feat: close phase15 governance loops
This commit is contained in:
@@ -77,7 +77,10 @@ def recommend_missing_clauses(
|
||||
"document_type_code": result.get("document_type_code"),
|
||||
"contract_type": payload.contract_type,
|
||||
"recommended_count": result.get("total"),
|
||||
"recommendation_keys": (result.get("summary") or {}).get("recommendation_keys") or [],
|
||||
"missing_clause_keys": (result.get("summary") or {}).get("missing_clause_keys") or [],
|
||||
"industry_practice_keys": (result.get("summary") or {}).get("industry_practice_keys") or [],
|
||||
"industry_practice_count": (result.get("summary") or {}).get("industry_practice_count") or 0,
|
||||
"source_count": (result.get("summary") or {}).get("source_count") or 0,
|
||||
"text_length": len(payload.contract_text or ""),
|
||||
},
|
||||
|
||||
@@ -105,6 +105,8 @@ def scan_legal_change_impact(
|
||||
"article_id": payload.article_id,
|
||||
"change_type": result["event"]["change_type"],
|
||||
"impact_summary": result["event"]["impact_summary"],
|
||||
"semantic_match_count": result["event"]["impact_summary"].get("semantic_match_count", 0),
|
||||
"scan_strategy": result["event"]["impact_summary"].get("scan_strategy", []),
|
||||
},
|
||||
ip_address=request.client.host if request.client else None,
|
||||
user_agent=request.headers.get("user-agent"),
|
||||
@@ -127,7 +129,11 @@ def run_legal_change_event_scan(
|
||||
action="run_legal_change_event_scan",
|
||||
resource_type="legal_change_event",
|
||||
resource_id=event_id,
|
||||
detail_json={"impact_summary": result["event"]["impact_summary"]},
|
||||
detail_json={
|
||||
"impact_summary": result["event"]["impact_summary"],
|
||||
"semantic_match_count": result["event"]["impact_summary"].get("semantic_match_count", 0),
|
||||
"scan_strategy": result["event"]["impact_summary"].get("scan_strategy", []),
|
||||
},
|
||||
ip_address=request.client.host if request.client else None,
|
||||
user_agent=request.headers.get("user-agent"),
|
||||
)
|
||||
@@ -153,6 +159,7 @@ def run_pending_legal_change_scans(
|
||||
"requested_limit": result["requested_limit"],
|
||||
"completed_count": result["completed_count"],
|
||||
"failed_count": result["failed_count"],
|
||||
"worker_compatible": True,
|
||||
},
|
||||
ip_address=request.client.host if request.client else None,
|
||||
user_agent=request.headers.get("user-agent"),
|
||||
|
||||
@@ -178,6 +178,10 @@ def run_prompt_asset_evaluation(
|
||||
"evaluation_status": result["version"]["evaluation_status"],
|
||||
"evaluation_score": result["version"]["evaluation_score"],
|
||||
"suite_key": result["evaluation_report"].get("suite_key"),
|
||||
"evaluation_mode": result["evaluation_report"].get("evaluation_mode"),
|
||||
"sample_count": result["evaluation_report"].get("sample_count"),
|
||||
"passed_samples": result["evaluation_report"].get("passed_samples"),
|
||||
"failed_samples": result["evaluation_report"].get("failed_samples"),
|
||||
},
|
||||
)
|
||||
return success_response(result)
|
||||
|
||||
@@ -24,6 +24,14 @@ class Settings(BaseSettings):
|
||||
shared_ai_platform_timeout_seconds: float = 30.0
|
||||
shared_ai_platform_legal_review_enabled: bool = True
|
||||
shared_ai_platform_legal_review_fallback_to_local: bool = False
|
||||
prompt_golden_eval_mode: str = "auto"
|
||||
prompt_golden_eval_include_raw_model_response: bool = False
|
||||
legal_change_impact_semantic_enabled: bool = True
|
||||
legal_change_impact_semantic_threshold: float = 0.10
|
||||
legal_change_impact_worker_enabled: bool = True
|
||||
legal_change_impact_worker_batch_size: int = 10
|
||||
legal_change_impact_worker_interval_seconds: int = 300
|
||||
legal_change_impact_worker_user_id: str = "system-legal-change-worker"
|
||||
secret_key: str = "change-me"
|
||||
access_token_expire_minutes: int = 120
|
||||
database_url: str = "postgresql+psycopg://postgres:postgres@localhost:5432/ai_contract_platform"
|
||||
|
||||
@@ -81,6 +81,15 @@ class LegalChangeImpactRepository:
|
||||
.all()
|
||||
)
|
||||
|
||||
def list_pending_events_for_worker(self, limit: int = 10) -> list[LegalChangeEvent]:
|
||||
return (
|
||||
self.db.query(LegalChangeEvent)
|
||||
.filter(LegalChangeEvent.scan_status == "pending")
|
||||
.order_by(LegalChangeEvent.created_at.asc())
