    paragraph=p, char_start=0, char_end=len(p.text),
+                    actual=actual, expected=expect,
+                    message=f"字体或字号不符合（实际 {actual['font']} {actual['size']}pt，期望 {expect}）",
+                ))
+        return CheckResult(passed=not hits, hits=hits)
+
+
+@register("style_match")
+class StyleMatchCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        expect = ctx.stage.expect or {}
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            ok = True
+            actual = {
+                "font": p.style.font_eastasia,
+                "size": p.style.font_size_pt,
+                "bold": p.style.bold,
+                "italic": p.style.italic,
+                "alignment": p.style.alignment,
+            }
+            if "eastasia" in expect and not _font_match(p.style.font_eastasia, expect["eastasia"]):
+                ok = False
+            if "size_pt" in expect and not _size_match(
+                p.style.font_size_pt, float(expect["size_pt"])
+            ):
+                ok = False
+            if "bold" in expect and bool(p.style.bold) != bool(expect["bold"]):
+                ok = False
+            if "alignment" in expect and p.style.alignment != expect["alignment"]:
+                ok = False
+            if not ok:
+                hits.append(CheckHit(
+                    paragraph=p, char_start=0, char_end=len(p.text),
+                    actual=actual, expected=expect, message="样式不符合",
+                ))
+        return CheckResult(passed=not hits, hits=hits)
+
+
+_ATTACHMENT_MARKER_RE = re.compile(r"^\s*(附件[:：]|附件\d+)")
+
+
+@register("attachment_marker_style")
+class AttachmentMarkerStyleCheck(CheckBase):
+    """只校验“附件：”或“附件1”等标记本身，不校验后续附件名称。"""
+
+    DEFAULT_EXPECT = {"eastasia": "黑体", "size_pt": 16, "bold": False}
+
+    def run(self, ctx: CheckContext) -> CheckResult:
+        expect = ctx.stage.expect or self.DEFAULT_EXPECT
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            match = _ATTACHMENT_MARKER_RE.match(p.text)
+            if not match:
+                continue
+            marker_end = match.end(1)
+            marker_styles = _marker_run_styles(p, marker_end)
+            if not marker_styles:
+                marker_styles = [p.style]
+            bad_style = next(
+                (style for style in marker_styles if not _style_matches(style, expect)),
+                None,
+            )
+            if bad_style is not None:
+                hits.append(CheckHit(
+                    paragraph=p,
+                    char_start=match.start(1),
+                    char_end=marker_end,
+                    actual={
+                        "font": bad_style.font_eastasia,
+                        "size": bad_style.font_size_pt,
+                        "bold": bad_style.bold,
+                    },
+                    expected=expect,
+                    message="附件标记样式不符合",
+                ))
+        return CheckResult(passed=not hits, hits=hits)
+
+
+def _marker_run_styles(p: Paragraph, marker_end: int) -> list[ParagraphStyle]:
+    styles: list[ParagraphStyle] = []
+    cursor = 0
+    for run in p.runs:
+        run_start = cursor
+        run_end = cursor + len(run.text)
+        cursor = run_end
+        if run_end <= 0 or run_start >= marker_end:
+            continue
+        if run.text.strip():
+            styles.append(run.style)
+    return styles
+
+
+@register("line_spacing")
+class LineSpacingCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        expect = ctx.stage.expect or {}
+        target = float(expect.get("multiple", 1.5))
+        tol = float(expect.get("tol", 0.05))
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            actual = p.style.line_spacing
+            if actual is None or abs(actual - target) > tol:
+                hits.append(CheckHit(
+                    paragraph=p, char_start=0, char_end=len(p.text),
+                    actual={"line_spacing": actual},
+                    expected={"line_spacing": target},
+                    message=f"行距应为 {target}，实际 {actual}",
+                ))
+        return CheckResult(passed=not hits, hits=hits)

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/forbid.py

@@ -0,0 +1,42 @@
+"""短语/字符黑名单。"""
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckBase, CheckContext, CheckHit, CheckResult
+
+
+@register("forbid_phrase")
+class ForbidPhraseCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        phrases = ctx.stage.phrases or []
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            for phr in phrases:
+                start = p.text.find(phr)
+                if start >= 0:
+                    hits.append(CheckHit(
+                        paragraph=p, char_start=start, char_end=start + len(phr),
+                        actual={"text": phr}, expected={"forbid": phr},
+                        message=f"出现禁用短语 \"{phr}\"",
+                    ))
+        return CheckResult(passed=not hits, hits=hits)
+
+
+@register("forbid_chars")
+class ForbidCharsCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        chars = ctx.stage.chars or []
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            for c in chars:
+                start = 0
+                while True:
+                    idx = p.text.find(c, start)
+                    if idx < 0:
+                        break
+                    hits.append(CheckHit(
+                        paragraph=p, char_start=idx, char_end=idx + len(c),
+                        actual={"char": c}, expected={"forbid": c},
+                        message=f"禁用字符 \"{c}\" 出现在 idx {idx}",
+                    ))
+                    start = idx + len(c)
+        return CheckResult(passed=not hits, hits=hits)

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/hierarchy.py

@@ -0,0 +1,29 @@
+"""层级序号格式 check。"""
+
+import re
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckBase, CheckContext, CheckHit, CheckResult
+
+
+@register("hierarchy")
+class HierarchyCheck(CheckBase):
+    """检查层级序号格式：
+    - expected_order: 各级允许的模式（正向白名单，按 level 升序）
+    - forbid_patterns: 禁用模式（黑名单，命中即报错）
+    """
+
+    def run(self, ctx: CheckContext) -> CheckResult:
+        forbid = [re.compile(p) for p in (ctx.stage.forbid_patterns or [])]
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            text = p.text.strip()
+            for f in forbid:
+                m = f.search(text)
+                if m:
+                    hits.append(CheckHit(
+                        paragraph=p, char_start=m.start(), char_end=m.end(),
+                        actual={"text": m.group(0)},
+                        expected={"forbid_pattern": f.pattern},
+                        message=f"层级序号格式错误：命中禁用模式 {f.pattern}",
+                    ))
+        return CheckResult(passed=not hits, hits=hits)

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/punctuation.py

@@ -0,0 +1,46 @@
+"""标点符号专项规则。"""
+
+import re
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckBase, CheckContext, CheckHit, CheckResult
+
+
+# 多书名号或引号并列时不应用顿号分隔（中文/中文标点）
+_QUOTE_DUNHAO_RE = re.compile(r"([”》])、([“《])")
+# 句内括号末尾（除问号/叹号/省略号外）不应有标点
+_PAREN_PUNCT_RE = re.compile(r"[（(][^）)]*?[，。；：、][）)]")
+# 引号嵌套：双引号内含单引号包裹的强调短语（如 "卓'粤'创一流"）
+_NESTED_QUOTE_RE = re.compile(r"“[^“”]*?‘[^‘’]+’[^“”]*?”")
+
+
+@register("punctuation")
+class PunctuationCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        rules = ctx.stage.rules or []
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            for r in rules:
+                t = r.get("type")
+                if t == "no_dunhao_between_quotes":
+                    for m in _QUOTE_DUNHAO_RE.finditer(p.text):
+                        hits.append(CheckHit(
+                            paragraph=p, char_start=m.start(), char_end=m.end(),
+                            actual={"text": m.group(0)},
+                            expected={"text": m.group(0).replace("、", "")},
+                            message="多个引号/书名号并列不应用顿号分隔",
+                        ))
+                elif t == "no_punct_inside_inline_paren":
+                    for m in _PAREN_PUNCT_RE.finditer(p.text):
+                        hits.append(CheckHit(
+                            paragraph=p, char_start=m.start(), char_end=m.end(),
+                            actual={"text": m.group(0)},
+                            message="句内括号末尾通常不应含标点",
+                        ))
+                elif t == "no_outer_quote_when_inner_quote":
+                    for m in _NESTED_QUOTE_RE.finditer(p.text):
+                        hits.append(CheckHit(
+                            paragraph=p, char_start=m.start(), char_end=m.end(),
+                            actual={"text": m.group(0)},
+                            message="双引号内已含单引号强调时，外层不应再加双引号",
+                        ))
+        return CheckResult(passed=not hits, hits=hits)

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/regex_check.py

@@ -0,0 +1,36 @@
+"""regex_require / regex_forbid。"""
+
+import re
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckBase, CheckContext, CheckHit, CheckResult
+
+
+@register("regex_require")
+class RegexRequireCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        pat = re.compile(ctx.stage.pattern or "")
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            if not pat.search(p.text):
+                hits.append(CheckHit(
+                    paragraph=p, char_start=0, char_end=len(p.text),
+                    actual={"text": p.text}, expected={"pattern": ctx.stage.pattern},
+                    message=f"未匹配模式 {ctx.stage.pattern}",
+                ))
+        return CheckResult(passed=not hits, hits=hits)
+
+
+@register("regex_forbid")
+class RegexForbidCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        pat = re.compile(ctx.stage.pattern or "")
+        hits: list[CheckHit] = []
+        for p in ctx.paragraphs:
+            for m in pat.finditer(p.text):
+                hits.append(CheckHit(
+                    paragraph=p, char_start=m.start(), char_end=m.end(),
+                    actual={"text": m.group(0)},
+                    expected={"forbid_pattern": ctx.stage.pattern},
+                    message=f"出现禁止模式 {ctx.stage.pattern}（命中 \"{m.group(0)}\"）",
+                ))
+        return CheckResult(passed=not hits, hits=hits)

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/required.py

@@ -0,0 +1,28 @@
+"""required check：目标实体或选中段落必须有非空文本。"""
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import (
+    CheckBase, CheckContext, CheckHit, CheckResult,
+)
+
+
+@register("required")
+class RequiredCheck(CheckBase):
+    def run(self, ctx: CheckContext) -> CheckResult:
+        # target 通道：检查实体 text 是否非空
+        if ctx.target is not None:
+            if ctx.target.text and ctx.target.text.strip():
+                return CheckResult(passed=True, hits=[])
+            anchor = ctx.paragraphs[0] if ctx.paragraphs else None
+            return CheckResult(passed=False, hits=[
+                CheckHit(paragraph=anchor, message=f"实体 {ctx.target.name} 缺失或为空")
+            ])
+
+        # applies_to 通道：所有段落必须非空
+        empty = [p for p in ctx.paragraphs if not p.text.strip()]
+        if empty:
+            return CheckResult(
+                passed=False,
+                hits=[CheckHit(paragraph=p, message="段落为空") for p in empty],
+            )
+        return CheckResult(passed=True, hits=[])

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/checks/wenzhong.py

@@ -0,0 +1,42 @@
+"""文种白名单（15 种法定公文文种）。"""
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import (
+    CheckBase, CheckContext, CheckHit, CheckResult,
+)
+
+
+LEGAL_WENZHONG = {
+    "决议", "决定", "命令", "令", "公报", "公告", "通告",
+    "意见", "通知", "通报", "报告", "请示", "批复",
+    "议案", "函", "纪要",
+}
+
+
+@register("wenzhong_whitelist")
+class WenzhongWhitelistCheck(CheckBase):
+    """检查文种是否在 15 种法定文种白名单内。
+
+    数据来源：
+    1. ctx.entities["wenzhong"].text   ← 推荐
+    2. ctx.target.text  （当 rule.target = wenzhong 时）
+    """
+
+    def run(self, ctx: CheckContext) -> CheckResult:
+        wz = ""
+        wz_entity = ctx.entities.get("wenzhong") if ctx.entities else None
+        if wz_entity is not None:
+            wz = (wz_entity.text or "").strip()
+        elif ctx.target is not None and ctx.target.name == "wenzhong":
+            wz = (ctx.target.text or "").strip()
+
+        if not wz:
+            return CheckResult(passed=True, hits=[])
+        if wz in LEGAL_WENZHONG:
+            return CheckResult(passed=True, hits=[])
+        return CheckResult(passed=False, hits=[CheckHit(
+            paragraph=None,
+            actual={"wenzhong": wz},
+            expected={"wenzhong_whitelist": sorted(LEGAL_WENZHONG)},
+            message=f"非法定文种 \"{wz}\"，应为 15 种法定公文文种之一",
+        )])

