X-Financial/server/src/app/services/risk_rule_generation.py

from __future__ import annotations

import json
import re
from datetime import UTC, datetime
from typing import Any

from sqlalchemy.orm import Session

from app.core.agent_enums import AgentAssetDomain, AgentAssetStatus, AgentAssetType
from app.models.agent_asset import AgentAsset, AgentAssetVersion
from app.schemas.agent_asset import AgentAssetRiskRuleGenerateRequest
from app.services.agent_asset_rule_library import AgentAssetRuleLibraryManager
from app.services.agent_asset_spreadsheet import RISK_RULES_LIBRARY
from app.services.audit import AuditLogService
from app.services.risk_rule_explainability import build_risk_rule_explainability_artifacts
from app.services.risk_rule_generation_ontology import (
    BUSINESS_DOMAIN_LABELS,
    DOMAIN_FIELD_PREFIXES,
    EXPENSE_BUSINESS_STAGE_LABELS,
    EXPENSE_RISK_CATEGORY_ALIASES,
    EXPENSE_RISK_CATEGORY_LABELS,
    FIELD_ONTOLOGY,
    RISK_LEVEL_LABELS,
    RiskRuleField,
)
from app.services.risk_rule_generation_prompt import build_risk_rule_compiler_messages
from app.services.risk_rule_generation_interpreter import COMPOSITE_RULE_TEMPLATE_KEY
from app.services.risk_rule_generation_markdown import build_risk_rule_version_markdown
from app.services.risk_rule_generation_semantics import (
    CITY_CONSISTENCY_SEMANTIC_TYPE,
    CITY_CONSISTENCY_SEMANTIC_TYPES,
    build_city_consistency_draft,
    build_city_consistency_params,
)
from app.services.risk_rule_generation_semantic_plan import unwrap_semantic_plan_payload
from app.services.risk_rule_dsl_validator import validate_risk_rule_draft
from app.services.risk_rule_scoring import apply_risk_score_to_draft, calculate_risk_rule_score
from app.services.runtime_chat import RuntimeChatService


class RiskRuleGenerationService:
    def __init__(
        self,
        db: Session,
        *,
        rule_library_manager: AgentAssetRuleLibraryManager | None = None,
        runtime_chat_service: RuntimeChatService | None = None,
    ) -> None:
        self.db = db
        self.rule_library_manager = rule_library_manager or AgentAssetRuleLibraryManager()
        self.runtime_chat_service = runtime_chat_service or RuntimeChatService(db)
        self.audit_service = AuditLogService(db)

    def generate_rule_asset(
        self,
        body: AgentAssetRiskRuleGenerateRequest,
        *,
        actor: str,
        request_id: str | None = None,
    ) -> str:
        domain = body.business_domain.value
        if domain not in BUSINESS_DOMAIN_LABELS:
            raise ValueError("当前仅支持报销、应收、应付业务域的新建风险规则。")

        natural_language = self._clean_text(body.natural_language)
        if len(natural_language) < 8:
            raise ValueError("请至少输入 8 个字的风险规则描述。")

        rule_title = self._clean_text(body.rule_title)
        if rule_title and len(rule_title) < 2:
            raise ValueError("规则标题至少需要 2 个字。")

        requires_attachment = bool(body.requires_attachment)
        business_stage = self._normalize_business_stage(body.business_stage, domain)
        business_stage_label = EXPENSE_BUSINESS_STAGE_LABELS.get(business_stage, "费用报销")
        expense_category = self._normalize_expense_category(body.expense_category, domain)
        expense_category_label = EXPENSE_RISK_CATEGORY_LABELS.get(expense_category or "", "")