|
||||
.limit(max(1, min(limit, 100)))
|
||||
.all()
|
||||
)
|
||||
|
||||
def delete_impacts_by_event(self, tenant_id: str, event_id: str) -> int:
|
||||
deleted = (
|
||||
self.db.query(LegalChangeImpact)
|
||||
|
||||
@@ -498,6 +498,199 @@ DOCUMENT_TYPE_RULES["software_development"] = [
|
||||
},
|
||||
]
|
||||
|
||||
INDUSTRY_PRACTICE_RULES: dict[str, list[dict[str, Any]]] = {
|
||||
"service_general": [
|
||||
{
|
||||
"clause_key": "change_order_confirmation",
|
||||
"clause_title": "服务范围变更确认",
|
||||
"clause_type": "industry_practice",
|
||||
"priority": "medium",
|
||||
"keywords": ["变更单", "服务范围变更", "change order", "需求变更"],
|
||||
"anchor_keywords": ["服务范围", "交付物", "验收"],
|
||||
"suggested_text": (
|
||||
"服务范围变更确认：服务内容、交付物、实施计划或费用口径发生变更的，双方应通过"
|
||||
"书面补充协议、变更单或双方确认的电子流程确认后执行；未经确认的变更不当然构成"
|
||||
"乙方新增义务或甲方付款义务。"
|
||||
),
|
||||
"reason": "服务项目中范围变更高频发生，提前约定变更确认方式能减少口头需求和费用争议。",
|
||||
"legal_basis": "无直接法条，仅为行业惯例；建议结合具体服务范围、审批流程和费用调整机制确认。",
|
||||
"practice_basis": "服务类项目通常通过变更单、补充协议或可追溯电子确认固化新增需求、延期和费用调整。",
|
||||
"source_rule": "LAW-CLAUSE-PRACTICE-SERVICE-CHANGE-ORDER",
|
||||
"default_position": "建议插入在服务范围或交付物条款之后。",
|
||||
"recommendation_kind": "industry_practice",
|
||||
"industry_context": "通用服务",
|
||||
"basis_type": "industry_practice_only",
|
||||
},
|
||||
{
|
||||
"clause_key": "service_progress_record",
|
||||
"clause_title": "服务过程记录与阶段确认",
|
||||
"clause_type": "industry_practice",
|
||||
"priority": "low",
|
||||
"keywords": ["服务记录", "周报", "月报", "阶段确认", "meeting minutes"],
|
||||
"anchor_keywords": ["服务范围", "验收", "付款"],
|
||||
"suggested_text": (
|
||||
"服务过程记录与阶段确认：乙方应按约定周期提交服务记录、会议纪要、阶段成果或"
|
||||
"工作报告；甲方应在收到后合理期限内确认或提出书面异议，逾期未反馈的，双方可"
|
||||
"按约定视为阶段事项已完成记录。"
|
||||
),
|
||||
"reason": "持续服务合同往往缺少可验证过程证据，阶段记录可以支撑验收、付款和争议处理。",
|
||||
"legal_basis": "无直接法条，仅为行业惯例；建议结合双方沟通机制和验收触发条件设置。",
|
||||
"practice_basis": "咨询、运维、外包服务通常用周报、月报、会议纪要或阶段确认单沉淀履约证据。",
|
||||
"source_rule": "LAW-CLAUSE-PRACTICE-SERVICE-PROGRESS-RECORD",
|
||||
"default_position": "建议插入在服务范围或验收条款之后。",
|
||||
"recommendation_kind": "industry_practice",
|
||||
"industry_context": "持续服务",
|
||||
"basis_type": "industry_practice_only",
|
||||
},
|
||||
],
|
||||
"technical_service": [
|
||||
{
|
||||
"clause_key": "technical_assumptions_boundary",
|
||||
"clause_title": "技术前提与接口边界",
|
||||
"clause_type": "industry_practice",
|
||||
"priority": "medium",
|
||||
"keywords": ["技术前提", "接口边界", "系统环境", "第三方接口", "technical assumptions"],
|
||||
"anchor_keywords": ["服务范围", "交付物", "验收"],
|
||||
"suggested_text": (
|
||||
"技术前提与接口边界：双方应在附件中列明系统环境、接口权限、第三方平台、数据格式"
|
||||
"和测试条件等技术前提；因前提条件变化、第三方接口调整或甲方环境不满足要求导致"
|
||||
"延期或返工的，双方应按变更流程处理。"
|
||||
),
|
||||
"reason": "技术服务常受系统环境、接口和第三方平台影响，提前固化边界能降低履约争议。",
|
||||
"legal_basis": "无直接法条，仅为行业惯例；建议结合项目技术附件和变更流程确认。",
|
||||
"practice_basis": "技术服务项目通常用技术附件、接口清单、环境确认单和测试记录管理履约前提。",
|
||||
"source_rule": "LAW-CLAUSE-PRACTICE-TECH-ASSUMPTIONS",
|
||||
"default_position": "建议插入在服务范围或交付物条款之后。",
|
||||
"recommendation_kind": "industry_practice",
|
||||
"industry_context": "技术服务",
|
||||
"basis_type": "industry_practice_only",
|
||||
},
|
||||
],
|
||||
"software_development": [
|
||||
{
|
||||
"clause_key": "scope_change_control",
|
||||
"clause_title": "需求变更与版本范围控制",
|
||||
"clause_type": "industry_practice",
|
||||
"priority": "high",
|
||||
"keywords": ["需求变更", "版本范围", "迭代范围", "需求确认", "scope change"],
|
||||
"anchor_keywords": ["服务范围", "交付物", "验收", "维护支持"],
|
||||
"suggested_text": (
|
||||
"需求变更与版本范围控制：软件功能、页面、接口、数据结构、第三方集成或验收标准"
|
||||
"发生新增、删减或调整的，双方应通过需求变更单、迭代确认单或可追溯电子流程确认"
|
||||
"范围、工期、费用和验收影响后实施。"
|
||||
),
|
||||
"reason": "软件开发项目中需求变更最容易导致延期和费用争议，建议把变更流程写入合同正文或附件。",
|
||||
"legal_basis": "无直接法条，仅为行业惯例；建议结合项目管理流程、需求清单和验收标准确认。",
|
||||
"practice_basis": "软件开发行业通常用需求基线、迭代计划、变更单和版本验收清单控制范围蔓延。",
|
||||
"source_rule": "LAW-CLAUSE-PRACTICE-SOFTWARE-SCOPE-CHANGE",
|
||||
"default_position": "建议插入在服务范围、交付物或验收条款之后。",
|
||||
"recommendation_kind": "industry_practice",
|
||||
"industry_context": "软件开发",
|
||||
"basis_type": "industry_practice_only",
|
||||
},
|
||||
{
|
||||
"clause_key": "open_source_third_party_components",
|
||||
"clause_title": "开源及第三方组件清单",
|
||||
"clause_type": "industry_practice",
|
||||
"priority": "medium",
|
||||
