"""异常调查的确定性完成条件。模型选择下一次查询,系统核对证据与停止原因。""" from __future__ import annotations import math from typing import Any from pydantic import BaseModel, ConfigDict, Field INVESTIGATION_VERSION = 1 class InvestigationPolicy(BaseModel): model_config = ConfigDict(extra="forbid") enabled: bool = True max_rounds: int = Field(default=2, ge=0, le=5) minimum_depth: int = Field(default=1, ge=0, le=5) min_coverage: float = Field(default=0.8, ge=0, le=1) def _number(value: Any) -> float: try: number = float(value) return number if math.isfinite(number) else 0.0 except (ValueError, TypeError, OverflowError): return 0.0 def assess(scope: Any, question: str, policy: InvestigationPolicy) -> dict[str, Any]: from app.analysis.intent import analysis_intent results = list(scope.results.values()) if scope is not None else [] intent = analysis_intent(question) active = bool(policy.enabled and scope is not None and ( (intent is not None and intent.kind == "why") or any( r.get("status") == "ok" and (r.get("verdict") or {}).get("code") in ("outside_band", "far_outside_band") for r in results))) out: dict[str, Any] = {"version": INVESTIGATION_VERSION, "active": active, "status": "not_required", "depth": 0, "coverage": None, "query_count": len(scope.queries) if scope is not None else 0, "analysis_count": len(results), "reason": "", "next_step": ""} if not active: return out def pending(reason: str, next_step: str, stage: str) -> dict[str, Any]: return {**out, "status": "partial", "stage": stage, "reason": reason, "next_step": next_step} done = [r for r in results if r.get("status") == "ok"] decomposed = [r for r in done if r.get("op") == "decompose"] if not decomposed: refused = [r for r in results if r.get("status") == "refused"] if refused: return pending(str(refused[-1].get("reason") or "现有数据不满足分析条件"), "检查缺数、截断、时间范围和口径;可以补齐时重新查询,无法补齐时明确证据缺口。", "data_quality") return pending("尚未对偏差进行贡献分解", "先核对口径、数据完整性和两期可比性,再选择业务维度或因子进行分解。", "decompose") last = decomposed[-1] params = last.get("params") or {} depth = int(params.get("depth") or 0) out["depth"] = depth chain = [last] seen = set() while (parent := (chain[-1].get("params") or {}).get("parent")): name = parent.get("of") if isinstance(parent, dict) else None previous = scope.results.get(name) if isinstance(name, str) else None if (name in seen or not isinstance(previous, dict) or previous.get("status") != "ok" or previous.get("op") != "decompose" or previous.get("mode") != last.get("mode")): return pending("下钻与上层结果的关联不完整", "重新取得上层分解,用 parent 关联真实的贡献项。", "lineage") seen.add(name) chain.append(previous) if depth != len(chain) - 1: return pending("下钻层数与证据链不一致", "从总指标开始,用 parent 逐层关联分解结果。", "lineage") caveats = {c.get("code") for result in chain for c in result.get("caveats") or [] if isinstance(c, dict)} code = (last.get("verdict") or {}).get("code") if any((r.get("verdict") or {}).get("code") == "unclosed" for r in chain) or "headline_basis" in caveats: return pending("分解尚未与总指标核对闭合", "取同一口径的总数作为 headline,核对分组和总数的差异,再继续解释。", "closure") if "comparability" in caveats: return pending("两期可比性尚未确认", "先核对两期的窗口、过滤条件和数据完整性;无法确认时保留有条件结论。", "comparability") comparable = [r for r in done if r.get("op") == "comparable"] # 算子已有内部可比性检查的结果不要求模型重复调用;缺少声明则不能据此证明可比。 if comparable and (comparable[-1].get("verdict") or {}).get("code") != "comparable": return pending("可比性检查未通过", "先解决可比性缺口,再基于同口径数据重新分解。", "comparability") values = last.get("values") or {} coverage = max(_number(values.get("top_share")), _number(values.get("top_k_share"))) if last.get("mode") == "additive": from app.analysis.inputs import same_value coverage = 1.0 if code in ("flat", "dispersed") else max(0.0, min(1.0, coverage)) for child, previous in zip(chain, chain[1:]): key = child["params"]["parent"].get("key") or [] table = previous.get("table") or {} columns = table.get("columns") or [] row = next((r for r in table.get("rows") or [] if key and len(r) >= len(columns) and all(same_value(k, v) for k, v in zip(key, r[:len(key)]))), None) if row is None or "share" not in columns: return pending("无法核对下钻贡献占整体偏差的比例", "保留上层贡献表及下钻的分组键,再核对整体覆盖程度。", "coverage") coverage *= max(0.0, min(1.0, _number(row[columns.index("share")]))) out["coverage"] = coverage if last.get("mode") == "additive" else None if code in ("flat", "dispersed") and depth == 0: return {**out, "status": "complete", "stage": "contribution", "reason": "分解显示变化平稳或贡献分散"} if last.get("mode") != "factor" and depth < policy.minimum_depth: return pending("主要贡献项尚未进一步下钻", "对主要贡献项选择更细的业务维度,用 parent 关联上层结果继续下钻;没有合适维度时说明限制。", "drill") if last.get("mode") == "additive" and coverage < policy.min_coverage: return pending("主要贡献项的解释覆盖程度不足", "检查其他候选维度和剩余差异,保留未解释的部分,不把贡献分解写成因果证明。", "coverage") return {**out, "status": "complete", "stage": "contribution", "reason": "完成贡献分析,因果关系仍需独立证据"}