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/result.py

@@ -0,0 +1,81 @@
+"""审查结果数据结构。"""
+
+from __future__ import annotations
+from collections import Counter
+from typing import Literal
+from pydantic import BaseModel, Field
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Finding
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.entities import SemanticEntity
+
+
+class CheckedRule(BaseModel):
+    rule_id: str
+    name: str
+    severity: str
+    category: str
+    status: Literal["pass", "fail", "skipped"]
+    skip_reason: str = ""
+
+
+class StructureItem(BaseModel):
+    """文档结构里一种 role 的统计。"""
+
+    role: str
+    label: str
+    count: int
+    expected: bool
+    paragraph_indices: list[int] = Field(default_factory=list)
+    samples: list[str] = Field(default_factory=list)
+    char_total: int = 0
+    dominant_font: str | None = None
+    dominant_size_pt: float | None = None
+    style_uniform: bool = True
+
+
+class OutlineNode(BaseModel):
+    """大纲节点（heading_1~4 的层级树）。"""
+
+    paragraph_index: int
+    level: int
+    text: str
+    children: list["OutlineNode"] = Field(default_factory=list)
+
+
+class AuditSummary(BaseModel):
+    score: int = 100
+    total_findings: int = 0
+    by_severity: dict[str, int] = Field(default_factory=dict)
+    by_category: dict[str, int] = Field(default_factory=dict)
+    passed_count: int = 0
+    failed_count: int = 0
+    skipped_count: int = 0
+
+
+class AuditResult(BaseModel):
+    audit_id: str
+    document: dict = Field(default_factory=dict)
+    summary: AuditSummary = Field(default_factory=AuditSummary)
+    findings: list[Finding] = Field(default_factory=list)
+    checked_rules: list[CheckedRule] = Field(default_factory=list)
+    structure: list[StructureItem] = Field(default_factory=list)
+    outline: list[OutlineNode] = Field(default_factory=list)
+    entities: dict[str, SemanticEntity | None] = Field(default_factory=dict)
+
+    def compute_summary(self) -> None:
+        sev_count = Counter(f.severity for f in self.findings)
+        cat_count = Counter(f.category for f in self.findings)
+        score = 100
+        score -= 10 * sev_count.get("error", 0)
+        score -= 3 * sev_count.get("warning", 0)
+        passed = sum(1 for r in self.checked_rules if r.status == "pass")
+        failed = sum(1 for r in self.checked_rules if r.status == "fail")
+        skipped = sum(1 for r in self.checked_rules if r.status == "skipped")
+        self.summary = AuditSummary(
+            score=max(0, score),
+            total_findings=len(self.findings),
+            by_severity=dict(sev_count),
+            by_category=dict(cat_count),
+            passed_count=passed,
+            failed_count=failed,
+            skipped_count=skipped,
+        )

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/runner.py

@@ -0,0 +1,242 @@
+"""规则评估引擎：跑一条规则的多 stage。"""
+
+from __future__ import annotations
+import asyncio
+import uuid
+from dataclasses import dataclass, field
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document, Finding, Location
+from fastapi_modules.fastapi_leaudit.govdoc_engine.dsl.schema import Rule
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import get_check  # noqa: F401  (确保注册)
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.base import CheckContext, CheckResult, CheckHit
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks.ai_check import AiCheck
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.selector import select_paragraphs
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.entities import SemanticEntity
+from fastapi_modules.fastapi_leaudit.govdoc_engine.llm.client import LlmClient
+
+# 触发所有 check 类的 @register
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import required as _r  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import font as _f  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import regex_check as _rc  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import confused_pair as _cp  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import forbid as _fb  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import wenzhong as _wz  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import hierarchy as _h  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import punctuation as _p  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import cross_role as _cr  # noqa: F401
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.checks import ai_check as _ai  # noqa: F401
+
+
+@dataclass
+class RuleOutcome:
+    """单条规则的执行结果（含 skipped 状态）。"""
+
+    rule: Rule
+    findings: list[Finding] = field(default_factory=list)
+    skipped: bool = False
+    skip_reason: str = ""
+
+
+class RuleRunner:
+    def __init__(self, llm_client: LlmClient | None = None):
+        self.llm = llm_client
+
+    # -- 上下文装配 -----------------------------------------------------
+    def _resolve_target(
+        self,
+        rule: Rule,
+        doc: Document,
+        entities: dict[str, SemanticEntity | None],
+    ) -> tuple[list, SemanticEntity | None, RuleOutcome | None]:
+        """根据 rule.target 或 rule.applies_to 选段落。
+
+        返回 (paragraphs, target_entity, early_outcome)；
+        若 early_outcome 非 None，调用方应直接返回（命中 on_missing 提前结束）。
+        """
+        if rule.target:
+            target_entity = entities.get(rule.target)
+            if target_entity is None:
+                return [], None, self._handle_missing(rule)
+            paragraphs = [
+                doc.paragraphs[i]
+                for i in target_entity.paragraph_indices
+                if 0 <= i < len(doc.paragraphs)
+            ]
+            return paragraphs, target_entity, None
+        # applies_to 通道（多段扫描）
+        return select_paragraphs(doc, rule.applies_to), None, None
+
+    def _handle_missing(self, rule: Rule) -> RuleOutcome:
+        mode = rule.on_missing
+        if mode == "pass":
+            return RuleOutcome(rule=rule)
+        reason = f"目标实体「{rule.target}」未识别到"
+        if mode == "skip":
+            return RuleOutcome(rule=rule, skipped=True, skip_reason=reason)
+        severity = "error" if mode == "fail" else "warning"
+        finding = Finding(
+            finding_id=f"F-{uuid.uuid4().hex[:8]}",
+            rule_id=rule.rule_id,
+            rule_name=rule.name,
+            severity=severity,
+            category=rule.category,
+            location=Location(paragraph_index=-1),
+            message=reason,
+            suggestion=rule.messages.fail or "",
+            evidence="", confidence=0.9,
+        )
+        return RuleOutcome(rule=rule, findings=[finding])
+
+    @staticmethod
+    def _merge_skip(outcome: RuleOutcome, result: CheckResult) -> None:
+        if not outcome.skip_reason:
+            outcome.skip_reason = result.skip_reason or "stage skipped"
+        outcome.skipped = True
+
+    # -- 同步路径 -------------------------------------------------------
+    def run_rule(
+        self,
+        rule: Rule,
+        doc: Document,
+        entities: dict[str, SemanticEntity | None] | None = None,
+    ) -> RuleOutcome:
+        entities = entities or {}
+        paragraphs, target, early = self._resolve_target(rule, doc, entities)
+        if early is not None:
+            return early
+
+        outcome = RuleOutcome(rule=rule)
+        for stage in rule.stages:
+            if stage.check == "ai":
+                check = AiCheck(llm_client=self.llm)
+            else:
+                check_cls = get_check(stage.check)
+                check = check_cls()
+
+            ctx = CheckContext(
+                document=doc,
+                paragraphs=paragraphs,
+                stage=stage,
+                entities=entities,
+                target=target,
+                rule_id=rule.rule_id,
+            )
+            result: CheckResult = check.run(ctx)
+            if result.skipped:
+                self._merge_skip(outcome, result)
+                continue
+            if not result.passed:
+                outcome.findings = [self._hit_to_finding(rule, h) for h in result.hits]
+                outcome.skipped = False
+                outcome.skip_reason = ""
+                return outcome
+        return outcome
+
+    def run_all(
+        self,
+        rules: list[Rule],
+        doc: Document,
+        entities: dict[str, SemanticEntity | None] | None = None,
+    ) -> list[Finding]:
+        flat, _ = self.evaluate(rules, doc, entities)
+        return flat
+
+    def evaluate(
+        self,
+        rules: list[Rule],
+        doc: Document,
+        entities: dict[str, SemanticEntity | None] | None = None,
+    ) -> tuple[list[Finding], list[RuleOutcome]]:
+        flat: list[Finding] = []
+        outcomes: list[RuleOutcome] = []
+        for r in rules:
+            o = self.run_rule(r, doc, entities)
+            flat.extend(o.findings)
+            outcomes.append(o)
+        return flat, outcomes
+
+    # -- 异步路径 -------------------------------------------------------
+    async def run_rule_async(
+        self,
+        rule: Rule,
+        doc: Document,
+        entities: dict[str, SemanticEntity | None] | None = None,
+    ) -> RuleOutcome:
+        entities = entities or {}
+        paragraphs, target, early = self._resolve_target(rule, doc, entities)
+        if early is not None:
+            return early
+
+        outcome = RuleOutcome(rule=rule)
+        for stage in rule.stages:
+            ctx = CheckContext(
+                document=doc,
+                paragraphs=paragraphs,
+                stage=stage,
+                entities=entities,
+                target=target,
+                rule_id=rule.rule_id,
+            )
+            if stage.check == "ai":
+                result = await AiCheck(llm_client=self.llm).run_async(ctx)
+            else:
+                check_cls = get_check(stage.check)
+                result = check_cls().run(ctx)
+            if result.skipped:
+                self._merge_skip(outcome, result)
+                continue
+            if not result.passed:
+                outcome.findings = [self._hit_to_finding(rule, h) for h in result.hits]
+                outcome.skipped = False
+                outcome.skip_reason = ""
+                return outcome
+        return outcome
+
+    async def run_all_async(
+        self,
+        rules: list[Rule],
+        doc: Document,
+        entities: dict[str, SemanticEntity | None] | None = None,
+    ) -> list[Finding]:
+        flat, _ = await self.evaluate_async(rules, doc, entities)
+        return flat
+
+    async def evaluate_async(
+        self,
+        rules: list[Rule],
+        doc: Document,
+        entities: dict[str, SemanticEntity | None] | None = None,
+    ) -> tuple[list[Finding], list[RuleOutcome]]:
+        outcomes_list = await asyncio.gather(
+            *(self.run_rule_async(r, doc, entities) for r in rules)
+        )
+        flat: list[Finding] = []
+        outcomes: list[RuleOutcome] = []
+        for o in outcomes_list:
+            flat.extend(o.findings)
+            outcomes.append(o)
+        return flat, outcomes
+
+    def _hit_to_finding(self, rule: Rule, hit: CheckHit) -> Finding:
+        para = hit.paragraph
+        loc = Location(
+            paragraph_index=para.index if para else -1,
+            role=para.role if para else None,
+            char_start=hit.char_start,
+            char_end=hit.char_end,
+            context=para.text if para else "",
+        )
+        msg = hit.message or rule.messages.fail
+        return Finding(
+            finding_id=f"F-{uuid.uuid4().hex[:8]}",
+            rule_id=rule.rule_id,
+            rule_name=rule.name,
+            severity=rule.severity,
+            category=rule.category,
+            location=loc,
+            actual=hit.actual or {},
+            expected=hit.expected or {},
+            message=msg,
+            suggestion=rule.messages.fail or "",
+            evidence=rule.messages.fail or "",
+            confidence=hit.confidence,
+        )

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/selector.py

@@ -0,0 +1,27 @@
+"""applies_to → 段落集合。"""
+
+from __future__ import annotations
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document, Paragraph
+from fastapi_modules.fastapi_leaudit.govdoc_engine.dsl.schema import AppliesTo
+
+
+def select_paragraphs(doc: Document, applies_to: AppliesTo) -> list[Paragraph]:
+    if applies_to.paragraph_index is not None:
+        idx = applies_to.paragraph_index
+        if 0 <= idx < len(doc.paragraphs):
+            return [doc.paragraphs[idx]]
+        return []
+
+    if applies_to.role == "any":
+        return list(doc.paragraphs)
+
+    targets: set[str] = set()
+    if applies_to.role:
+        targets.add(applies_to.role)
+    if applies_to.roles:
+        targets.update(applies_to.roles)
+
+    if not targets:
+        return list(doc.paragraphs)
+
+    return [p for p in doc.paragraphs if p.role in targets]