"keywords": ["开源组件", "第三方组件", "开源软件", "组件清单", "open source"],
|
||||
"anchor_keywords": ["知识产权", "交付物", "维护支持"],
|
||||
"suggested_text": (
|
||||
"开源及第三方组件清单：乙方交付的软件如使用开源软件、第三方 SDK、字体、模型、"
|
||||
"插件或云服务，应向甲方提供组件清单、许可类型、使用范围及已知限制；涉及额外授权、"
|
||||
"费用或合规义务的，应由双方另行确认承担方式。"
|
||||
),
|
||||
"reason": "软件交付中第三方组件会影响授权、后续商用和安全维护，建议在合同中要求清单化披露。",
|
||||
"legal_basis": "无直接法条，仅为行业惯例；如涉及著作权、商业许可或数据处理，应另行结合法律依据审查。",
|
||||
"practice_basis": "软件交付通常要求 SBOM、第三方组件清单或开源许可披露，以便后续运维和合规审计。",
|
||||
"source_rule": "LAW-CLAUSE-PRACTICE-SOFTWARE-THIRD-PARTY",
|
||||
"default_position": "建议插入在知识产权或交付物条款之后。",
|
||||
"recommendation_kind": "industry_practice",
|
||||
"industry_context": "软件开发",
|
||||
"basis_type": "industry_practice_only",
|
||||
},
|
||||
],
|
||||
"labor_contract": [
|
||||
{
|
||||
"clause_key": "employee_handbook_acknowledgement",
|
||||
"clause_title": "规章制度告知与签收",
|
||||
"clause_type": "industry_practice",
|
||||
"priority": "medium",
|
||||
"keywords": ["规章制度", "员工手册", "制度签收", "培训确认"],
|
||||
"anchor_keywords": ["岗位", "劳动纪律", "解除", "试用期"],
|
||||
"suggested_text": (
|
||||
"规章制度告知与签收：甲方依法制定并公示的员工手册、考勤、奖惩、信息安全、保密"
|
||||
"和岗位管理制度作为劳动管理依据；乙方确认已通过书面、电子签收或培训记录知悉相关制度。"
|
||||
),
|
||||
"reason": "劳动合同履行中，制度告知和签收记录是后续考核、奖惩和争议处理的重要证据。",
|
||||
"legal_basis": "无直接法条，仅为行业惯例；涉及规章制度适用时仍需满足民主程序、公示告知和合法合理要求。",
|
||||
"practice_basis": "企业人力资源管理通常通过员工手册签收、入职培训记录和制度更新通知沉淀管理证据。",
|
||||
"source_rule": "LAW-CLAUSE-PRACTICE-LABOR-HANDBOOK",
|
||||
"default_position": "建议插入在岗位职责、劳动纪律或解除终止条款附近。",
|
||||
"recommendation_kind": "industry_practice",
|
||||
"industry_context": "劳动人事",
|
||||
"basis_type": "industry_practice_only",
|
||||
},
|
||||
{
|
||||
"clause_key": "handover_materials_record",
|
||||
"clause_title": "离职交接清单",
|
||||
"clause_type": "industry_practice",
|
||||
"priority": "low",
|
||||
"keywords": ["离职交接", "交接清单", "资料返还", "账号移交"],
|
||||
"anchor_keywords": ["解除", "终止", "保密", "竞业限制"],
|
||||
"suggested_text": (
|
||||
"离职交接清单：劳动合同解除或终止时，乙方应按甲方合理要求办理工作资料、设备、"
|
||||
"账号权限、客户信息和未结事项交接；双方可通过交接清单、系统记录或书面确认留存证据。"
|
||||
),
|
||||
"reason": "离职交接是劳动争议和商业资料保护的高频风险点，清单化处理更便于执行。",
|
||||
"legal_basis": "无直接法条，仅为行业惯例；涉及工资结算、经济补偿或保密义务时应另行依法处理。",
|
||||
"practice_basis": "企业通常用离职交接单、资产归还记录和账号权限回收记录完成离职闭环。",
|
||||
"source_rule": "LAW-CLAUSE-PRACTICE-LABOR-HANDOVER",
|
||||
"default_position": "建议插入在解除终止或保密条款之后。",
|
||||
"recommendation_kind": "industry_practice",
|
||||
"industry_context": "劳动人事",
|
||||
"basis_type": "industry_practice_only",
|
||||
},
|
||||
],
|
||||
"nda_commercial": [
|
||||
{
|
||||
"clause_key": "need_to_know_access",
|
||||
"clause_title": "最小知悉范围",
|
||||
"clause_type": "industry_practice",
|
||||
"priority": "medium",
|
||||
"keywords": ["最小知悉", "need to know", "访问权限", "授权人员"],
|
||||
"anchor_keywords": ["保密义务", "保密信息", "披露"],
|
||||
"suggested_text": (
|
||||
"最小知悉范围：接收方仅可在实现合作目的所必需的范围内向其员工、顾问或关联人员"
|
||||
"披露保密信息，并应确保接触人员承担不低于本协议约定的保密义务。"
|
||||
),
|
||||
"reason": "保密协议通常需要控制内部流转范围，避免保密信息因过度共享而失控。",
|
||||
"legal_basis": "无直接法条，仅为行业惯例；涉及商业秘密保护时应结合保密措施和访问控制证据确认。",
|
||||
"practice_basis": "商业合作中常按 need-to-know 原则控制保密信息访问权限并留存授权记录。",
|
||||
"source_rule": "LAW-CLAUSE-PRACTICE-NDA-NEED-TO-KNOW",
|
||||
"default_position": "建议插入在保密义务或保密信息定义之后。",
|
||||
"recommendation_kind": "industry_practice",
|
||||
"industry_context": "保密合作",
|
||||
"basis_type": "industry_practice_only",
|
||||
},
|
||||
],
|
||||
}
|
||||
|
||||
SERVICE_GENERAL_PRACTICE_RULES = INDUSTRY_PRACTICE_RULES["service_general"]
|
||||
TECHNICAL_SERVICE_BASE_PRACTICE_RULES = INDUSTRY_PRACTICE_RULES["technical_service"]
|
||||
SOFTWARE_DEVELOPMENT_BASE_PRACTICE_RULES = INDUSTRY_PRACTICE_RULES["software_development"]
|
||||
TECHNICAL_SERVICE_PRACTICE_RULES = [
|
||||
*TECHNICAL_SERVICE_BASE_PRACTICE_RULES,
|
||||
*SERVICE_GENERAL_PRACTICE_RULES,
|
||||
]
|
||||
SOFTWARE_DEVELOPMENT_PRACTICE_RULES = [
|
||||
*SOFTWARE_DEVELOPMENT_BASE_PRACTICE_RULES,
|
||||
*TECHNICAL_SERVICE_BASE_PRACTICE_RULES,
|
||||
*SERVICE_GENERAL_PRACTICE_RULES,
|
||||
]
|
||||
INDUSTRY_PRACTICE_RULES["technical_service"] = TECHNICAL_SERVICE_PRACTICE_RULES
|
||||
INDUSTRY_PRACTICE_RULES["software_development"] = SOFTWARE_DEVELOPMENT_PRACTICE_RULES
|
||||
|
||||
|
||||
class ClauseRecommendationService:
|
||||