        created_at = datetime.now(UTC)
        fields = self._resolve_fields(natural_language, domain=domain)
        draft = self._compile_with_model(
            natural_language=natural_language,
            domain=domain,
            business_stage=business_stage,
            business_stage_label=business_stage_label,
            expense_category=expense_category,
            expense_category_label=expense_category_label,
            fields=fields,
        ) or self._build_fallback_draft(
            natural_language=natural_language,
            domain=domain,
            expense_category_label=expense_category_label,
            risk_level="medium",
            fields=fields,
        )
        draft = validate_risk_rule_draft(draft, fields=fields, natural_language=natural_language)
        draft = self._align_draft_fields(
            draft,
            natural_language=natural_language,
            risk_level="medium",
            fields=fields,
        )
        draft = validate_risk_rule_draft(draft, fields=fields, natural_language=natural_language)
        risk_score = calculate_risk_rule_score(
            natural_language=natural_language,
            draft=draft,
            fields=fields,
            expense_category=expense_category,
            expense_category_label=expense_category_label,
            requires_attachment=requires_attachment,
        )
        risk_level = str(risk_score["level"])
        draft = apply_risk_score_to_draft(draft, risk_score)
        payload = self._build_rule_payload(
            draft,
            natural_language=natural_language,
            domain=domain,
            business_stage=business_stage,
            business_stage_label=business_stage_label,
            expense_category=expense_category,
            expense_category_label=expense_category_label,
            risk_level=risk_level,
            fields=fields,
            created_at=created_at,
            actor=actor,
            requires_attachment=requires_attachment,
            rule_title=rule_title,
            risk_score=risk_score,
        )
        rule_code = str(payload["rule_code"])
        file_name = f"{rule_code}.json"

        self.rule_library_manager.write_rule_library_json(
            library=RISK_RULES_LIBRARY,
            file_name=file_name,
            payload=payload,
        )

        asset = AgentAsset(
            asset_type=AgentAssetType.RULE.value,
            code=rule_code,
            name=str(payload["name"]),
            description=str(payload["description"]),
            domain=domain,
            scenario_json=[str(payload.get("risk_category") or BUSINESS_DOMAIN_LABELS[domain])],
            owner=actor,
            reviewer=None,
            status=AgentAssetStatus.DRAFT.value,
            current_version="v0.1.0",
            published_version=None,
            working_version="v0.1.0",
            config_json={
                "severity": risk_level,
                "risk_score": risk_score["score"],
                "risk_level": risk_level,
                "risk_level_label": risk_score["level_label"],
                "risk_score_detail": risk_score,
                "enabled": True,
                "requires_attachment": requires_attachment,
                "tag": "风险规则",
                "detail_mode": "json_risk",
                "business_stage": business_stage,
                "business_stage_label": business_stage_label,
                "expense_category": expense_category,
                "expense_category_label": expense_category_label,
                "risk_category": payload.get("risk_category"),
                "rule_library": RISK_RULES_LIBRARY,
                "rule_document": {
                    "file_name": file_name,
                    "storage_key": f"rules/{RISK_RULES_LIBRARY}/{file_name}",
                },
                "ontology_signal": payload.get("ontology_signal"),
                "evaluator": payload.get("evaluator"),
                "generated_by": "natural_language",
                "source_ref": "自然语言风险规则",
                "last_operation": {
                    "action": "create",
                    "actor": actor,
                    "at": datetime.now(UTC).isoformat(),
                },
            },
        )
        self.db.add(asset)
        self.db.flush()
        self.db.add(
            AgentAssetVersion(
                asset_id=asset.id,
                version="v0.1.0",
                content=build_risk_rule_version_markdown(payload),
                content_type="markdown",
                change_note="通过自然语言新建风险规则草稿。",
                created_by=actor,
            )
        )
        self.audit_service.log_action(
            actor=actor,
            action="generate_agent_asset_risk_rule",
            resource_type=AgentAssetType.RULE.value,
            resource_id=asset.id,
            before_json=None,
            after_json={
                "rule_code": rule_code,
                "risk_level": risk_level,
                "risk_score": risk_score["score"],
                "domain": domain,
                "business_stage": business_stage,
                "expense_category": expense_category,
                "requires_attachment": requires_attachment,
            },
            request_id=request_id,
        )
        self.db.refresh(asset)
        return asset.id