fastapi_modules/fastapi_leaudit/govdoc_engine/engine/structure.py

@@ -0,0 +1,93 @@
+"""从 Document 派生出 structure（按 role 分类统计）+ outline（heading 层级树）。"""
+
+from __future__ import annotations
+from collections import Counter
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document, Paragraph
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.result import OutlineNode, StructureItem
+
+
+_ROLE_LABELS: list[tuple[str, str, bool]] = [
+    # (role, 中文标签, 是否常规公文必备)
+    ("title", "标题", True),
+    ("doc_number", "发文字号", True),
+    ("recipient", "主送机关", True),
+    ("heading_1", "一级标题", False),
+    ("heading_2", "二级标题", False),
+    ("heading_3", "三级标题", False),
+    ("heading_4", "四级标题", False),
+    ("body", "正文", True),
+    ("attachment_marker", "附件标记", False),
+    ("attachment_title", "附件标题", False),
+    ("signature", "署名", True),
+    ("date", "成文日期", True),
+    ("no_text_marker", "(此页无正文)", False),
+    ("unknown", "未识别", False),
+]
+
+_HEADING_LEVELS = {
+    "heading_1": 1,
+    "heading_2": 2,
+    "heading_3": 3,
+    "heading_4": 4,
+}
+
+
+def _dominant_style(paragraphs: list[Paragraph]) -> tuple[str | None, float | None, bool]:
+    """返回 (字体众数, 字号众数, 是否所有段落样式一致)。"""
+    if not paragraphs:
+        return None, None, True
+    fonts = Counter(p.style.font_eastasia for p in paragraphs if p.style.font_eastasia)
+    sizes = Counter(p.style.font_size_pt for p in paragraphs if p.style.font_size_pt is not None)
+    dom_font = fonts.most_common(1)[0][0] if fonts else None
+    dom_size = sizes.most_common(1)[0][0] if sizes else None
+    uniform = len(fonts) <= 1 and len(sizes) <= 1
+    return dom_font, dom_size, uniform
+
+
+def build_structure(doc: Document) -> list[StructureItem]:
+    items: list[StructureItem] = []
+    for role, label, expected in _ROLE_LABELS:
+        paragraphs = [p for p in doc.paragraphs if p.role == role]
+        if not paragraphs and not expected:
+            # 非必备 role 没出现就不展示，保持面板紧凑
+            continue
+        samples = [p.text[:60] for p in paragraphs[:3]]
+        font, size, uniform = _dominant_style(paragraphs)
+        items.append(StructureItem(
+            role=role,
+            label=label,
+            count=len(paragraphs),
+            expected=expected,
+            paragraph_indices=[p.index for p in paragraphs],
+            samples=samples,
+            char_total=sum(len(p.text) for p in paragraphs),
+            dominant_font=font,
+            dominant_size_pt=size,
+            style_uniform=uniform,
+        ))
+    return items
+
+
+def build_outline(doc: Document) -> list[OutlineNode]:
+    """按段落顺序 + heading 层级生成树。"""
+    headings = [
+        (p.index, _HEADING_LEVELS[p.role], p.text)
+        for p in doc.paragraphs
+        if p.role in _HEADING_LEVELS
+    ]
+    if not headings:
+        return []
+
+    roots: list[OutlineNode] = []
+    stack: list[OutlineNode] = []
+    for idx, level, text in headings:
+        node = OutlineNode(paragraph_index=idx, level=level, text=text)
+        # 弹出比当前 level 更深的祖先
+        while stack and stack[-1].level >= level:
+            stack.pop()
+        if stack:
+            stack[-1].children.append(node)
+        else:
+            roots.append(node)
+        stack.append(node)
+    return roots

fastapi_modules/fastapi_leaudit/govdoc_engine/llm/cache.py

@@ -0,0 +1,101 @@
+"""LLM 响应缓存（SQLite）。
+
+缓存键 = sha256(model + canonical_json(messages, temperature, top_p, max_tokens))。
+仅缓存成功返回的文本；JSON 解析失败、API 错误、超时一律不入库。
+"""
+
+from __future__ import annotations
+import hashlib
+import json
+import logging
+import sqlite3
+import time
+from pathlib import Path
+from threading import Lock
+from typing import Any
+
+_log = logging.getLogger(__name__)
+
+_SCHEMA = """
+CREATE TABLE IF NOT EXISTS llm_cache (
+    cache_key TEXT PRIMARY KEY,
+    model TEXT NOT NULL,
+    response_text TEXT NOT NULL,
+    created_at REAL NOT NULL,
+    hit_count INTEGER NOT NULL DEFAULT 0,
+    last_hit_at REAL
+);
+CREATE INDEX IF NOT EXISTS idx_llm_cache_created ON llm_cache(created_at);
+"""
+
+# 影响响应的关键参数。其他 kwargs 不入 hash（如 stream/timeout）。
+_KEY_PARAMS = ("temperature", "top_p", "max_tokens", "response_format")
+
+
+def _canonical(messages: list[dict[str, str]], **kwargs: Any) -> str:
+    payload = {
+        "messages": messages,
+        "params": {k: kwargs.get(k) for k in _KEY_PARAMS},
+    }
+    return json.dumps(payload, sort_keys=True, ensure_ascii=False)
+
+
+def make_key(model: str, messages: list[dict[str, str]], **kwargs: Any) -> str:
+    raw = model + "\x00" + _canonical(messages, **kwargs)
+    return hashlib.sha256(raw.encode("utf-8")).hexdigest()
+
+
+class LlmCache:
+    def __init__(self, path: str | Path):
+        self.path = Path(path)
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        self._lock = Lock()
+        self._conn = sqlite3.connect(str(self.path), check_same_thread=False)
+        self._conn.executescript(_SCHEMA)
+        self._conn.commit()
+
+    def get(self, key: str) -> str | None:
+        with self._lock:
+            row = self._conn.execute(
+                "SELECT response_text FROM llm_cache WHERE cache_key = ?",
+                (key,),
+            ).fetchone()
+            if row is None:
+                return None
+            self._conn.execute(
+                "UPDATE llm_cache "
+                "SET hit_count = hit_count + 1, last_hit_at = ? "
+                "WHERE cache_key = ?",
+                (time.time(), key),
+            )
+            self._conn.commit()
+            return row[0]
+
+    def put(self, key: str, model: str, response_text: str) -> None:
+        if not response_text:
+            return
+        with self._lock:
+            self._conn.execute(
+                "INSERT OR IGNORE INTO llm_cache "
+                "(cache_key, model, response_text, created_at) "
+                "VALUES (?, ?, ?, ?)",
+                (key, model, response_text, time.time()),
+            )
+            self._conn.commit()
+
+    def stats(self) -> dict[str, int]:
+        with self._lock:
+            row = self._conn.execute(
+                "SELECT COUNT(*), COALESCE(SUM(hit_count), 0) FROM llm_cache"
+            ).fetchone()
+            return {"entries": int(row[0] or 0), "total_hits": int(row[1] or 0)}
+
+    def clear(self) -> int:
+        with self._lock:
+            cur = self._conn.execute("DELETE FROM llm_cache")
+            self._conn.commit()
+            return cur.rowcount
+
+    def close(self) -> None:
+        with self._lock:
+            self._conn.close()

fastapi_modules/fastapi_leaudit/govdoc_engine/llm/client.py

@@ -0,0 +1,258 @@
+"""Qwen LLM 客户端（OpenAI 兼容协议）。
+
+包含：超时（asyncio.wait_for）、重试（指数退避）、并发上限（Semaphore）。
+"""
+
+from __future__ import annotations
+import asyncio
+import json
+import logging
+import re
+import time
+from typing import Any
+
+from openai import AsyncOpenAI, OpenAI, APIError, APIConnectionError, RateLimitError
+
+from fastapi_admin.config import (
+    LLM_API_KEY,
+    LLM_BASE_URL,
+    LLM_MODEL,
+    LEAUDIT_LLM_MAX_CONCURRENCY,
+    LEAUDIT_LLM_REQUEST_TIMEOUT,
+    LEAUDIT_LLM_RETRY_MAX_ATTEMPTS,
+    LEAUDIT_LLM_RETRY_BACKOFF_BASE_SECONDS,
+)
+from fastapi_modules.fastapi_leaudit.govdoc_engine.llm.cache import LlmCache, make_key
+
+_log = logging.getLogger(__name__)
+
+
+_FENCE_RE = re.compile(r"```(?:json)?\s*([\s\S]+?)\s*```", re.MULTILINE)
+
+# 这些异常会触发重试；JSON 解析错误等业务错误不重试
+_RETRYABLE = (
+    asyncio.TimeoutError,
+    TimeoutError,
+    APIConnectionError,
+    RateLimitError,
+)
+
+
+class LlmJsonError(Exception):
+    """LLM 返回内容无法解析为 JSON。"""
+
+
+class LlmConfigError(Exception):
+    """LLM 客户端缺少必要配置。"""
+
+
+def _parse_json_text(text: str) -> dict[str, Any]:
+    text = text.strip()
+    m = _FENCE_RE.search(text)
+    if m:
+        text = m.group(1)
+    try:
+        return json.loads(text)
+    except json.JSONDecodeError:
+        start = text.find("{")
+        end = text.rfind("}")
+        if start >= 0 and end > start:
+            try:
+                return json.loads(text[start : end + 1])
+            except json.JSONDecodeError as e:
+                raise LlmJsonError(f"failed to parse LLM JSON: {text!r}") from e
+        raise LlmJsonError(f"LLM returned non-JSON content: {text!r}")
+
+
+def _is_retryable_status(exc: Exception) -> bool:
+    """APIError 中只重试 5xx 与 429。"""
+    if isinstance(exc, RateLimitError):
+        return True
+    if isinstance(exc, APIError):
+        status = getattr(exc, "status_code", None)
+        return status is not None and (status >= 500 or status == 429)
+    return False
+
+
+def _clip_text(value: Any, limit: int = 400) -> str:
+    text = str(value).strip()
+    if len(text) <= limit:
+        return text
+    return text[: limit - 3] + "..."
+
+
+def _format_exc(exc: Exception) -> str:
+    text = str(exc).strip()
+    parts = [exc.__class__.__name__]
+    if text:
+        parts.append(text)
+
+    status = getattr(exc, "status_code", None)
+    if status is not None:
+        parts.append(f"status={status}")
+
+    body = getattr(exc, "body", None)
+    if body not in (None, "", b""):
+        parts.append(f"body={_clip_text(body)}")
+
+    response = getattr(exc, "response", None)
+    if response is not None:
+        try:
+            request = getattr(response, "request", None)
+            if request is not None and getattr(request, "url", None):
+                parts.append(f"url={request.url}")
+        except Exception:
+            pass
+
+    request = getattr(exc, "request", None)
+    if request is not None and getattr(request, "url", None):
+        parts.append(f"url={request.url}")
+
+    return ": ".join(parts[:2]) + ("" if len(parts) <= 2 else " | " + " | ".join(parts[2:]))
+
+
+class LlmClient:
+    def __init__(
+        self,
+        api_key: str | None = None,
+        base_url: str | None = None,
+        model: str | None = None,
+        max_concurrency: int | None = None,
+        timeout_seconds: float | None = None,
+        max_retries: int | None = None,
+        cache: LlmCache | None = None,
+        cache_enabled: bool | None = None,
+    ):
+        key = api_key or LLM_API_KEY
+        self._misconfigured_error: LlmConfigError | None = None
+        if not key:
+            self._client = None
+            self._aclient = None
+            self._misconfigured_error = LlmConfigError(
+                "LLM_API_KEY is not configured. Set LLM_API_KEY in platform config."
+            )
+        else:
+            self._client = OpenAI(api_key=key, base_url=base_url or LLM_BASE_URL)
+            self._aclient = AsyncOpenAI(api_key=key, base_url=base_url or LLM_BASE_URL)
+        self.model = model or LLM_MODEL
+        self.timeout = timeout_seconds if timeout_seconds is not None else LEAUDIT_LLM_REQUEST_TIMEOUT
+        self.max_retries = max_retries if max_retries is not None else LEAUDIT_LLM_RETRY_MAX_ATTEMPTS
+        conc = max_concurrency if max_concurrency is not None else LEAUDIT_LLM_MAX_CONCURRENCY
+        self._sem = asyncio.Semaphore(conc)
+        # 缓存：cache 显式传入则用之；否则默认关闭。
+        if cache is not None:
+            self.cache: LlmCache | None = cache
+        elif cache_enabled is not False:
+            self.cache = None
+        else:
+            self.cache = None
+
+    def _ensure_ready(self) -> None:
+        if self._misconfigured_error is not None:
+            raise self._misconfigured_error
+
+    @staticmethod
+    def _prompt_text(messages: list[dict[str, str]]) -> str:
+        return "\n\n".join(
+            f"[{m.get('role', 'user')}]\n{m.get('content', '')}"
+            for m in messages
+        )
+
+    # -- 同步路径 -------------------------------------------------------
+    def chat(self, messages: list[dict[str, str]], **kwargs) -> str:
+        self._ensure_ready()
+        use_cache = kwargs.pop("use_cache", True)
+        label = kwargs.pop("label", "llm_call")
+        cache_kwargs = {k: kwargs.get(k) for k in ("temperature", "top_p", "max_tokens", "response_format")}
+        cache_key: str | None = None
+        prompt_text = self._prompt_text(messages)
+        t0 = time.monotonic()
+
+        if use_cache and self.cache is not None:
+            cache_key = make_key(self.model, messages, **cache_kwargs)
+            hit = self.cache.get(cache_key)
+            if hit is not None:
+                _log.debug("LLM cache HIT key=%s", cache_key[:12])
+                return hit
+
+        kwargs.setdefault("timeout", self.timeout)
+        last_exc: Exception | None = None
+        for attempt in range(self.max_retries + 1):
+            try:
+                resp = self._client.chat.completions.create(
+                    model=self.model, messages=messages, **kwargs
+                )
+                content = resp.choices[0].message.content or ""
+                if cache_key is not None and content:
+                    self.cache.put(cache_key, self.model, content)
+                return content
+            except _RETRYABLE as e:
+                last_exc = e
+            except APIError as e:
+                if not _is_retryable_status(e):
+                    raise
+                last_exc = e
+            if attempt < self.max_retries:
+                wait = min(8.0, 2 ** attempt)
+                _log.warning(
+                    "LLM call failed (%s); retry %d/%d after %.1fs",
+                    _format_exc(last_exc), attempt + 1, self.max_retries, wait,
+                )
+                time.sleep(wait)
+        assert last_exc is not None
+        raise last_exc
+
+    def chat_json(self, messages: list[dict[str, str]], **kwargs) -> dict[str, Any]:
+        return _parse_json_text(self.chat(messages, **kwargs))
+
+    # -- 异步路径 -------------------------------------------------------
+    async def chat_async(self, messages: list[dict[str, str]], **kwargs) -> str:
+        self._ensure_ready()
+        use_cache = kwargs.pop("use_cache", True)
+        label = kwargs.pop("label", "llm_call")
+        cache_kwargs = {k: kwargs.get(k) for k in ("temperature", "top_p", "max_tokens", "response_format")}
+        cache_key: str | None = None
+        prompt_text = self._prompt_text(messages)
+        t0 = time.monotonic()
+
+        if use_cache and self.cache is not None:
+            cache_key = make_key(self.model, messages, **cache_kwargs)
+            hit = self.cache.get(cache_key)
+            if hit is not None:
+                _log.debug("LLM cache HIT key=%s", cache_key[:12])
+                return hit
+
+        last_exc: Exception | None = None
+        for attempt in range(self.max_retries + 1):
+            try:
+                async with self._sem:
+                    resp = await asyncio.wait_for(
+                        self._aclient.chat.completions.create(
+                            model=self.model, messages=messages, **kwargs,
+                        ),
+                        timeout=self.timeout,
+                    )
+                content = resp.choices[0].message.content or ""
+                if cache_key is not None and content:
+                    self.cache.put(cache_key, self.model, content)
+                return content
+            except _RETRYABLE as e:
+                last_exc = e
+            except APIError as e:
+                if not _is_retryable_status(e):
+                    raise
+                last_exc = e
+            if attempt < self.max_retries:
+                wait = min(8.0, 2 ** attempt)
+                _log.warning(
+                    "LLM async call failed (%s); retry %d/%d after %.1fs",
+                    _format_exc(last_exc), attempt + 1, self.max_retries, wait,
+                )
+                await asyncio.sleep(wait)
+        assert last_exc is not None
+        raise last_exc
+
+    async def chat_json_async(
+        self, messages: list[dict[str, str]], **kwargs
+    ) -> dict[str, Any]:
+        return _parse_json_text(await self.chat_async(messages, **kwargs))