def __init__(self, db: Optional[Session], legal_article_service: Optional[LegalArticleService] = None):
|
||||
@@ -531,6 +724,7 @@ class ClauseRecommendationService:
|
||||
max_count = self._normalize_max_items(max_items)
|
||||
|
||||
detected_keys: list[str] = []
|
||||
detected_industry_practice_keys: list[str] = []
|
||||
recommendations: list[dict[str, Any]] = []
|
||||
for rule in rules:
|
||||
if self._rule_is_present(normalized_text, rule):
|
||||
@@ -548,7 +742,35 @@ class ClauseRecommendationService:
|
||||
if len(recommendations) >= max_count:
|
||||
break
|
||||
|
||||
missing_keys = [item["clause_key"] for item in recommendations]
|
||||
if len(recommendations) < max_count:
|
||||
for rule in self._industry_practice_rules_for_document_type(normalized_document_type):
|
||||
if self._rule_is_present(normalized_text, rule):
|
||||
detected_industry_practice_keys.append(rule["clause_key"])
|
||||
continue
|
||||
legal_article_refs = self._search_legal_articles(current_user, rule)
|
||||
recommendations.append(
|
||||
self._build_recommendation_item(
|
||||
contract_text=normalized_text,
|
||||
document_type_code=normalized_document_type,
|
||||
rule=rule,
|
||||
legal_article_refs=legal_article_refs,
|
||||
)
|
||||
)
|
||||
if len(recommendations) >= max_count:
|
||||
break
|
||||
|
||||
recommendation_keys = [item["clause_key"] for item in recommendations]
|
||||
required_missing_keys = [
|
||||
item["clause_key"]
|
||||
for item in recommendations
|
||||
if item.get("recommendation_kind") != "industry_practice"
|
||||
]
|
||||
missing_keys = required_missing_keys
|
||||
industry_practice_keys = [
|
||||
item["clause_key"]
|
||||
for item in recommendations
|
||||
if item.get("recommendation_kind") == "industry_practice"
|
||||
]
|
||||
result = {
|
||||
"contract_id": contract_id,
|
||||
"document_type_code": normalized_document_type,
|
||||
@@ -558,16 +780,22 @@ class ClauseRecommendationService:
|
||||
"total": len(recommendations),
|
||||
"summary": {
|
||||
"recommended_count": len(recommendations),
|
||||
"recommendation_keys": recommendation_keys,
|
||||
"detected_clause_keys": detected_keys,
|
||||
"missing_clause_keys": missing_keys,
|
||||
"required_missing_clause_keys": required_missing_keys,
|
||||
"industry_practice_keys": industry_practice_keys,
|
||||
"detected_industry_practice_keys": detected_industry_practice_keys,
|
||||
"industry_practice_count": len(industry_practice_keys),
|
||||
"source_count": sum(len(item.get("legal_article_refs") or []) for item in recommendations),
|
||||
"strategy": "rule_based_required_clause_gap",
|
||||
"strategy": "rule_based_required_clause_gap_and_industry_practice",
|
||||
},
|
||||
"sources": [
|
||||
"document_type_required_clause_list",
|
||||
"standard_clause_library",
|
||||
"legal_articles",
|
||||
"review_rules",
|
||||
"industry_practice_rules",
|
||||
],
|
||||
}
|
||||
self.prompt_runtime_service.record_run_link(
|
||||
@@ -581,7 +809,10 @@ class ClauseRecommendationService:
|
||||
"document_type_code": normalized_document_type,
|
||||
"contract_type": contract_type,
|
||||
"recommended_count": len(recommendations),
|
||||
"recommendation_keys": recommendation_keys,
|
||||
"missing_clause_keys": missing_keys,
|
||||
"required_missing_clause_keys": required_missing_keys,
|
||||
"industry_practice_keys": industry_practice_keys,
|
||||
"text_length": len(normalized_text),
|
||||
},
|
||||
)
|
||||
@@ -705,20 +936,19 @@ class ClauseRecommendationService:
|
||||
"clause_key": rule["clause_key"],
|
||||
"clause_title": rule["clause_title"],
|
||||
"clause_type": rule["clause_type"],
|
||||
"recommendation_kind": rule.get("recommendation_kind", "missing_clause"),
|
||||
"basis_type": rule.get("basis_type") or ("legal_basis" if legal_article_refs else "rule_fallback"),
|
||||
"industry_context": rule.get("industry_context"),
|
||||
"priority": rule.get("priority", "medium"),
|
||||
"confidence": "high" if legal_article_refs else "medium",
|
||||
"insert_position": insertion_position,
|
||||
"suggested_text": rule["suggested_text"],
|
||||
"reason": rule["reason"],
|
||||
"legal_basis": cls._build_legal_basis(rule, legal_article_refs),
|
||||
"practice_basis": rule.get("practice_basis"),
|
||||
"legal_article_refs": legal_article_refs,
|
||||
"source_rule": rule["source_rule"],