    def _compile_with_model(
        self,
        *,
        natural_language: str,
        domain: str,
        business_stage: str,
        business_stage_label: str,
        expense_category: str | None,
        expense_category_label: str,
        fields: list[RiskRuleField],
    ) -> dict[str, Any] | None:
        field_payload = [
            {
                "key": item.key,
                "label": item.label,
                "type": item.field_type,
                "source": item.source,
            }
            for item in fields
        ]
        messages = build_risk_rule_compiler_messages(
            domain=domain,
            domain_label=BUSINESS_DOMAIN_LABELS[domain],
            business_stage=business_stage,
            business_stage_label=business_stage_label,
            expense_category=expense_category,
            expense_category_label=expense_category_label,
            natural_language=natural_language,
            available_fields=field_payload,
        )
        answer = self.runtime_chat_service.complete(
            messages,
            max_tokens=1400,
            temperature=0.1,
            timeout_seconds=12,
            max_attempts=1,
        )
        if not answer:
            return None

        try:
            payload = json.loads(self._extract_json_object(answer))
        except (json.JSONDecodeError, ValueError):
            return None
        if not isinstance(payload, dict):
            return None
        payload = unwrap_semantic_plan_payload(payload)
        return self._sanitize_model_draft(payload, fields=fields)

    def _sanitize_model_draft(
        self,
        payload: dict[str, Any],
        *,
        fields: list[RiskRuleField],
    ) -> dict[str, Any]:
        allowed_fields = {item.key for item in fields}
        template_key = str(payload.get("template_key") or "").strip()
        if template_key not in {
            "field_required_v1",
            "field_compare_v1",
            "keyword_match_v1",
            COMPOSITE_RULE_TEMPLATE_KEY,
        }:
            template_key = "field_required_v1"

        raw_field_keys = payload.get("field_keys")
        field_keys = [
            str(item or "").strip()
            for item in (raw_field_keys if isinstance(raw_field_keys, list) else [])
            if str(item or "").strip() in allowed_fields
        ]
        if not field_keys and fields:
            field_keys = [fields[0].key]

        keywords = [
            str(item or "").strip()
            for item in (
                payload.get("keywords") if isinstance(payload.get("keywords"), list) else []
            )
            if str(item or "").strip()
        ]
        exception_keywords = [
            str(item or "").strip()
            for item in (
                payload.get("exception_keywords")
                if isinstance(payload.get("exception_keywords"), list)
                else []
            )
            if str(item or "").strip()
        ]
        unsupported_fields = [
            str(item or "").strip()
            for item in (
                payload.get("unsupported_fields")
                if isinstance(payload.get("unsupported_fields"), list)
                else []
            )
            if str(item or "").strip()
        ]
        flow = payload.get("flow") if isinstance(payload.get("flow"), dict) else {}
        rule_ir = payload.get("rule_ir") if isinstance(payload.get("rule_ir"), dict) else {}
        draft = {
            "name": self._clean_text(payload.get("name"))[:80],
            "description": self._clean_text(payload.get("description")),
            "template_key": template_key,
            "semantic_type": self._clean_text(payload.get("semantic_type")),
            "field_keys": field_keys,
            "condition_summary": self._clean_text(payload.get("condition_summary")),
            "keywords": keywords[:12],
            "exception_keywords": exception_keywords[:12],
            "unsupported_fields": unsupported_fields[:20],
            "rule_ir": rule_ir,
            "flow": {
                "start": self._clean_text(flow.get("start")) or "提交业务单据",
                "evidence": self._clean_text(flow.get("evidence")) or "读取规则字段",
                "decision": self._clean_text(flow.get("decision")) or "判断是否命中风险",
                "pass": self._clean_text(flow.get("pass")) or "继续流转",
                "fail": self._clean_text(flow.get("fail")) or "提示风险并进入复核",
            },
        }
        for key in ("conditions", "hit_logic", "field_groups"):
            value = payload.get(key)
            if isinstance(value, (list, dict)):
                draft[key] = value
        scoring_evidence = payload.get("risk_scoring_evidence")
        if isinstance(scoring_evidence, dict):
            draft["risk_scoring_evidence"] = scoring_evidence
        if isinstance(payload.get("model_semantic_plan"), dict):
            draft["model_semantic_plan"] = payload["model_semantic_plan"]
        for key in ("formula", "message_template"):
            value = self._clean_text(payload.get(key))
            if value:
                draft[key] = value
        return draft