fastapi_modules/fastapi_leaudit/govdoc_engine/models.py

@@ -0,0 +1,77 @@
+"""公文审查的核心数据模型。"""
+
+from __future__ import annotations
+from typing import Any, Literal
+from pydantic import BaseModel, Field
+
+
+Role = Literal[
+    "title", "doc_number", "recipient",
+    "heading_1", "heading_2", "heading_3", "heading_4",
+    "body", "attachment_marker", "attachment_title", "signature", "date",
+    "no_text_marker", "unknown", "any",
+]
+Severity = Literal["error", "warning", "info"]
+
+
+class ParagraphStyle(BaseModel):
+    font_eastasia: str | None = None
+    font_ascii: str | None = None
+    font_size_pt: float | None = None
+    bold: bool = False
+    italic: bool = False
+    line_spacing: float | None = None
+    line_spacing_rule: str | None = None
+    alignment: str = "left"
+    first_line_indent_pt: float = 0.0
+
+
+class Run(BaseModel):
+    text: str
+    style: ParagraphStyle
+
+
+class Paragraph(BaseModel):
+    index: int
+    text: str
+    runs: list[Run]
+    style: ParagraphStyle
+    role: Role | None = None
+    role_confidence: float = 1.0
+    in_table: bool = False
+    in_header: bool = False
+    in_footer: bool = False
+
+
+class Table(BaseModel):
+    index: int
+    rows: list[list[str]]
+
+
+class Document(BaseModel):
+    meta: dict[str, Any] = Field(default_factory=dict)
+    paragraphs: list[Paragraph]
+    tables: list[Table] = Field(default_factory=list)
+
+
+class Location(BaseModel):
+    paragraph_index: int
+    role: Role | None = None
+    char_start: int = 0
+    char_end: int = 0
+    context: str = ""
+
+
+class Finding(BaseModel):
+    finding_id: str
+    rule_id: str
+    rule_name: str
+    severity: Severity
+    category: str
+    location: Location
+    actual: dict[str, Any] = Field(default_factory=dict)
+    expected: dict[str, Any] = Field(default_factory=dict)
+    message: str
+    suggestion: str = ""
+    evidence: str = ""
+    confidence: float = 1.0

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/docx_parser.py

@@ -0,0 +1,152 @@
+"""解析 .docx → Document 对象。
+
+文档顺序遍历 body：顶级段落 + 表格内段落都纳入 paragraphs，
+后续 role tagging 与规则评估都能扫到表格内的内容。
+"""
+
+from __future__ import annotations
+from pathlib import Path
+from docx import Document as DocxDocument
+from docx.oxml.ns import qn
+from docx.text.paragraph import Paragraph as DocxParagraph
+from lxml import etree
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document, Paragraph, ParagraphStyle, Run, Table
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.style_resolver import StyleResolver
+
+
+_ALIGN_MAP = {0: "left", 1: "center", 2: "right", 3: "justify"}
+
+
+def _read_run_style(run, p_elem, resolver: StyleResolver) -> ParagraphStyle:
+    rs = resolver.resolve_run(p_elem, run._element)
+    return ParagraphStyle(
+        font_eastasia=rs.font_eastasia,
+        font_ascii=rs.font_ascii,
+        font_size_pt=rs.size_pt,
+        bold=bool(rs.bold) if rs.bold is not None else False,
+        italic=bool(rs.italic) if rs.italic is not None else False,
+    )
+
+
+def _read_paragraph_style(p, resolver: StyleResolver) -> ParagraphStyle:
+    pf = p.paragraph_format
+    alignment = (
+        _ALIGN_MAP.get(pf.alignment, "left") if pf.alignment is not None else "left"
+    )
+    spacing_pt = float(pf.line_spacing) if pf.line_spacing is not None else None
+    indent = pf.first_line_indent
+    indent_pt = float(indent.pt) if indent is not None else 0.0
+
+    if p.runs:
+        base = _read_run_style(p.runs[0], p._element, resolver)
+    else:
+        rs = resolver.resolve_paragraph(p._element)
+        base = ParagraphStyle(
+            font_eastasia=rs.font_eastasia,
+            font_ascii=rs.font_ascii,
+            font_size_pt=rs.size_pt,
+            bold=bool(rs.bold) if rs.bold is not None else False,
+            italic=bool(rs.italic) if rs.italic is not None else False,
+        )
+    base.alignment = alignment
+    base.line_spacing = spacing_pt
+    base.first_line_indent_pt = indent_pt
+    return base
+
+
+def _is_in_table(p_elem) -> bool:
+    parent = p_elem.getparent()
+    while parent is not None:
+        if etree.QName(parent).localname == "tbl":
+            return True
+        parent = parent.getparent()
+    return False
+
+
+def _iter_body_paragraphs(docx):
+    """文档顺序遍历 body 下所有 w:p（含表格内）。"""
+    for p_elem in docx.element.body.iter(qn("w:p")):
+        yield p_elem
+
+
+def _iter_header_footer_paragraphs(docx):
+    """yield (DocxParagraph, p_elem, in_header, in_footer)，跨 section 去重。"""
+    seen: set[int] = set()
+    for section in docx.sections:
+        targets = [
+            ("header", section.header),
+            ("first_header", section.first_page_header),
+            ("even_header", section.even_page_header),
+            ("footer", section.footer),
+            ("first_footer", section.first_page_footer),
+            ("even_footer", section.even_page_footer),
+        ]
+        for kind, hf in targets:
+            if hf is None:
+                continue
+            try:
+                if hf.is_linked_to_previous:
+                    continue
+            except Exception:
+                pass
+            in_header = "header" in kind
+            for p in hf.paragraphs:
+                pid = id(p._element)
+                if pid in seen:
+                    continue
+                seen.add(pid)
+                yield p, p._element, in_header, not in_header
+
+
+def parse_docx(path: str | Path) -> Document:
+    path = Path(path)
+    docx = DocxDocument(path)
+    resolver = StyleResolver(docx)
+
+    paragraphs: list[Paragraph] = []
+    idx = 0
+    # 1) body：含表格内段落
+    for p_elem in _iter_body_paragraphs(docx):
+        p = DocxParagraph(p_elem, docx.part)
+        runs = [
+            Run(text=r.text, style=_read_run_style(r, p_elem, resolver))
+            for r in p.runs
+        ]
+        style = _read_paragraph_style(p, resolver)
+        paragraphs.append(Paragraph(
+            index=idx,
+            text=p.text,
+            runs=runs,
+            style=style,
+            in_table=_is_in_table(p_elem),
+        ))
+        idx += 1
+    # 2) headers / footers：附在末尾，role tagger 也能扫到
+    for p, p_elem, in_header, in_footer in _iter_header_footer_paragraphs(docx):
+        runs = [
+            Run(text=r.text, style=_read_run_style(r, p_elem, resolver))
+            for r in p.runs
+        ]
+        style = _read_paragraph_style(p, resolver)
+        paragraphs.append(Paragraph(
+            index=idx,
+            text=p.text,
+            runs=runs,
+            style=style,
+            in_table=_is_in_table(p_elem),
+            in_header=in_header,
+            in_footer=in_footer,
+        ))
+        idx += 1
+
+    tables = []
+    for tidx, t in enumerate(docx.tables):
+        rows = [[cell.text for cell in row.cells] for row in t.rows]
+        tables.append(Table(index=tidx, rows=rows))
+
+    return Document(
+        meta={"path": str(path), "page_count": len(docx.sections)},
+        paragraphs=paragraphs,
+        tables=tables,
+    )

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/entities.py

@@ -0,0 +1,27 @@
+"""语义实体：把段落 + 字段值 + 样式合在一起。"""
+
+from __future__ import annotations
+from typing import Any, Literal
+from pydantic import BaseModel, Field
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import ParagraphStyle
+
+
+EntitySource = Literal["structural", "llm", "derived"]
+
+
+class SemanticEntity(BaseModel):
+    """公文中的一个语义单元（标题 / 发文字号 / 主送机关 / ...）。
+
+    - structural：name 与某个 role 一一对应，paragraph_indices 非空，style 可用。
+    - derived：从其他实体推导（如 wenzhong 从 title 末尾），paragraph_indices 借用源段落。
+    - llm：仅当结构 / 派生路径都失败时启用，paragraph_indices 可能为空。
+    """
+
+    name: str
+    text: str = ""
+    paragraph_indices: list[int] = Field(default_factory=list)
+    primary_role: str | None = None
+    style: ParagraphStyle | None = None
+    extra: dict[str, Any] = Field(default_factory=dict)
+    source: EntitySource = "structural"
+    confidence: float = 1.0