|
||||
"source_types": [
|
||||
"document_type_required_clause_list",
|
||||
"standard_clause_library",
|
||||
"legal_articles" if legal_article_refs else "legal_basis_fallback",
|
||||
"review_rules",
|
||||
],
|
||||
"source_types": cls._build_recommendation_source_types(rule, legal_article_refs),
|
||||
"detected_keywords": [],
|
||||
}
|
||||
|
||||
@@ -742,6 +972,24 @@ class ClauseRecommendationService:
|
||||
return str(rule.get("legal_basis") or "")
|
||||
return f"{rule.get('legal_basis') or '建议结合法条来源核对。'} 关联依据：{'；'.join(labels)}。"
|
||||
|
||||
@staticmethod
|
||||
def _build_recommendation_source_types(rule: dict[str, Any], legal_article_refs: list[dict[str, Any]]) -> list[str]:
|
||||
if rule.get("recommendation_kind") == "industry_practice":
|
||||
source_types = [
|
||||
"industry_practice_rules",
|
||||
"standard_clause_library",
|
||||
"legal_articles" if legal_article_refs else "industry_practice_fallback",
|
||||
]
|
||||
if rule.get("basis_type") == "industry_practice_only":
|
||||
source_types.append("no_direct_legal_article")
|
||||
return source_types
|
||||
return [
|
||||
"document_type_required_clause_list",
|
||||
"standard_clause_library",
|
||||
"legal_articles" if legal_article_refs else "legal_basis_fallback",
|
||||
"review_rules",
|
||||
]
|
||||
|
||||
@classmethod
|
||||
def _resolve_insertion_position(cls, contract_text: str, rule: dict[str, Any]) -> dict[str, Any]:
|
||||
anchor = cls._find_anchor_line(contract_text, rule.get("anchor_keywords") or [])
|
||||
@@ -989,6 +1237,8 @@ class ClauseRecommendationService:
|
||||
def _search_legal_articles(self, current_user: dict[str, Any], rule: dict[str, Any]) -> list[dict[str, Any]]:
|
||||
if self.legal_article_service is None:
|
||||
return []
|
||||
if rule.get("basis_type") == "industry_practice_only" and not str(rule.get("legal_basis_query") or "").strip():
|
||||
return []
|
||||
query = str(rule.get("legal_basis_query") or rule.get("clause_title") or "").strip()
|
||||
if not query:
|
||||
return []
|
||||
@@ -1024,6 +1274,10 @@ class ClauseRecommendationService:
|
||||
return rules
|
||||
return GENERAL_RULES
|
||||
|
||||
@staticmethod
|
||||
def _industry_practice_rules_for_document_type(document_type_code: str) -> list[dict[str, Any]]:
|
||||
return INDUSTRY_PRACTICE_RULES.get(document_type_code, [])
|
||||
|
||||
@staticmethod
|
||||
def _normalize_document_type(document_type_code: Optional[str], contract_type: Optional[str]) -> str:
|
||||
raw = (document_type_code or contract_type or "general").strip().lower().replace("-", "_")
|
||||
|
||||
@@ -8,6 +8,7 @@ from typing import Any, Iterable, Optional
|
||||
from fastapi import HTTPException, status
|
||||
from sqlalchemy.orm import Session
|
||||
|
||||
from app.core.config import settings
|
||||
from app.models.contract_template import ContractTemplate
|
||||
from app.models.generated_contract import GeneratedContract
|
||||
from app.models.legal_article import LegalArticle
|
||||
@@ -16,6 +17,7 @@ from app.models.legal_change_impact import LegalChangeImpact
|
||||
from app.models.review_rule import ReviewRule
|
||||
from app.repositories.legal_article_repository import LegalArticleRepository
|
||||
from app.repositories.legal_change_impact_repository import LegalChangeImpactRepository
|
||||
from app.services.legal_semantic_matcher import LegalSemanticMatch, LegalSemanticMatcher
|
||||
|
||||
|
||||
VALID_CHANGE_TYPES = {"created", "updated", "repealed", "effective_changed"}
|
||||
@@ -27,6 +29,7 @@ class LegalChangeImpactService:
|
||||
self.db = db
|
||||
self.repository = LegalChangeImpactRepository(db)
|
||||
self.legal_article_repository = LegalArticleRepository(db)
|
||||
self.semantic_matcher = LegalSemanticMatcher(threshold=settings.legal_change_impact_semantic_threshold)
|
||||
|
||||
def list_events(
|
||||
self,
|
||||
@@ -219,6 +222,15 @@ class LegalChangeImpactService:
|
||||
direct_terms = self._unique_terms([article.source_id, article.source_title, article.article_no])
|
||||
tag_terms = self._unique_terms(article.tags_json or [])
|
||||
article_keywords = self._extract_article_keywords(article.article_text)
|
||||
article_semantic_text = self._join_text(
|
||||
event.change_summary,
|
||||
article.source_title,