    def _build_fallback_draft(
        self,
        *,
        natural_language: str,
        domain: str,
        expense_category_label: str,
        risk_level: str,
        fields: list[RiskRuleField],
    ) -> dict[str, Any]:
        field_keys = [item.key for item in fields[:4]]
        template_key = self._infer_template_key(natural_language)
        condition_summary = self._build_condition_summary(
            natural_language,
            template_key=template_key,
            fields=fields,
        )
        name = self._infer_rule_name(natural_language)
        business_label = expense_category_label or BUSINESS_DOMAIN_LABELS[domain]
        description = (
            f"当{business_label}业务满足“{natural_language}”时，系统会按"
            f"{RISK_LEVEL_LABELS[risk_level]}进行提示，并要求经办人或审核人补充核对依据。"
        )
        return {
            "name": name,
            "description": description,
            "template_key": template_key,
            "field_keys": field_keys,
            "condition_summary": condition_summary,
            "keywords": self._infer_keywords(natural_language),
            "flow": {
                "start": f"{business_label}单据提交",
                "evidence": "读取" + "、".join(item.label for item in fields[:3]),
                "decision": condition_summary,
                "pass": "未命中风险，继续业务流转",
                "fail": f"命中{RISK_LEVEL_LABELS[risk_level]}，提示复核",
            },
        }

    def _build_rule_payload(
        self,
        draft: dict[str, Any],
        *,
        natural_language: str,
        domain: str,
        business_stage: str,
        business_stage_label: str,
        expense_category: str | None,
        expense_category_label: str,
        risk_level: str,
        fields: list[RiskRuleField],
        created_at: datetime,
        actor: str,
        requires_attachment: bool,
        rule_title: str = "",
        risk_score: dict[str, Any] | None = None,
    ) -> dict[str, Any]:
        created_stamp = created_at.strftime("%Y%m%d%H%M%S%f")
        domain_slug = {"expense": "expense", "ar": "ar", "ap": "ap"}[domain]
        category_slug = f".{expense_category}" if expense_category else ""
        rule_code = f"risk.{domain_slug}{category_slug}.generated_{created_stamp}"
        template_key = str(draft.get("template_key") or "field_required_v1").strip()
        field_keys = [
            str(item or "").strip()
            for item in list(draft.get("field_keys") or [])
            if str(item or "").strip()
        ]
        condition_summary = (
            self._clean_text(draft.get("condition_summary")) or "判断是否符合自然语言规则描述"
        )
        risk_category = expense_category_label or BUSINESS_DOMAIN_LABELS[domain]
        risk_score_payload = dict(risk_score or {})
        risk_score_value = int(risk_score_payload.get("score") or 0)
        risk_level_label = str(
            risk_score_payload.get("level_label") or RISK_LEVEL_LABELS.get(risk_level, "风险")
        )
        keywords = list(draft.get("keywords") or [])
        field_by_key = {item.key: item for item in fields}
        params: dict[str, Any] = {
            "template_key": template_key,
            "field_keys": field_keys,
            "condition_summary": condition_summary,
            "natural_language": natural_language,
            "business_stage": business_stage,
            "business_stage_label": business_stage_label,
        }
        semantic_type = str(draft.get("semantic_type") or "").strip()
        if semantic_type:
            params["semantic_type"] = semantic_type
        if isinstance(draft.get("dsl_validation"), dict):
            params["dsl_validation"] = draft["dsl_validation"]
        if template_key == COMPOSITE_RULE_TEMPLATE_KEY and isinstance(draft.get("rule_ir"), dict):
            params["rule_ir"] = draft["rule_ir"]
            for key in ("conditions", "hit_logic", "field_groups", "formula", "message_template"):
                if key in draft:
                    params[key] = draft[key]
            for key in ("keywords", "exception_keywords", "unsupported_fields"):
                values = draft.get(key)
                if isinstance(values, list):
                    params[key] = values
        if draft.get("semantic_type") == CITY_CONSISTENCY_SEMANTIC_TYPE:
            params.update(build_city_consistency_params(draft))
        if template_key == "field_required_v1":
            params["required_fields"] = field_keys
        if template_key == "field_compare_v1" and "conditions" not in params:
            params["conditions"] = self._build_compare_conditions(field_keys)
        if template_key == "keyword_match_v1":
            params["keywords"] = keywords
            params["search_fields"] = field_keys
        applies_to: dict[str, Any] = {"domains": [domain]}
        if business_stage:
            applies_to["business_stages"] = [business_stage]
        if expense_category:
            applies_to["expense_categories"] = [expense_category]