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/entity_builder.py

@@ -0,0 +1,195 @@
+"""从已 tag 的 Document 抽取语义实体（结构化优先）。"""
+
+from __future__ import annotations
+import re
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.entities import SemanticEntity
+
+
+# 8 个内置实体名（也用于 schema 校验冲突）
+BUILTIN_ENTITY_NAMES: frozenset[str] = frozenset({
+    "title", "doc_number", "recipient", "date",
+    "signature", "attachments", "wenzhong", "issuer",
+})
+
+
+# 内置实体的 LLM 兜底 prompt 描述（Phase B 使用）
+BUILTIN_LLM_DESCRIPTION: dict[str, str] = {
+    "title":       "公文主标题（不含发文字号）",
+    "doc_number":  "X发〔YYYY〕N号 形式的发文字号",
+    "recipient":   "公文抬头的接收机关名称",
+    "date":        "末尾的成文日期原文",
+    "signature":   "末尾的发文机关署名",
+    "attachments": "附件清单（数组，每项含 序号 与 名称）",
+    "wenzhong":    "公文文种（决议/决定/通知/通报/请示/批复 等 15 种之一）",
+    "issuer":      "发文机关全称",
+}
+
+
+# role → entity name 的 1:1 映射
+_ROLE_ENTITY_MAP = {
+    "title": "title",
+    "doc_number": "doc_number",
+    "recipient": "recipient",
+    "date": "date",
+    "signature": "signature",
+}
+
+
+_ATTACHMENT_HEAD_RE = re.compile(r"^附件\d*[:：]\s*")
+_ATTACHMENT_ITEM_RE = re.compile(r"^\s*(\d+)[\.．、)]\s*(.+)$")
+
+# 15 种法定文种（参照《党政机关公文处理工作条例》）
+_WENZHONG_LIST = (
+    "决议", "决定", "命令", "公报", "公告", "通告",
+    "意见", "通知", "通报", "报告", "请示", "批复",
+    "议案", "函", "纪要",
+)
+_WENZHONG_RE = re.compile("(" + "|".join(_WENZHONG_LIST) + ")$")
+
+# 「XX关于...的YY」 → issuer = XX
+_ISSUER_PREFIX_RE = re.compile(r"^(.+?)关于")
+
+
+class EntityBuilder:
+    """从已 tag 的 Document 抽取 8 个内置语义实体。"""
+
+    def build(self, doc: Document) -> dict[str, SemanticEntity | None]:
+        entities: dict[str, SemanticEntity | None] = {
+            name: None for name in BUILTIN_ENTITY_NAMES
+        }
+
+        # ① 一对一 role → entity
+        for role, name in _ROLE_ENTITY_MAP.items():
+            paras = [p for p in doc.paragraphs if p.role == role]
+            if not paras:
+                continue
+            target = paras[-1] if name == "signature" else paras[0]
+            entities[name] = SemanticEntity(
+                name=name,
+                text=target.text.strip(),
+                paragraph_indices=[target.index],
+                primary_role=role,
+                style=target.style,
+                source="structural",
+                confidence=target.role_confidence,
+            )
+
+        # ② attachments：attachment_marker + 跟随行
+        entities["attachments"] = self._build_attachments(doc)
+
+        # ③ 派生：wenzhong / issuer
+        title_e = entities.get("title")
+        if title_e:
+            entities["wenzhong"] = self._derive_wenzhong(title_e)
+            entities["issuer"] = self._derive_issuer(
+                title_e, entities.get("signature")
+            )
+        elif entities.get("signature"):
+            entities["issuer"] = self._derive_issuer(
+                None, entities["signature"]
+            )
+
+        return entities
+
+    # ---------- attachments ----------
+    def _build_attachments(self, doc: Document) -> SemanticEntity | None:
+        markers = [
+            i for i, p in enumerate(doc.paragraphs)
+            if p.role == "attachment_marker"
+        ]
+        if not markers:
+            return None
+
+        m = markers[0]
+        items: list[dict] = []
+        para_idxs: list[int] = [m]
+
+        first = doc.paragraphs[m].text.strip()
+        head = _ATTACHMENT_HEAD_RE.sub("", first)
+        if head:
+            mt = _ATTACHMENT_ITEM_RE.match(head)
+            if mt:
+                items.append(
+                    {"序号": int(mt.group(1)), "名称": mt.group(2).strip()}
+                )
+            else:
+                items.append({"序号": 1, "名称": head})
+
+        # 后续顺序行：直到遇到非 body / unknown 的段
+        for j in range(m + 1, len(doc.paragraphs)):
+            p = doc.paragraphs[j]
+            if p.role and p.role not in ("body", "unknown", "attachment_marker"):
+                break
+            t = p.text.strip()
+            if not t:
+                continue
+            mt = _ATTACHMENT_ITEM_RE.match(t)
+            if not mt:
+                break
+            items.append(
+                {"序号": int(mt.group(1)), "名称": mt.group(2).strip()}
+            )
+            para_idxs.append(p.index)
+
+        if not items:
+            return None
+
+        text = "; ".join(f"{it['序号']}. {it['名称']}" for it in items)
+        return SemanticEntity(
+            name="attachments",
+            text=text,
+            paragraph_indices=para_idxs,
+            primary_role="attachment_marker",
+            style=doc.paragraphs[m].style,
+            extra={"items": items},
+            source="structural",
+            confidence=0.9,
+        )
+
+    # ---------- 派生 ----------
+    def _derive_wenzhong(
+        self, title: SemanticEntity
+    ) -> SemanticEntity | None:
+        m = _WENZHONG_RE.search(title.text)
+        if not m:
+            return None
+        return SemanticEntity(
+            name="wenzhong",
+            text=m.group(1),
+            paragraph_indices=list(title.paragraph_indices),
+            primary_role="title",
+            extra={"derived_from": "title.suffix"},
+            source="derived",
+            confidence=0.95,
+        )
+
+    def _derive_issuer(
+        self,
+        title: SemanticEntity | None,
+        signature: SemanticEntity | None,
+    ) -> SemanticEntity | None:
+        if title:
+            m = _ISSUER_PREFIX_RE.match(title.text)
+            if m:
+                return SemanticEntity(
+                    name="issuer",
+                    text=m.group(1),
+                    paragraph_indices=list(title.paragraph_indices),
+                    primary_role="title",
+                    extra={"derived_from": "title.prefix"},
+                    source="derived",
+                    confidence=0.9,
+                )
+        if signature:
+            return SemanticEntity(
+                name="issuer",
+                text=signature.text,
+                paragraph_indices=list(signature.paragraph_indices),
+                primary_role="signature",
+                style=signature.style,
+                extra={"derived_from": "signature"},
+                source="derived",
+                confidence=0.8,
+            )
+        return None

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/extractor.py

@@ -0,0 +1,104 @@
+"""LLM 字段抽取：差量模式（仅对未知字段构造 prompt）。"""
+
+from __future__ import annotations
+import logging
+from typing import Any
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document
+from fastapi_modules.fastapi_leaudit.govdoc_engine.llm.client import LlmClient, _format_exc
+
+_log = logging.getLogger(__name__)
+
+
+_PROMPT_HEAD = """从下面的公文中抽取以下指定字段，仅以 JSON 输出。
+
+【公文内容（顺序段落）】
+{text}
+
+【需要抽取的字段】
+{spec_block}
+
+【输出格式】
+仅 JSON：{{{example}}}
+未识别的字段填 ""（list 类型填 []）。
+"""
+
+
+def _build_doc_text(doc: Document) -> str:
+    return "\n".join(f"[{p.index}] {p.text}" for p in doc.paragraphs)
+
+
+def _example_for(spec: dict[str, dict]) -> str:
+    parts = []
+    for name, meta in spec.items():
+        t = meta.get("type", "string")
+        if t == "list":
+            parts.append(f'"{name}": []')
+        else:
+            parts.append(f'"{name}": ""')
+    return ", ".join(parts)
+
+
+class FieldExtractor:
+    """LLM 差量字段抽取。
+
+    extract_missing(doc, spec): spec 指定需要抽哪些字段；空 spec 不调 LLM。
+    """
+
+    def __init__(self, llm_client: LlmClient):
+        self.client = llm_client
+
+    def _build_messages_for_spec(
+        self, doc: Document, spec: dict[str, dict]
+    ) -> list[dict[str, str]]:
+        spec_lines = [
+            f"- {name}: {meta.get('description', name)}"
+            f"（{meta.get('type', 'string')}）"
+            for name, meta in spec.items()
+        ]
+        prompt = _PROMPT_HEAD.format(
+            text=_build_doc_text(doc),
+            spec_block="\n".join(spec_lines) or "（无）",
+            example=_example_for(spec),
+        )
+        return [{"role": "user", "content": prompt}]
+
+    def _shape_missing(
+        self, spec: dict[str, dict], resp: dict
+    ) -> dict[str, Any]:
+        out: dict[str, Any] = {}
+        for name, meta in spec.items():
+            if meta.get("type") == "list":
+                out[name] = resp.get(name) or []
+            else:
+                out[name] = resp.get(name) or ""
+        return out
+
+    def extract_missing(
+        self, doc: Document | None, spec: dict[str, dict]
+    ) -> dict[str, Any]:
+        if not spec or doc is None:
+            return {}
+        label = "extract_missing__" + ",".join(spec.keys())
+        try:
+            resp = self.client.chat_json(
+                self._build_messages_for_spec(doc, spec), label=label,
+            )
+        except Exception as e:
+            _log.warning("Differential extraction failed: %s", _format_exc(e))
+            resp = {}
+        return self._shape_missing(spec, resp)
+
+    async def extract_missing_async(
+        self, doc: Document | None, spec: dict[str, dict]
+    ) -> dict[str, Any]:
+        if not spec or doc is None:
+            return {}
+        label = "extract_missing__" + ",".join(spec.keys())
+        try:
+            resp = await self.client.chat_json_async(
+                self._build_messages_for_spec(doc, spec), label=label,
+            )
+        except Exception as e:
+            _log.warning("Differential extraction failed: %s", _format_exc(e))
+            resp = {}
+        return self._shape_missing(spec, resp)

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/loader.py

@@ -0,0 +1,83 @@
+"""doc / wps → docx 转换。"""
+
+from __future__ import annotations
+
+import shutil
+import subprocess
+from pathlib import Path
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.config import get_settings
+
+
+class UnsupportedFormat(Exception):
+    pass
+
+
+class ConversionError(Exception):
+    pass
+
+
+_SUPPORTED_DIRECT = {".docx"}
+_SUPPORTED_CONVERT = {".doc", ".wps"}
+_SOFFICE_FALLBACK_PATHS = (
+    "/opt/homebrew/bin/soffice",
+    "/usr/local/bin/soffice",
+    "/Applications/LibreOffice.app/Contents/MacOS/soffice",
+    "/usr/bin/soffice",
+)
+
+
+def load_to_docx(src: Path) -> Path:
+    """统一返回 .docx 路径。.doc/.wps 调 soffice 转换。"""
+    ext = src.suffix.lower()
+    if ext in _SUPPORTED_DIRECT:
+        return src
+    if ext in _SUPPORTED_CONVERT:
+        return _convert_via_soffice(src)
+    raise UnsupportedFormat(f"unsupported file type: {ext}")
+
+
+def _convert_via_soffice(src: Path) -> Path:
+    soffice = _resolve_soffice_path(get_settings().soffice_path)
+
+    out_dir = src.parent
+    cmd = [
+        soffice, "--headless", "--convert-to", "docx",
+        "--outdir", str(out_dir), str(src),
+    ]
+    try:
+        result = subprocess.run(
+            cmd, capture_output=True, timeout=60,
+        )
+    except subprocess.TimeoutExpired as e:
+        raise ConversionError("soffice timeout") from e
+
+    if result.returncode != 0:
+        raise ConversionError(
+            f"soffice exit {result.returncode}: {result.stderr.decode(errors='ignore')}"
+        )
+
+    out = out_dir / (src.stem + ".docx")
+    if not out.exists():
+        raise ConversionError(f"expected output not found: {out}")
+    return out
+
+
+def _resolve_soffice_path(configured: str) -> str:
+    candidates = [configured, *_SOFFICE_FALLBACK_PATHS]
+    checked: list[str] = []
+    for candidate in candidates:
+        if candidate in checked:
+            continue
+        checked.append(candidate)
+
+        resolved = shutil.which(candidate)
+        if resolved:
+            return resolved
+        if Path(candidate).exists():
+            return candidate
+
+    raise ConversionError(
+        f"soffice not found; checked: {', '.join(checked)}. "
+        "Install LibreOffice or set SOFFICE_PATH."
+    )

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/role_tagger.py

@@ -0,0 +1,50 @@
+"""组合规则 tagger + LLM tagger 的总入口。"""
+
+from __future__ import annotations
+import asyncio
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.role_tagger_rule import RuleBasedTagger
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.role_tagger_llm import LlmTagger
+from fastapi_modules.fastapi_leaudit.govdoc_engine.llm.client import LlmClient
+
+
+class RoleTagger:
+    """两段式：先规则打标，置信度 < threshold 的段落送 LLM 兜底。"""
+
+    def __init__(
+        self,
+        llm_client: LlmClient | None = None,
+        threshold: float = 0.8,
+    ):
+        self.rule = RuleBasedTagger()
+        self.llm = LlmTagger(llm_client) if llm_client else None
+        self.threshold = threshold
+
+    def _low_conf_indices(self, doc: Document) -> list[int]:
+        return [
+            i for i, p in enumerate(doc.paragraphs)
+            if p.role_confidence < self.threshold
+        ]
+
+    def tag(self, doc: Document) -> None:
+        self.rule.tag(doc)
+        if self.llm is None:
+            return
+        for i in self._low_conf_indices(doc):
+            role, conf = self.llm.disambiguate(doc, i)
+            doc.paragraphs[i].role = role
+            doc.paragraphs[i].role_confidence = conf
+
+    async def tag_async(self, doc: Document) -> None:
+        self.rule.tag(doc)
+        if self.llm is None:
+            return
+        targets = self._low_conf_indices(doc)
+        if not targets:
+            return
+        results = await asyncio.gather(
+            *(self.llm.disambiguate_async(doc, i) for i in targets)
+        )
+        for i, (role, conf) in zip(targets, results):
+            doc.paragraphs[i].role = role
+            doc.paragraphs[i].role_confidence = conf