|
||||
article.article_no,
|
||||
article.chapter_title,
|
||||
article.section_title,
|
||||
article.article_text,
|
||||
article.tags_json,
|
||||
)
|
||||
impacts: list[LegalChangeImpact] = []
|
||||
|
||||
templates = self.db.query(ContractTemplate).filter(ContractTemplate.tenant_id == tenant_id).all()
|
||||
@@ -237,7 +249,8 @@ class LegalChangeImpactService:
|
||||
matched_direct = self._matched_terms(text, direct_terms)
|
||||
matched_tags = self._matched_terms(text, tag_terms)
|
||||
matched_keywords = self._matched_terms(text, article_keywords[:6])
|
||||
if matched_direct or matched_tags or matched_keywords:
|
||||
semantic_match = self._semantic_match(article_semantic_text, text)
|
||||
if matched_direct or matched_tags or matched_keywords or semantic_match.matched:
|
||||
impact_level = "immediate_update" if matched_direct else "reference_only"
|
||||
impacts.append(
|
||||
self._new_impact(
|
||||
@@ -247,11 +260,14 @@ class LegalChangeImpactService:
|
||||
target_id=template.id,
|
||||
target_name=template.name,
|
||||
target_version_label=template.current_version_label,
|
||||
match_terms=self._unique_terms([*matched_direct, *matched_tags, *matched_keywords]),
|
||||
match_reason=self._build_match_reason(matched_direct, matched_tags, matched_keywords),
|
||||
match_terms=self._unique_terms([*matched_direct, *matched_tags, *matched_keywords, *semantic_match.shared_terms]),
|
||||
match_reason=self._build_match_reason(matched_direct, matched_tags, matched_keywords, semantic_match),
|
||||
impact_level=impact_level,
|
||||
suggestion=self._build_suggestion("contract_template", impact_level, article),
|
||||
evidence={"template_code": template.template_code, "document_type_code": template.document_type_code},
|
||||
evidence=self._build_evidence(
|
||||
{"template_code": template.template_code, "document_type_code": template.document_type_code},
|
||||
semantic_match,
|
||||
),
|
||||
)
|
||||
)
|
||||
|
||||
@@ -261,7 +277,8 @@ class LegalChangeImpactService:
|
||||
matched_direct = self._matched_terms(text, direct_terms)
|
||||
matched_tags = self._matched_terms(text, tag_terms)
|
||||
matched_keywords = self._matched_terms(text, article_keywords[:8])
|
||||
if matched_direct or matched_tags or matched_keywords:
|
||||
semantic_match = self._semantic_match(article_semantic_text, text)
|
||||
if matched_direct or matched_tags or matched_keywords or semantic_match.matched:
|
||||
impacts.append(
|
||||
self._new_impact(
|
||||
tenant_id,
|
||||
@@ -270,11 +287,11 @@ class LegalChangeImpactService:
|
||||
target_id=rule.id,
|
||||
target_name=rule.name,
|
||||
target_version_label=rule.current_version_label,
|
||||
match_terms=self._unique_terms([*matched_direct, *matched_tags, *matched_keywords]),
|
||||
match_reason=self._build_match_reason(matched_direct, matched_tags, matched_keywords),
|
||||
match_terms=self._unique_terms([*matched_direct, *matched_tags, *matched_keywords, *semantic_match.shared_terms]),
|
||||
match_reason=self._build_match_reason(matched_direct, matched_tags, matched_keywords, semantic_match),
|
||||
impact_level="immediate_update",
|
||||
suggestion=self._build_suggestion("review_rule", "immediate_update", article),
|
||||
evidence={"rule_code": rule.rule_code, "risk_level": rule.risk_level},
|
||||
evidence=self._build_evidence({"rule_code": rule.rule_code, "risk_level": rule.risk_level}, semantic_match),
|
||||
)
|
||||
)
|
||||
|
||||
@@ -293,7 +310,8 @@ class LegalChangeImpactService:
|
||||
matched_direct = self._matched_terms(text, direct_terms)
|
||||
matched_tags = self._matched_terms(text, tag_terms)
|
||||
matched_keywords = self._matched_terms(text, article_keywords[:6])
|
||||
if matched_direct or matched_tags or matched_keywords:
|
||||
semantic_match = self._semantic_match(article_semantic_text, text)
|
||||
if matched_direct or matched_tags or matched_keywords or semantic_match.matched:
|
||||
impact_level = "next_renewal_update" if matched_direct else "reference_only"
|
||||
impacts.append(
|
||||
self._new_impact(
|
||||
@@ -303,11 +321,14 @@ class LegalChangeImpactService:
|
||||
target_id=contract.id,
|
||||
target_name=contract.file_name,
|
||||
target_version_label=None,
|
||||
match_terms=self._unique_terms([*matched_direct, *matched_tags, *matched_keywords]),