        payload = {
            "schema_version": "2.0",
            "rule_code": rule_code,
            "name": rule_title
            or self._clean_text(draft.get("name"))
            or self._infer_rule_name(natural_language),
            "description": self._clean_text(draft.get("description")) or natural_language,
            "enabled": True,
            "requires_attachment": requires_attachment,
            "risk_dimension": "natural_language_rule",
            "risk_category": risk_category,
            "ontology_signal": "natural_language_risk",
            "evaluator": "template_rule",
            "template_key": template_key,
            "semantic_type": str(draft.get("semantic_type") or "").strip() or None,
            "applies_to": applies_to,
            "inputs": {
                "fields": [
                    {
                        "key": item.key,
                        "label": item.label,
                        "type": item.field_type,
                        "source": item.source,
                    }
                    for item in [field_by_key[key] for key in field_keys if key in field_by_key]
                ],
            },
            "params": params,
            "outcomes": {
                "pass": {"severity": "none", "action": "continue"},
                "fail": {
                    "severity": risk_level,
                    "action": "manual_review",
                    "risk_score": risk_score_value,
                },
            },
            "metadata": {
                "owner": actor,
                "stability": "generated_draft",
                "source_ref": "自然语言风险规则",
                "created_at": created_at.isoformat(),
                "created_by": actor,
                "requires_attachment": requires_attachment,
                "risk_score": risk_score_value,
                "risk_level": risk_level,
                "risk_level_label": risk_level_label,
                "risk_score_model": risk_score_payload.get("model"),
                "risk_score_detail": risk_score_payload,
                "rule_title": rule_title,
                "expense_category": expense_category,
                "expense_category_label": expense_category_label,
                "business_stage": business_stage,
                "business_stage_label": business_stage_label,
                "natural_language": natural_language,
                "business_explanation": self._clean_text(draft.get("description")),
                "condition_summary": condition_summary,
                "rule_ir": draft.get("rule_ir") if isinstance(draft.get("rule_ir"), dict) else {},
                "model_semantic_plan": draft.get("model_semantic_plan") if isinstance(draft.get("model_semantic_plan"), dict) else {},
                "flow": draft.get("flow") if isinstance(draft.get("flow"), dict) else {},
            },
        }
        explainability = build_risk_rule_explainability_artifacts(
            payload,
            fields=[field_by_key[key] for key in field_keys if key in field_by_key],
            domain_label=risk_category,
            risk_level=risk_level,
            risk_level_label=risk_level_label,
        )
        payload.update(explainability)
        payload["metadata"].update(
            {
                "semantic_plan": explainability["semantic_plan"],
                "flow_model": explainability["flow_model"],
                "flow_explanation": explainability["flow_explanation"],
                "flow_diagram_svg": explainability["flow_diagram_svg"],
            }
        )
        return payload

    @staticmethod
    def _normalize_expense_category(value: str | None, domain: str) -> str | None:
        if domain != AgentAssetDomain.EXPENSE.value:
            return None

        normalized = str(value or "").strip().lower()
        if not normalized:
            return None

        normalized = EXPENSE_RISK_CATEGORY_ALIASES.get(normalized, normalized)
        if normalized not in EXPENSE_RISK_CATEGORY_LABELS:
            allowed = "、".join(EXPENSE_RISK_CATEGORY_LABELS.values())
            raise ValueError(f"费用领域仅支持：{allowed}。")
        return normalized