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/role_tagger_llm.py

@@ -0,0 +1,90 @@
+"""LLM 兜底打 role：对低置信段落做二次确认。"""
+
+from __future__ import annotations
+import logging
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document, Role
+from fastapi_modules.fastapi_leaudit.govdoc_engine.llm.client import LlmClient, _format_exc
+
+_log = logging.getLogger(__name__)
+
+
+VALID_ROLES = [
+    "title", "doc_number", "recipient",
+    "heading_1", "heading_2", "heading_3", "heading_4",
+    "body", "attachment_marker", "signature", "date",
+    "no_text_marker", "unknown",
+]
+
+
+_PROMPT = """你是公文格式专家。下面是一份公文的段落列表，请为指定的"待定段落"判断其角色。
+
+【全文段落（带索引和当前规则推测）】
+{context}
+
+【待定段落 idx={idx}】
+文本: {text}
+当前推测角色: {current_role}（置信度 {conf:.2f}）
+
+【角色取值范围】
+{roles}
+
+请综合公文结构判断该段落最可能的角色。
+
+仅以 JSON 输出：
+{{"role": "<角色>", "confidence": <0-1 浮点数>, "reason": "<简短理由>"}}
+"""
+
+
+class LlmTagger:
+    def __init__(self, client: LlmClient):
+        self.client = client
+
+    def _build_prompt(self, doc: Document, target_idx: int) -> tuple[str, "object"]:
+        ctx_lines = []
+        for p in doc.paragraphs:
+            tag = "← 待定" if p.index == target_idx else ""
+            ctx_lines.append(f"[{p.index}] role={p.role} text={p.text[:60]} {tag}")
+        ctx = "\n".join(ctx_lines)
+        target = doc.paragraphs[target_idx]
+        prompt = _PROMPT.format(
+            context=ctx,
+            idx=target_idx,
+            text=target.text,
+            current_role=target.role or "unknown",
+            conf=target.role_confidence,
+            roles=", ".join(VALID_ROLES),
+        )
+        return prompt, target
+
+    def _interpret(self, resp: dict, target) -> tuple[Role, float]:
+        role = resp.get("role", "unknown")
+        if role not in VALID_ROLES:
+            role = "unknown"
+        conf = float(resp.get("confidence", 0.5))
+        return role, conf  # type: ignore[return-value]
+
+    def disambiguate(self, doc: Document, target_idx: int) -> tuple[Role, float]:
+        prompt, target = self._build_prompt(doc, target_idx)
+        label = f"role_tag_p{target_idx}"
+        try:
+            resp = self.client.chat_json(
+                [{"role": "user", "content": prompt}], label=label,
+            )
+        except Exception as e:
+            _log.warning("Role disambiguation skipped (LLM error): %s", _format_exc(e))
+            return target.role or "unknown", target.role_confidence  # type: ignore[return-value]
+        return self._interpret(resp, target)
+
+    async def disambiguate_async(
+        self, doc: Document, target_idx: int
+    ) -> tuple[Role, float]:
+        prompt, target = self._build_prompt(doc, target_idx)
+        label = f"role_tag_p{target_idx}"
+        try:
+            resp = await self.client.chat_json_async(
+                [{"role": "user", "content": prompt}], label=label,
+            )
+        except Exception as e:
+            _log.warning("Role disambiguation skipped (LLM error): %s", _format_exc(e))
+            return target.role or "unknown", target.role_confidence  # type: ignore[return-value]
+        return self._interpret(resp, target)

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/role_tagger_rule.py

@@ -0,0 +1,132 @@
+"""基于位置 + 文字模式 + 字体样式的段落角色识别。"""
+
+from __future__ import annotations
+import re
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document, Paragraph, Role
+
+
+HEADING_1_RE = re.compile(r"^[一二三四五六七八九十百]+、")
+HEADING_2_RE = re.compile(r"^（[一二三四五六七八九十]+）")
+HEADING_3_RE = re.compile(r"^\d+[\.．]")
+HEADING_4_RE = re.compile(r"^（\d+）")
+DOC_NUMBER_RE = re.compile(r"[一-龥]+[〔\[]\d{4}[〕\]]第?\d+号")
+DATE_RE = re.compile(
+    r"^\d{4}年\d{1,2}月\d{1,2}日$"
+    r"|^[一二三四五六七八九十○〇零]+年[一二三四五六七八九十○〇零]+月[一二三四五六七八九十○〇零]+日$"
+)
+ATTACHMENT_RE = re.compile(r"^附件[:：1-9]")
+NO_TEXT_RE = re.compile(r"^[\(（]\s*此页无正文\s*[\)）]")
+RECIPIENT_TAIL_RE = re.compile(r"[：:]\s*$")
+RECIPIENT_HINTS = (
+    "局", "委", "府", "厅", "办", "公司", "各", "处室",
+    "委员会", "署", "院", "部", "司", "处",
+)
+RECIPIENT_BLOCKLIST = (
+    "现将", "兹", "经研究", "为做好", "为深入", "为进一步",
+    "根据", "如下", "汇报", "通知如下", "请示如下",
+)
+
+
+class RuleBasedTagger:
+    def tag(self, doc: Document) -> None:
+        n = len(doc.paragraphs)
+        for i, p in enumerate(doc.paragraphs):
+            role, conf = self._classify(p, i, n, doc)
+            p.role = role
+            p.role_confidence = conf
+
+    def _classify(
+        self, p: Paragraph, idx: int, total: int, doc: Document
+    ) -> tuple[Role, float]:
+        text = p.text.strip()
+
+        if not text:
+            return ("unknown", 0.5)
+
+        if NO_TEXT_RE.match(text):
+            return ("no_text_marker", 1.0)
+
+        if ATTACHMENT_RE.match(text):
+            return ("attachment_marker", 0.95)
+
+        if DATE_RE.match(text):
+            return ("date", 0.9)
+
+        if DOC_NUMBER_RE.search(text) and idx <= 5:
+            return ("doc_number", 0.95)
+
+        if idx == 0 or (
+            idx <= 2
+            and p.style.alignment == "center"
+            and (p.style.font_size_pt or 0) >= 18
+        ):
+            return ("title", 0.95)
+
+        font = (p.style.font_eastasia or "").strip()
+        size = p.style.font_size_pt or 0
+
+        if self._is_attachment_title(p, idx, doc):
+            return ("attachment_title", 0.9)
+
+        if HEADING_1_RE.match(text):
+            conf = 0.95 if "黑体" in font else 0.7
+            return ("heading_1", conf)
+
+        if HEADING_2_RE.match(text):
+            conf = 0.95 if "楷体" in font else 0.7
+            return ("heading_2", conf)
+
+        if HEADING_3_RE.match(text):
+            conf = 0.9 if "仿宋" in font else 0.65
+            return ("heading_3", conf)
+
+        if HEADING_4_RE.match(text):
+            return ("heading_4", 0.85)
+
+        if (
+            idx <= 6
+            and 3 <= len(text) <= 50
+            and RECIPIENT_TAIL_RE.search(text)
+            and any(kw in text for kw in RECIPIENT_HINTS)
+            and not any(kw in text for kw in RECIPIENT_BLOCKLIST)
+        ):
+            return ("recipient", 0.9)
+
+        if total - idx <= 3 and 5 <= len(text) <= 30 and any(
+            kw in text
+            for kw in ["局", "公司", "委员会", "人民政府", "办公厅", "办公室"]
+        ):
+            return ("signature", 0.7)
+
+        if size >= 14 or font:
+            return ("body", 0.85)
+
+        return ("unknown", 0.4)
+
+    @staticmethod
+    def _is_attachment_title(p: Paragraph, idx: int, doc: Document) -> bool:
+        """识别附件正文首页标题，避免按普通正文套用 GW-F-004。"""
+        if idx <= 0:
+            return False
+        text = p.text.strip()
+        font = (p.style.font_eastasia or "").strip()
+        if (
+            p.style.alignment != "center"
+            or (p.style.font_size_pt or 0) < 18
+            or "小标宋" not in font
+        ):
+            return False
+
+        marker_index = None
+        marker_text = ""
+        for prev in reversed(doc.paragraphs[:idx]):
+            if prev.role == "attachment_marker" or ATTACHMENT_RE.match(prev.text.strip()):
+                marker_index = prev.index
+                marker_text = prev.text.strip()
+                break
+        if marker_index is None or idx - marker_index > 12:
+            return False
+
+        attachment_name = re.sub(r"^附件\d*[:：]\s*", "", marker_text).strip()
+        attachment_name = re.sub(r"^\d+[\.．、)]\s*", "", attachment_name).strip()
+        return not attachment_name or text == attachment_name or text in attachment_name or attachment_name in text