|
||||
match_reason=self._build_match_reason(matched_direct, matched_tags, matched_keywords),
|
||||
match_terms=self._unique_terms([*matched_direct, *matched_tags, *matched_keywords, *semantic_match.shared_terms]),
|
||||
match_reason=self._build_match_reason(matched_direct, matched_tags, matched_keywords, semantic_match),
|
||||
impact_level=impact_level,
|
||||
suggestion=self._build_suggestion("generated_contract", impact_level, article),
|
||||
evidence={"template_id": contract.template_id, "document_type_code": contract.document_type_code},
|
||||
evidence=self._build_evidence(
|
||||
{"template_id": contract.template_id, "document_type_code": contract.document_type_code},
|
||||
semantic_match,
|
||||
),
|
||||
)
|
||||
)
|
||||
|
||||
@@ -351,18 +372,34 @@ class LegalChangeImpactService:
|
||||
def _build_impact_summary(impacts: list[LegalChangeImpact]) -> dict[str, Any]:
|
||||
by_level: dict[str, int] = {}
|
||||
by_target_type: dict[str, int] = {}
|
||||
by_match_reason: dict[str, int] = {}
|
||||
semantic_match_count = 0
|
||||
for impact in impacts:
|
||||
by_level[impact.impact_level] = by_level.get(impact.impact_level, 0) + 1
|
||||
by_target_type[impact.target_type] = by_target_type.get(impact.target_type, 0) + 1
|
||||
for reason in str(impact.match_reason or "").split(","):
|
||||
if not reason:
|
||||
continue
|
||||
by_match_reason[reason] = by_match_reason.get(reason, 0) + 1
|
||||
if reason == "semantic_vector_match":
|
||||
semantic_match_count += 1
|
||||
return {
|
||||
"total": len(impacts),
|
||||
"by_level": by_level,
|
||||
"by_target_type": by_target_type,
|
||||
"by_match_reason": by_match_reason,
|
||||
"blocking_count": by_level.get("immediate_update", 0),
|
||||
"semantic_match_count": semantic_match_count,
|
||||
"scan_strategy": ["direct_reference", "tag_match", "keyword_match", "semantic_vector_match"],
|
||||
}
|
||||
|
||||
@staticmethod
|
||||
def _build_match_reason(direct_terms: list[str], tag_terms: list[str], keyword_terms: list[str]) -> str:
|
||||
def _build_match_reason(
|
||||
direct_terms: list[str],
|
||||
tag_terms: list[str],
|
||||
keyword_terms: list[str],
|
||||
semantic_match: Optional[LegalSemanticMatch] = None,
|
||||
) -> str:
|
||||
reasons: list[str] = []
|
||||
if direct_terms:
|
||||
reasons.append("direct_legal_reference")
|
||||
@@ -370,8 +407,26 @@ class LegalChangeImpactService:
|
||||
reasons.append("tag_match")
|
||||
if keyword_terms:
|
||||
reasons.append("keyword_match")
|
||||
if semantic_match and semantic_match.matched:
|
||||
reasons.append("semantic_vector_match")
|
||||
return ",".join(reasons) or "unknown"
|
||||
|
||||
def _semantic_match(self, source_text: str, target_text: str) -> LegalSemanticMatch:
|
||||
if not settings.legal_change_impact_semantic_enabled:
|
||||
return LegalSemanticMatch(False, 0.0, [], [])
|
||||
return self.semantic_matcher.match(source_text, target_text)
|
||||
|
||||
@staticmethod
|
||||
def _build_evidence(base: dict[str, Any], semantic_match: LegalSemanticMatch) -> dict[str, Any]:
|
||||
evidence = dict(base)
|
||||
if semantic_match.matched:
|
||||
evidence["semantic_vector"] = {
|
||||
"score": semantic_match.score,
|
||||
"shared_terms": semantic_match.shared_terms,
|
||||
"shared_concepts": semantic_match.shared_concepts,
|
||||
}
|
||||
return evidence
|
||||
|
||||
@staticmethod
|
||||
def _build_suggestion(target_type: str, impact_level: str, article: LegalArticle) -> str:
|
||||
article_label = f"{article.source_title}{article.article_no}"
|
||||
|
||||
@@ -0,0 +1,96 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import math
|
||||
import re
|
||||
from collections import Counter
|
||||
from dataclasses import dataclass
|
||||
from typing import Iterable
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class LegalSemanticMatch:
|
||||
matched: bool
|
||||
score: float
|
||||
shared_terms: list[str]
|
||||
shared_concepts: list[str]
|
||||
|
||||
|
||||
class LegalSemanticMatcher:
|
||||
"""Small deterministic vector matcher for legal-impact scanning.
|
||||
|
||||
This is intentionally local and dependency-free. It gives Phase15-G a real
|
||||
vector-style semantic scan path without requiring a remote embedding service.