    @staticmethod
    def _normalize_business_stage(value: str | None, domain: str) -> str:
        if domain != AgentAssetDomain.EXPENSE.value:
            return "reimbursement"

        normalized = str(value or "reimbursement").strip().lower()
        if not normalized:
            normalized = "reimbursement"
        if normalized not in EXPENSE_BUSINESS_STAGE_LABELS:
            allowed = "、".join(EXPENSE_BUSINESS_STAGE_LABELS.values())
            raise ValueError(f"业务环节仅支持：{allowed}。")
        return normalized

    def _resolve_fields(self, text: str, *, domain: str) -> list[RiskRuleField]:
        prefixes = DOMAIN_FIELD_PREFIXES.get(domain, ())
        candidates = [field for field in FIELD_ONTOLOGY if field.key.startswith(prefixes)]
        normalized = text.lower()
        matched: list[tuple[int, RiskRuleField]] = []
        for field in candidates:
            score = self._score_field_match(field, text, normalized)
            if score > 0:
                matched.append((score, field))

        if domain == AgentAssetDomain.EXPENSE.value:
            if any(keyword in text for keyword in ("住宿", "酒店", "行程", "城市", "出差")):
                matched.extend(
                    (10, field)
                    for field in candidates
                    if field.key
                    in {
                        "claim.reason",
                        "claim.location",
                        "employee.location",
                        "item.item_date",
                        "item.item_reason",
                        "item.item_location",
                        "attachment.hotel_city",
                        "attachment.route_cities",
                        "attachment.issue_date",
                        "attachment.stay_start_date",
                        "attachment.stay_end_date",
                    }
                )
            if any(keyword in text for keyword in ("发票", "票据", "品名", "抬头", "开票")):
                matched.extend(
                    (6, field)
                    for field in candidates
                    if field.key
                    in {
                        "attachment.invoice_no",
                        "attachment.buyer_name",
                        "attachment.goods_name",
                        "attachment.ocr_text",
                    }
                )

        matched.sort(key=lambda item: item[0], reverse=True)
        deduped: list[RiskRuleField] = []
        seen: set[str] = set()
        for _, field in matched:
            if field.key in seen:
                continue
            seen.add(field.key)
            deduped.append(field)
        if deduped:
            return deduped[:10]
        return candidates[:4]

    @staticmethod
    def _score_field_match(field: RiskRuleField, text: str, normalized: str) -> int:
        score = 0
        if field.label in text:
            score += 8
        for alias in field.aliases:
            if alias.lower() in normalized:
                score += 4 + min(len(alias), 6)

        if field.key == "attachment.hotel_city" and any(term in text for term in ("酒店", "住宿")):
            score += 12
        if field.key == "attachment.route_cities" and any(
            term in text for term in ("行程", "交通票", "路线", "途经")
        ):
            score += 10
        if field.key in {
            "claim.trip_start_date",
            "claim.trip_end_date",
            "item.item_date",
            "attachment.stay_start_date",
            "attachment.stay_end_date",
        } and any(term in text for term in ("日期", "时间", "出差开始", "出差结束", "入住", "离店")):
            score += 10
        if field.key == "claim.location" and any(
            term in text for term in ("申报目的地", "申报地点", "目的地", "出差地")
        ):
            score += 10
        if field.key.startswith("attachment.") and any(term in text for term in ("发票", "票据")):
            score += 2
        return score

    def _align_draft_fields(
        self,
        draft: dict[str, Any],
        *,
        natural_language: str,
        risk_level: str,
        fields: list[RiskRuleField],
    ) -> dict[str, Any]:
        if str(draft.get("semantic_type") or "").strip() in CITY_CONSISTENCY_SEMANTIC_TYPES:
            return build_city_consistency_draft(
                draft,
                natural_language=natural_language,
                fields=fields,
                risk_level=risk_level,
            )

        field_by_key = {field.key: field for field in fields}
        original_keys = [
            str(item or "").strip()
            for item in list(draft.get("field_keys") or [])
            if str(item or "").strip() in field_by_key
        ]
        if draft.get("template_key") == COMPOSITE_RULE_TEMPLATE_KEY:
            return {**draft, "field_keys": original_keys or [field.key for field in fields[:8]]}

        preferred_keys: list[str] = []

        def add_preferred(key: str, *terms: str) -> None:
            if key in field_by_key and any(term in natural_language for term in terms):
                preferred_keys.append(key)