fastapi_modules/fastapi_leaudit/govdoc_engine/parser/style_resolver.py

@@ -0,0 +1,241 @@
+"""OOXML 字体解析：处理样式继承链 + 主题字体。
+
+Word 把字体属性分散在四个层级：
+1. 直接 run rPr：`<w:r><w:rPr><w:rFonts/></w:rPr>...`
+2. 段落 rPr（段落标记字体）：`<w:p><w:pPr><w:rPr><w:rFonts/></w:rPr></w:pPr>`
+3. 段落引用样式：`<w:p><w:pPr><w:pStyle w:val="Heading1"/></w:pPr>`
+   样式定义在 styles.xml，可经 `<w:basedOn>` 链向上继承
+4. 全局默认：styles.xml 的 `<w:docDefaults>`
+
+此外 `<w:rFonts>` 的 `*Theme` 属性指向 theme1.xml 中的字体方案
+（majorEastAsia / minorEastAsia 等），需要做二次解析。
+"""
+
+from __future__ import annotations
+from dataclasses import dataclass
+from docx.oxml.ns import qn
+from lxml import etree
+
+# theme1.xml 命名空间
+_DML_NS = "http://schemas.openxmlformats.org/drawingml/2006/main"
+
+
+@dataclass
+class ResolvedRunStyle:
+    font_eastasia: str | None = None
+    font_ascii: str | None = None
+    size_pt: float | None = None
+    bold: bool | None = None
+    italic: bool | None = None
+
+
+def _empty_to_none(s: str | None) -> str | None:
+    if s is None:
+        return None
+    s = s.strip()
+    return s or None
+
+
+class StyleResolver:
+    """构造时一次性解析样式表 + 主题；之后 resolve_run() 是 O(链长)。"""
+
+    def __init__(self, docx):
+        self._theme = self._load_theme(docx)
+        self._styles, self._doc_defaults = self._load_styles(docx)
+
+    # ---- 主题 ---------------------------------------------------------
+    def _load_theme(self, docx) -> dict[tuple[str, str], str | None]:
+        """返回 {(axis, scheme_attr): font_name}。
+
+        scheme_attr 形如 'majorEastAsia' / 'minorAscii'，axis 是 rFonts 的轴。
+        """
+        out: dict[tuple[str, str], str | None] = {}
+        try:
+            theme_part = next(
+                p for p in docx.part.package.parts
+                if p.partname.endswith("/theme/theme1.xml")
+            )
+        except StopIteration:
+            return out
+        try:
+            root = etree.fromstring(theme_part.blob)
+        except etree.XMLSyntaxError:
+            return out
+
+        ns = {"a": _DML_NS}
+        for kind, font_tag in (("major", "majorFont"), ("minor", "minorFont")):
+            font_elem = root.find(f".//a:fontScheme/a:{font_tag}", ns)
+            if font_elem is None:
+                continue
+            latin = font_elem.find("a:latin", ns)
+            ea = font_elem.find("a:ea", ns)
+            cs = font_elem.find("a:cs", ns)
+            # ea 为空时用简中 Hans 兜底
+            ea_val = _empty_to_none(ea.get("typeface")) if ea is not None else None
+            if ea_val is None:
+                hans = font_elem.find('a:font[@script="Hans"]', ns)
+                if hans is not None:
+                    ea_val = _empty_to_none(hans.get("typeface"))
+            latin_val = _empty_to_none(latin.get("typeface")) if latin is not None else None
+            cs_val = _empty_to_none(cs.get("typeface")) if cs is not None else None
+
+            out[("ascii", f"{kind}Ascii")] = latin_val
+            out[("ascii", f"{kind}HAnsi")] = latin_val  # asciiTheme=majorHAnsi 也可能出现
+            out[("hAnsi", f"{kind}HAnsi")] = latin_val
+            out[("hAnsi", f"{kind}Ascii")] = latin_val
+            out[("eastAsia", f"{kind}EastAsia")] = ea_val
+            out[("cs", f"{kind}Bidi")] = cs_val
+        return out
+
+    # ---- 样式表 -------------------------------------------------------
+    def _load_styles(
+        self, docx
+    ) -> tuple[dict[str, dict], ResolvedRunStyle | None]:
+        out: dict[str, dict] = {}
+        defaults: ResolvedRunStyle | None = None
+        try:
+            styles_root = docx.part._styles_part.element
+        except (AttributeError, KeyError):
+            return out, defaults
+        if styles_root is None:
+            return out, defaults
+
+        # docDefaults
+        ddef = styles_root.find(qn("w:docDefaults"))
+        if ddef is not None:
+            rdef = ddef.find(qn("w:rPrDefault"))
+            if rdef is not None:
+                defaults = self._read_rpr(rdef.find(qn("w:rPr")))
+
+        # 各 style
+        for style in styles_root.findall(qn("w:style")):
+            sid = style.get(qn("w:styleId"))
+            if not sid:
+                continue
+            rpr = style.find(qn("w:rPr"))
+            ppr = style.find(qn("w:pPr"))
+            ppr_rpr = ppr.find(qn("w:rPr")) if ppr is not None else None
+            based_on = None
+            bo = style.find(qn("w:basedOn"))
+            if bo is not None:
+                based_on = bo.get(qn("w:val"))
+            link = style.find(qn("w:link"))
+            link_id = link.get(qn("w:val")) if link is not None else None
+            out[sid] = {
+                "rpr": rpr,
+                "ppr_rpr": ppr_rpr,
+                "based_on": based_on,
+                "link": link_id,
+            }
+        return out, defaults
+
+    # ---- 读 rPr -------------------------------------------------------
+    def _read_rpr(self, rpr) -> ResolvedRunStyle | None:
+        if rpr is None:
+            return None
+        rs = ResolvedRunStyle()
+        rfonts = rpr.find(qn("w:rFonts"))
+        if rfonts is not None:
+            rs.font_eastasia = self._resolve_font_axis(rfonts, "eastAsia")
+            rs.font_ascii = self._resolve_font_axis(rfonts, "ascii")
+        sz = rpr.find(qn("w:sz"))
+        if sz is not None and sz.get(qn("w:val")):
+            try:
+                rs.size_pt = float(sz.get(qn("w:val"))) / 2.0
+            except ValueError:
+                pass
+        if rpr.find(qn("w:b")) is not None:
+            rs.bold = True
+        if rpr.find(qn("w:i")) is not None:
+            rs.italic = True
+        return rs
+
+    def _resolve_font_axis(self, rfonts, axis: str) -> str | None:
+        """同一根 rFonts 上 explicit > theme。"""
+        explicit = _empty_to_none(rfonts.get(qn(f"w:{axis}")))
+        if explicit:
+            return explicit
+        theme_attr = "cstheme" if axis == "cs" else f"{axis}Theme"
+        theme = _empty_to_none(rfonts.get(qn(f"w:{theme_attr}")))
+        if theme:
+            return self._theme.get((axis, theme))
+        return None
+
+    # ---- 合并 ---------------------------------------------------------
+    @staticmethod
+    def _fill(target: ResolvedRunStyle, source: ResolvedRunStyle | None) -> None:
+        """target 已有的字段保留；缺的从 source 取。"""
+        if source is None:
+            return
+        if target.font_eastasia is None:
+            target.font_eastasia = source.font_eastasia
+        if target.font_ascii is None:
+            target.font_ascii = source.font_ascii
+        if target.size_pt is None:
+            target.size_pt = source.size_pt
+        if target.bold is None:
+            target.bold = source.bold
+        if target.italic is None:
+            target.italic = source.italic
+
+    def _resolve_style_chain(
+        self, sid: str | None, _seen: set[str] | None = None
+    ) -> ResolvedRunStyle | None:
+        """段落样式 → 链向 basedOn → 沿途累积 rPr 与 pPr 的 rPr。"""
+        if sid is None:
+            return None
+        seen = _seen or set()
+        if sid in seen:
+            return None
+        seen = seen | {sid}
+        info = self._styles.get(sid)
+        if info is None:
+            return None
+        # 当前 style 的两个 rPr
+        rs = ResolvedRunStyle()
+        self._fill(rs, self._read_rpr(info.get("rpr")))
+        self._fill(rs, self._read_rpr(info.get("ppr_rpr")))
+        # 链接的 character style（如果有）
+        if info.get("link"):
+            self._fill(rs, self._resolve_style_chain(info["link"], seen))
+        # 父样式
+        if info.get("based_on"):
+            self._fill(rs, self._resolve_style_chain(info["based_on"], seen))
+        return rs
+
+    # ---- 主入口 -------------------------------------------------------
+    def resolve_run(self, p_elem, run_elem) -> ResolvedRunStyle:
+        """解析单个 run 的最终样式。p_elem 可为 None。"""
+        rs = ResolvedRunStyle()
+        # 1. 直接 run rPr
+        if run_elem is not None:
+            self._fill(rs, self._read_rpr(run_elem.find(qn("w:rPr"))))
+        # 2. 段落 rPr（段落标记字体）+ pStyle 链
+        if p_elem is not None:
+            ppr = p_elem.find(qn("w:pPr"))
+            if ppr is not None:
+                self._fill(rs, self._read_rpr(ppr.find(qn("w:rPr"))))
+                pstyle = ppr.find(qn("w:pStyle"))
+                if pstyle is not None and pstyle.get(qn("w:val")):
+                    self._fill(rs, self._resolve_style_chain(pstyle.get(qn("w:val"))))
+        # 3. 默认 style "Normal"（中文文档常见）
+        if "Normal" in self._styles:
+            self._fill(rs, self._resolve_style_chain("Normal"))
+        # 4. docDefaults
+        self._fill(rs, self._doc_defaults)
+        return rs
+
+    def resolve_paragraph(self, p_elem) -> ResolvedRunStyle:
+        """段落整体样式（不读 run，仅 pPr/style/默认）。"""
+        rs = ResolvedRunStyle()
+        if p_elem is not None:
+            ppr = p_elem.find(qn("w:pPr"))
+            if ppr is not None:
+                self._fill(rs, self._read_rpr(ppr.find(qn("w:rPr"))))
+                pstyle = ppr.find(qn("w:pStyle"))
+                if pstyle is not None and pstyle.get(qn("w:val")):
+                    self._fill(rs, self._resolve_style_chain(pstyle.get(qn("w:val"))))
+        if "Normal" in self._styles:
+            self._fill(rs, self._resolve_style_chain("Normal"))
+        self._fill(rs, self._doc_defaults)
+        return rs

fastapi_modules/fastapi_leaudit/govdoc_engine/pipeline.py

@@ -0,0 +1,248 @@
+"""Govdoc 引擎主编排入口。
+
+将旧 govdoc-audit 的 audit_file() 函数适配为异步 Pipeline 接口，
+供 govdoc_bridge.runner 调用。
+
+迁移自: govdoc-audit/src/govdoc_audit/pipeline.py
+移除依赖: RunRecorder, config.py (local file logging)
+适配平台: 异步执行、直接返回 AuditResult
+"""
+
+from __future__ import annotations
+
+import logging
+import uuid
+from pathlib import Path
+from typing import Any
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.docx_parser import parse_docx
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.role_tagger import RoleTagger
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.extractor import FieldExtractor
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.entity_builder import (
+    EntityBuilder,
+    BUILTIN_LLM_DESCRIPTION,
+)
+from fastapi_modules.fastapi_leaudit.govdoc_engine.parser.entities import SemanticEntity
+from fastapi_modules.fastapi_leaudit.govdoc_engine.dsl.loader import load_rules
+from fastapi_modules.fastapi_leaudit.govdoc_engine.dsl.schema import EntitySpec, RuleSet
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.runner import RuleRunner
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.result import AuditResult, CheckedRule
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.structure import build_outline, build_structure
+from fastapi_modules.fastapi_leaudit.govdoc_engine.llm.client import LlmClient
+
+_log = logging.getLogger(__name__)
+
+
+# ── 辅助函数 ────────────────────────────────────────────
+
+def _outcomes_to_checked(outcomes) -> list[CheckedRule]:
+    """将规则执行结果汇总为 CheckedRule 列表。"""
+    rows: list[CheckedRule] = []
+    for o in outcomes:
+        if o.skipped:
+            status = "skipped"
+        elif o.findings:
+            status = "fail"
+        else:
+            status = "pass"
+        rows.append(
+            CheckedRule(
+                rule_id=o.rule.rule_id,
+                name=o.rule.name,
+                severity=o.rule.severity,
+                category=o.rule.category,
+                status=status,
+                skip_reason=o.skip_reason,
+            )
+        )
+    return rows
+
+
+def _build_result(
+    docx_path: Path, doc, findings, entities, outcomes,
+) -> AuditResult:
+    """从审查产物构建 AuditResult。"""
+    document_meta = {
+        "filename": docx_path.name,
+        "path": str(docx_path),
+        "page_count": doc.meta.get("page_count", 1),
+        "paragraph_count": len(doc.paragraphs),
+    }
+    result = AuditResult(
+        audit_id=f"A-{uuid.uuid4().hex[:8]}",
+        document=document_meta,
+        findings=findings,
+        entities=entities,
+        checked_rules=_outcomes_to_checked(outcomes),
+        structure=build_structure(doc),
+        outline=build_outline(doc),
+    )
+    result.compute_summary()
+    return result
+
+
+def _compute_missing_spec(
+    entities: dict[str, SemanticEntity | None],
+    custom_entities: list[EntitySpec],
+) -> dict[str, dict]:
+    """计算哪些实体需要送 LLM 抽取。"""
+    spec: dict[str, dict] = {}
+    for name, desc in BUILTIN_LLM_DESCRIPTION.items():
+        if entities.get(name) is None:
+            spec[name] = {
+                "description": desc,
+                "type": "list" if name == "attachments" else "string",
+            }
+    for s in custom_entities:
+        spec[s.name] = {"description": s.description or s.name, "type": s.type}
+    return spec
+
+
+def _merge_llm_into_entities(
+    entities: dict[str, SemanticEntity | None],
+    llm_values: dict[str, Any],
+) -> None:
+    """将 LLM 抽取结果合并进 entities。"""
+    for name, val in llm_values.items():
+        if val in (None, "", []):
+            continue
+        if isinstance(val, list):
+            text = "; ".join(
+                f"{it.get('序号', i + 1)}. {it.get('名称', '')}"
+                if isinstance(it, dict) else str(it)
+                for i, it in enumerate(val)
+            )
+            extra = {"items": val}
+        else:
+            text = str(val)
+            extra = {}
+        entities[name] = SemanticEntity(
+            name=name,
+            text=text,
+            paragraph_indices=[],
+            primary_role=None,
+            source="llm",
+            confidence=0.7,
+            extra=extra,
+        )
+
+
+# ── 实体构建 (同步，供 sync 入口使用) ──────────────────
+
+def _build_entities(
+    doc, ruleset: RuleSet, llm: LlmClient,
+) -> dict[str, SemanticEntity | None]:
+    """构建实体 + 差量 LLM 抽取（同步）。"""
+    entities = EntityBuilder().build(doc)
+    spec = _compute_missing_spec(entities, ruleset.extract.entities)
+    if spec:
+        llm_vals = FieldExtractor(llm).extract_missing(doc, spec)
+        _merge_llm_into_entities(entities, llm_vals)
+    return entities
+
+
+# ── 实体构建 (异步，供 async 入口使用) ──────────────────
+
+async def _build_entities_async(
+    doc, ruleset: RuleSet, llm: LlmClient,
+) -> dict[str, SemanticEntity | None]:
+    """构建实体 + 差量 LLM 抽取（异步）。"""
+    entities = EntityBuilder().build(doc)
+    spec = _compute_missing_spec(entities, ruleset.extract.entities)
+    if spec:
+        llm_vals = await FieldExtractor(llm).extract_missing_async(doc, spec)
+        _merge_llm_into_entities(entities, llm_vals)
+    return entities
+
+
+# ── 同步入口 (保留兼容) ─────────────────────────────────
+
+def audit_file(
+    docx_path: str | Path,
+    rules_path: str | Path,
+    llm_client: LlmClient | None = None,
+) -> AuditResult:
+    """同步审查单个公文文件。
+
+    Args:
+        docx_path: DOCX 文件路径。
+        rules_path: YAML 规则文件路径。
+        llm_client: 可选 LLM 客户端实例。
+
+    Returns:
+        AuditResult 包含 findings, entities, checked_rules, summary 等。
+    """
+    docx_path = Path(docx_path)
+    rules_path = Path(rules_path)
+    llm = llm_client or LlmClient()
+
+    doc = parse_docx(docx_path)
+    RoleTagger(llm_client=llm).tag(doc)
+
+    ruleset = load_rules(rules_path)
+    entities = _build_entities(doc, ruleset, llm)
+
+    findings, outcomes = RuleRunner(llm_client=llm).evaluate(
+        ruleset.all_rules(), doc, entities
+    )
+
+    return _build_result(docx_path, doc, findings, entities, outcomes)
+
+
+# ── 异步入口 (推荐，供 bridge 调用) ──────────────────────
+
+async def run(
+    file_path: str | Path,
+    rules_path: str | Path,
+    llm_client: LlmClient | None = None,
+) -> AuditResult:
+    """异步审查单个公文文件。
+
+    这是 govdoc_bridge 的主要调用入口。
+
+    Args:
+        file_path: 文档文件路径 (DOCX 或 PDF)。
+        rules_path: YAML 规则文件路径。
+        llm_client: 可选 LLM 客户端实例。
+
+    Returns:
+        AuditResult 包含 findings, entities, checked_rules, summary 等。
+    """
+    file_path = Path(file_path)
+    rules_path = Path(rules_path)
+    llm = llm_client or LlmClient()
+
+    _log.info("Govdoc pipeline start: %s", file_path.name)
+
+    # 1. 解析文档
+    doc = parse_docx(file_path)
+    _log.info("  parsed: %d paragraphs", len(doc.paragraphs))
+
+    # 2. 段落角色标注
+    RoleTagger(llm_client=llm).tag(doc)
+
+    # 3. 加载规则
+    ruleset = load_rules(rules_path)
+    _log.info("  rules: %d groups, %d rules", len(ruleset.groups), len(ruleset.all_rules()))
+
+    # 4. 实体抽取 (含差量 LLM)
+    entities = await _build_entities_async(doc, ruleset, llm)
+    _log.info("  entities: %d/%d resolved", sum(1 for v in entities.values() if v), len(entities))
+
+    # 5. 规则评估
+    findings, outcomes = RuleRunner(llm_client=llm).evaluate(
+        ruleset.all_rules(), doc, entities
+    )
+    _log.info("  evaluated: %d findings from %d rules", len(findings), len(outcomes))
+
+    # 6. 构建结果
+    result = _build_result(file_path, doc, findings, entities, outcomes)
+    _log.info(
+        "Govdoc pipeline complete: score=%d, pass=%d, fail=%d, skip=%d",
+        result.summary.score,
+        result.summary.passed_count,
+        result.summary.failed_count,
+        result.summary.skipped_count,
+    )
+
+    return result

