# -*- coding: utf-8 -*- """iAOP-Core · LLM 网关 —— 幻觉/事实性校验中间件(EPIC #6 主体,Issue #47 雏形)。 对应 PRD 5.4「④ LLM 网关 + RAG」: - **事实性校验**:RAG 答案强制**引用溯源**(返回命中文档片段+来源); 对高利害输出(如处置建议)设置信度阈值,低于阈值触发"人工确认"; 定期用评测集检验事实一致性。 本模块实现 `HallucinationGuard`: - **引用溯源校验**:模型输出中声称引用的片段(`[来源: ]`)必须能在 RAG 检索命中的文档片段中找到对应来源,找不到即判定 `unsupported` (无源引用 = 幻觉嫌疑); - **信度阈值**:对高利害输出(处置建议 / 报警解释)要求信度 ≥ 阈值, 低于阈值返回 `human_review`(转人工确认,PRD 5.4 异常时转人工); - **评测集检验**:`evaluate()` 对 (prompt, answer, expected_sources) 样本 批量评估事实一致性(供"定期评测"脚本调用)。 设计说明(供子任务 #47 继续细化): - 本版实现校验核心(溯源 + 信度阈值 + 评测入口); - 子任务 #47 将在此基础上补齐与 Prompt 版本库的联动与评测报告脚本。 测试:`python -m unittest discover -s tests -v`(在 core/llm-gateway 目录下执行)。 """ from __future__ import annotations import re import uuid from dataclasses import dataclass, field from datetime import datetime, timezone from typing import Dict, List, Optional, Sequence # 输出中引用声明的格式:`[来源: 文档标题]` 或 `[src: doc_id]` _SOURCE_REF_RE = re.compile(r"\[来源[::]\s*([^\]]+)\]", re.IGNORECASE) @dataclass(frozen=True) class GuardVerdict: """一次事实性校验的结论。""" answer: str supported: bool # 所有引用声明均有真实来源 confidence: float # 调用方给出的信度(0~1) threshold: float # 本次校验使用的信度阈值 action: str # pass / human_review / unsupported missing_sources: List[str] = field(default_factory=list) verdict_id: str = field(default_factory=lambda: uuid.uuid4().hex[:12]) created_at: str = field(default_factory=lambda: datetime.now(timezone.utc).isoformat()) def to_dict(self) -> Dict[str, object]: return { "verdict_id": self.verdict_id, "created_at": self.created_at, "supported": self.supported, "confidence": self.confidence, "threshold": self.threshold, "action": self.action, "missing_sources": self.missing_sources, "answer": self.answer, } class HallucinationGuard: """幻觉/事实性校验中间件。 `check(answer, sources, confidence, high_stakes=False)`: - `sources`:本次 RAG 检索实际命中的文档标题列表; - `high_stakes=True`:启用信度阈值(处置建议 / 报警解释等), 低于阈值 → `human_review`; - 输出中所有 `[来源: X]` 声明必须出现在 `sources` 中, 否则 → `unsupported`(缺失引用列表随结论返回)。 """ def __init__(self, default_threshold: float = 0.8) -> None: self.default_threshold = default_threshold self._audit: List[Dict[str, object]] = [] def check(self, answer: str, sources: Sequence[str], confidence: float = 1.0, high_stakes: bool = False, threshold: Optional[float] = None) -> GuardVerdict: """校验一条模型输出。返回结论(不修改输出,由调用方决定如何处置)。""" th = threshold if threshold is not None else self.default_threshold # 1) 引用溯源:输出中声明的来源必须真实存在 declared = _SOURCE_REF_RE.findall(answer) available = set(sources) missing = [s.strip() for s in declared if s.strip() not in available] supported = not missing # 2) 高利害 → 信度阈值 if high_stakes and confidence < th: action = "human_review" elif not supported: action = "unsupported" else: action = "pass" verdict = GuardVerdict( answer=answer, supported=supported, confidence=confidence, threshold=th, action=action, missing_sources=missing, ) self._audit.append(verdict.to_dict()) return verdict # -- 评测集检验(定期事实一致性评测入口) ------------------------------ def evaluate(self, samples: List[Dict[str, object]]) -> Dict[str, object]: """批量评估事实一致性。 `samples`:`[{"answer", "sources", "confidence", "high_stakes"}, ...]`。 返回支持率 / 人工复核率 / 未支持率。子任务 #47 将扩展为评测报告。 """ total = len(samples) if total == 0: return {"supported_rate": 0.0, "human_review_rate": 0.0, "total": 0} supported = 0 human = 0 for s in samples: v = self.check( answer=str(s.get("answer", "")), sources=[str(x) for x in s.get("sources", [])], confidence=float(s.get("confidence", 1.0)), high_stakes=bool(s.get("high_stakes", False)), ) if v.supported: supported += 1 if v.action == "human_review": human += 1 return { "supported_rate": round(supported / total, 4), "human_review_rate": round(human / total, 4), "unsupported_rate": round((total - supported) / total, 4), "total": total, } # -- 审计 -------------------------------------------------------------- def drain_audit(self) -> List[Dict[str, object]]: out, self._audit = self._audit, [] return out def __repr__(self) -> str: # pragma: no cover - 调试辅助 return f""