        add_preferred("attachment.hotel_city", "酒店", "住宿")
        add_preferred("claim.location", "申报目的地", "申报地点", "目的地", "出差地")
        add_preferred("attachment.route_cities", "行程", "交通票", "路线", "途经")

        merged_keys: list[str] = []
        for key in [*preferred_keys, *original_keys, *[field.key for field in fields]]:
            if key in field_by_key and key not in merged_keys:
                merged_keys.append(key)
            if len(merged_keys) >= 4:
                break

        if draft.get("template_key") == "field_compare_v1" and len(merged_keys) < 2:
            for field in fields:
                if field.key not in merged_keys:
                    merged_keys.append(field.key)
                if len(merged_keys) >= 2:
                    break

        aligned = {**draft, "field_keys": merged_keys}
        selected_fields = [field_by_key[key] for key in merged_keys if key in field_by_key]
        if selected_fields:
            aligned["condition_summary"] = self._build_condition_summary(
                natural_language,
                template_key=str(aligned.get("template_key") or "field_required_v1"),
                fields=selected_fields,
            )
            flow = aligned.get("flow") if isinstance(aligned.get("flow"), dict) else {}
            aligned["flow"] = {
                **flow,
                "evidence": "读取" + "、".join(field.label for field in selected_fields[:3]),
                "decision": aligned["condition_summary"],
            }
        return aligned

    @staticmethod
    def _build_compare_conditions(field_keys: list[str]) -> list[dict[str, str]]:
        if len(field_keys) >= 2:
            return [{"left": field_keys[0], "operator": "overlap", "right": field_keys[1]}]
        if field_keys:
            return [{"left": field_keys[0], "operator": "is_empty", "right": ""}]
        return []

    @staticmethod
    def _infer_template_key(text: str) -> str:
        if any(keyword in text for keyword in ("超过", "超出", "超预算", "预算", "阈值", "早于", "晚于", "范围")):
            return COMPOSITE_RULE_TEMPLATE_KEY
        if any(
            keyword in text
            for keyword in ("一致", "匹配", "相同", "不一致", "不符", "对应", "出现在")
        ):
            return "field_compare_v1"
        if any(
            keyword in text
            for keyword in ("关键词", "包含", "出现", "品名", "摘要", "服务费", "咨询费")
        ):
            return "keyword_match_v1"
        return "field_required_v1"

    @staticmethod
    def _infer_keywords(text: str) -> list[str]:
        quoted = re.findall(r"[“\"']([^“”\"']{2,20})[”\"']", text)
        keywords = [item.strip() for item in quoted if item.strip()]
        for candidate in ("咨询费", "服务费", "其他", "办公用品", "招待", "红冲", "作废"):
            if candidate in text and candidate not in keywords:
                keywords.append(candidate)
        return keywords[:8]

    @staticmethod
    def _infer_rule_name(text: str) -> str:
        normalized = re.sub(r"\s+", "", str(text or ""))
        normalized = re.sub(r"[，。；;：:、,.!?！？]", "", normalized)
        if not normalized:
            return "自然语言风险规则"
        return f"{normalized[:18]}风险规则"

    @staticmethod
    def _build_condition_summary(
        natural_language: str,
        *,
        template_key: str,
        fields: list[RiskRuleField],
    ) -> str:
        field_text = "、".join(item.label for item in fields[:3]) or "业务字段"
        if template_key == "field_compare_v1":
            return f"对比{field_text}之间是否一致或存在交集"
        if template_key == "keyword_match_v1":
            return f"检查{field_text}是否出现规则描述中的风险关键词"
        return f"检查{field_text}是否满足必填和完整性要求"

    @staticmethod
    def _clean_text(value: Any) -> str:
        return re.sub(r"\s+", " ", str(value or "")).strip()

    @staticmethod
    def _extract_json_object(text: str) -> str:
        normalized = re.sub(r"^```(?:json)?|```$", "", str(text or "").strip(), flags=re.IGNORECASE)
        start = normalized.find("{")
        end = normalized.rfind("}")
        if start < 0 or end <= start:
            raise ValueError("JSON object not found.")
        return normalized[start : end + 1]