fastapi_modules/fastapi_leaudit/govdoc_engine/reporter/docx_annotator.py

@@ -0,0 +1,105 @@
+"""docx 标注：在原文加高亮 + 文末追加审核报告附页。"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from docx import Document as DocxDocument
+from docx.enum.text import WD_BREAK
+from docx.shared import Pt
+from docx.oxml.ns import qn
+from lxml import etree
+
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.result import AuditResult
+
+
+_HIGHLIGHT_NAME = {
+    "error": "red",
+    "warning": "yellow",
+    "info": "cyan",
+}
+
+
+def _highlight_run(run, color_name: str) -> None:
+    rpr = run._element.get_or_add_rPr()
+    hl = rpr.find(qn("w:highlight"))
+    if hl is None:
+        hl = etree.SubElement(rpr, qn("w:highlight"))
+    hl.set(qn("w:val"), color_name)
+
+
+def _highlight_paragraph_range(paragraph, start: int, end: int, color_name: str) -> None:
+    """简化策略：高亮整段（精准 char range 留 v0.2 实现）。"""
+    for run in paragraph.runs:
+        _highlight_run(run, color_name)
+
+
+def _add_heading_with_fallback(doc, text: str, level: int = 1):
+    try:
+        return doc.add_heading(text, level=level)
+    except KeyError:
+        # Some uploaded documents don't include Word's built-in heading styles.
+        p = doc.add_paragraph()
+        run = p.add_run(text)
+        run.bold = True
+        if level == 1:
+            run.font.size = Pt(16)
+        elif level == 2:
+            run.font.size = Pt(13)
+        else:
+            run.font.size = Pt(12)
+        return p
+
+
+def _append_appendix(doc, result: AuditResult) -> None:
+    p = doc.add_paragraph()
+    p.add_run().add_break(WD_BREAK.PAGE)
+    _add_heading_with_fallback(doc, "审核报告附页", level=1)
+
+    s = result.summary
+    doc.add_paragraph(
+        f"得分: {s.score}/100   错误: {s.by_severity.get('error', 0)}   "
+        f"警告: {s.by_severity.get('warning', 0)}   提示: {s.by_severity.get('info', 0)}"
+    )
+
+    if not result.findings:
+        doc.add_paragraph("未发现问题。")
+        return
+
+    table = doc.add_table(rows=1, cols=5)
+    try:
+        table.style = "Light Grid"
+    except KeyError:
+        # Some source documents don't ship with the built-in table style set.
+        pass
+    hdr = table.rows[0].cells
+    for i, h in enumerate(["编号", "规则", "严重度", "类别", "位置 / 说明"]):
+        hdr[i].text = h
+
+    for f in result.findings:
+        row = table.add_row().cells
+        row[0].text = f.finding_id
+        row[1].text = f.rule_id
+        row[2].text = f.severity
+        row[3].text = f.category
+        loc = f.location
+        ctx = (loc.context or "")[:30]
+        row[4].text = f"P{loc.paragraph_index} ({loc.role}): {f.message}\n  原文: {ctx}"
+
+
+def annotate_docx(src: str | Path, dst: str | Path, result: AuditResult) -> None:
+    src = Path(src)
+    dst = Path(dst)
+    doc = DocxDocument(src)
+
+    for f in result.findings:
+        idx = f.location.paragraph_index
+        if 0 <= idx < len(doc.paragraphs):
+            color = _HIGHLIGHT_NAME.get(f.severity, "yellow")
+            _highlight_paragraph_range(
+                doc.paragraphs[idx], f.location.char_start, f.location.char_end, color
+            )
+
+    _append_appendix(doc, result)
+    dst.parent.mkdir(parents=True, exist_ok=True)
+    doc.save(dst)

fastapi_modules/fastapi_leaudit/govdoc_engine/reporter/html_paragraph.py

@@ -0,0 +1,42 @@
+"""把 Document 渲染为带 inline style 的 HTML 段落，给前端用。"""
+
+from __future__ import annotations
+from html import escape
+from fastapi_modules.fastapi_leaudit.govdoc_engine.models import Document
+
+
+def _style(p) -> str:
+    s = p.style
+    parts = []
+    if s.font_size_pt:
+        sz = s.font_size_pt
+        sz_str = str(int(sz)) if sz == int(sz) else str(sz)
+        parts.append(f"font-size:{sz_str}pt")
+    if s.font_eastasia:
+        parts.append(f"font-family:'{s.font_eastasia}',serif")
+    if s.alignment and s.alignment != "left":
+        parts.append(f"text-align:{s.alignment}")
+    if s.bold:
+        parts.append("font-weight:700")
+    if s.first_line_indent_pt:
+        parts.append(f"text-indent:{s.first_line_indent_pt}pt")
+    return ";".join(parts)
+
+
+def paragraphs_to_html(doc: Document, finding_map: dict[int, list[str]]) -> str:
+    """把 doc 每个段落渲染成 <p> 带 data-pi / data-role / data-finding-ids。"""
+    out = ['<div class="doc-view">']
+    for p in doc.paragraphs:
+        style = _style(p)
+        finding_ids = finding_map.get(p.index, [])
+        attrs = [
+            f'data-pi="{p.index}"',
+            f'data-role="{escape(p.role or "")}"',
+        ]
+        if finding_ids:
+            attrs.append(f'data-finding-ids="{escape(",".join(finding_ids))}"')
+        if style:
+            attrs.append(f'style="{escape(style)}"')
+        out.append(f"<p {' '.join(attrs)}>{escape(p.text)}</p>")
+    out.append("</div>")
+    return "\n".join(out)

fastapi_modules/fastapi_leaudit/govdoc_engine/reporter/html_renderer.py

@@ -0,0 +1,76 @@
+"""把 AuditResult 渲染成单文件 HTML 报告。"""
+
+from __future__ import annotations
+from html import escape
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.result import AuditResult
+
+
+_CSS = """
+body { font-family: -apple-system, "PingFang SC", sans-serif; margin: 0; padding: 24px;
+       background: #f7f7f9; color: #1a1a1a; }
+.header { display: flex; align-items: center; gap: 16px; margin-bottom: 24px; }
+.score { width: 96px; height: 96px; border-radius: 50%;
+         background: conic-gradient(#22c55e var(--p), #e5e7eb var(--p));
+         display: grid; place-items: center; font-weight: 700; font-size: 22px; color: #111; }
+.score-inner { background: white; width: 76px; height: 76px; border-radius: 50%;
+               display: grid; place-items: center; }
+.tag { padding: 2px 8px; border-radius: 999px; font-size: 12px; }
+.error { background: #fee2e2; color: #b91c1c; }
+.warning { background: #fef9c3; color: #a16207; }
+.info { background: #dbeafe; color: #1d4ed8; }
+table { width: 100%; border-collapse: collapse; background: white; border-radius: 8px;
+        overflow: hidden; box-shadow: 0 1px 3px rgba(0,0,0,0.06); }
+th, td { padding: 10px 12px; text-align: left; border-bottom: 1px solid #f1f5f9; vertical-align: top; }
+th { background: #f8fafc; font-size: 13px; }
+td.msg { max-width: 480px; }
+.context { color: #64748b; font-size: 12px; margin-top: 4px; }
+"""
+
+
+def render_html(result: AuditResult) -> str:
+    s = result.summary
+    score = s.score
+    pct = f"{score}%"
+    rows = []
+    for f in result.findings:
+        loc = f.location
+        suggest = (
+            f'<div style="color:#0369a1">建议: {escape(f.suggestion)}</div>'
+            if f.suggestion else ""
+        )
+        rows.append(f"""
+<tr>
+  <td>{escape(f.finding_id)}</td>
+  <td>{escape(f.rule_id)}<br><span style="color:#64748b;font-size:12px">{escape(f.rule_name)}</span></td>
+  <td><span class="tag {f.severity}">{f.severity}</span></td>
+  <td>{escape(f.category)}</td>
+  <td>P{loc.paragraph_index} ({escape(loc.role or '')})</td>
+  <td class="msg">{escape(f.message)}
+    <div class="context">原文: {escape((loc.context or '')[:80])}</div>
+    {suggest}
+  </td>
+</tr>""")
+
+    body = f"""<!doctype html>
+<html lang="zh"><head><meta charset="utf-8"><title>公文审核报告</title>
+<style>{_CSS}</style></head><body>
+<div class="header">
+  <div class="score" style="--p:{pct}"><div class="score-inner">{score}</div></div>
+  <div>
+    <h1 style="margin:0">公文格式审核报告</h1>
+    <div style="color:#64748b">{escape(result.document.get('filename', ''))} · 共 {s.total_findings} 项</div>
+    <div style="margin-top:6px">
+      <span class="tag error">错误 {s.by_severity.get('error', 0)}</span>
+      <span class="tag warning">警告 {s.by_severity.get('warning', 0)}</span>
+      <span class="tag info">提示 {s.by_severity.get('info', 0)}</span>
+    </div>
+  </div>
+</div>
+<table>
+  <thead><tr>
+    <th>编号</th><th>规则</th><th>严重度</th><th>类别</th><th>位置</th><th>说明</th>
+  </tr></thead>
+  <tbody>{''.join(rows) or '<tr><td colspan=6>未发现问题</td></tr>'}</tbody>
+</table>
+</body></html>"""
+    return body

fastapi_modules/fastapi_leaudit/govdoc_engine/reporter/json_report.py

@@ -0,0 +1,12 @@
+"""把 AuditResult 序列化为 JSON 字符串。"""
+
+import json
+from fastapi_modules.fastapi_leaudit.govdoc_engine.engine.result import AuditResult
+
+
+def to_json(result: AuditResult, indent: int = 2) -> str:
+    return json.dumps(
+        result.model_dump(mode="json"),
+        ensure_ascii=False,
+        indent=indent,
+    )