|
||||
"""
|
||||
|
||||
LEGAL_CONCEPT_ALIASES: dict[str, tuple[str, ...]] = {
|
||||
"breach_liability": ("违约", "违约金", "违约责任", "未按约", "未履行", "迟延履行", "逾期交付", "逾期付款", "赔偿", "损失赔偿"),
|
||||
"payment": ("价款", "付款", "支付", "结算", "费用", "款项", "收款", "发票"),
|
||||
"termination": ("解除", "终止", "撤销", "合同解除", "提前终止", "退还"),
|
||||
"confidentiality": ("保密", "商业秘密", "信息披露", "泄露", "保密义务"),
|
||||
"intellectual_property": ("知识产权", "著作权", "专利", "商标", "许可", "侵权"),
|
||||
"labor": ("劳动", "用人单位", "劳动者", "工资", "加班", "社保", "劳动合同"),
|
||||
"data_privacy": ("个人信息", "数据", "隐私", "处理", "委托处理", "信息安全"),
|
||||
"dispute_resolution": ("争议", "仲裁", "诉讼", "管辖", "法院", "调解"),
|
||||
"delivery_acceptance": ("交付", "验收", "质量", "瑕疵", "整改", "交付物"),
|
||||
}
|
||||
|
||||
def __init__(self, *, threshold: float = 0.16):
|
||||
self.threshold = max(0.0, min(1.0, threshold))
|
||||
|
||||
def match(self, source_text: str, target_text: str) -> LegalSemanticMatch:
|
||||
source_vector = self.vectorize(source_text)
|
||||
target_vector = self.vectorize(target_text)
|
||||
if not source_vector or not target_vector:
|
||||
return LegalSemanticMatch(False, 0.0, [], [])
|
||||
|
||||
score = self._cosine(source_vector, target_vector)
|
||||
shared = [term for term in source_vector if term in target_vector]
|
||||
shared_terms = [term for term in shared if not term.startswith("concept:")][:12]
|
||||
shared_concepts = [term.replace("concept:", "", 1) for term in shared if term.startswith("concept:")]
|
||||
return LegalSemanticMatch(score >= self.threshold, round(score, 4), shared_terms, shared_concepts)
|
||||
|
||||
def vectorize(self, text: str) -> Counter[str]:
|
||||
normalized = self._normalize(text)
|
||||
vector: Counter[str] = Counter()
|
||||
if not normalized:
|
||||
return vector
|
||||
|
||||
for token in self._word_tokens(normalized):
|
||||
vector[token] += 2.0
|
||||
for token in self._han_ngrams(normalized, n=2):
|
||||
vector[token] += 0.35
|
||||
for token in self._han_ngrams(normalized, n=3):
|
||||
vector[token] += 0.5
|
||||
for concept, aliases in self.LEGAL_CONCEPT_ALIASES.items():
|
||||
matched_aliases = [alias for alias in aliases if alias in normalized]
|
||||
if matched_aliases:
|
||||
vector[f"concept:{concept}"] += 3.0
|
||||
for alias in matched_aliases:
|
||||
vector[alias] += 1.5
|
||||
return vector
|
||||
|
||||
@staticmethod
|
||||
def _normalize(text: str) -> str:
|
||||
return re.sub(r"\s+", " ", str(text or "").lower()).strip()
|
||||
|
||||
@staticmethod
|
||||
def _word_tokens(text: str) -> Iterable[str]:
|
||||
for token in re.findall(r"[a-z0-9_]{2,}|[\u4e00-\u9fff]{2,}", text):
|
||||
if len(token) <= 64:
|
||||
yield token
|
||||
|
||||
@staticmethod
|
||||
def _han_ngrams(text: str, *, n: int) -> Iterable[str]:
|
||||
for chunk in re.findall(r"[\u4e00-\u9fff]{%d,}" % n, text):
|
||||
for index in range(0, len(chunk) - n + 1):
|
||||
yield chunk[index : index + n]
|
||||
|
||||
@staticmethod
|
||||
def _cosine(left: Counter[str], right: Counter[str]) -> float:
|
||||
shared = set(left).intersection(right)
|
||||
numerator = sum(left[key] * right[key] for key in shared)
|
||||
left_norm = math.sqrt(sum(value * value for value in left.values()))
|
||||
right_norm = math.sqrt(sum(value * value for value in right.values()))
|
||||
if left_norm == 0 or right_norm == 0:
|
||||
return 0.0
|
||||
return numerator / (left_norm * right_norm)
|
||||
@@ -14,6 +14,7 @@ from app.repositories.prompt_asset_evaluation_run_repository import PromptAssetE
|
||||
from app.repositories.prompt_asset_repository import PromptAssetRepository
|
||||
from app.repositories.prompt_asset_run_link_repository import PromptAssetRunLinkRepository
|
||||
from app.repositories.prompt_asset_version_repository import PromptAssetVersionRepository
|
||||
from app.services.prompt_golden_sample_evaluator import PromptGoldenSampleEvaluator
|
||||
|
||||
|
||||
class PromptAssetService:
|
||||
@@ -42,6 +43,7 @@ class PromptAssetService:
|
||||
self.version_repository = PromptAssetVersionRepository(db)
|
||||
self.run_link_repository = PromptAssetRunLinkRepository(db)
|
||||
self.evaluation_run_repository = PromptAssetEvaluationRunRepository(db)
|
||||
self.golden_sample_evaluator = PromptGoldenSampleEvaluator()
|
||||
|
||||
def list_prompt_assets(
|
||||
self,
|
||||
@@ -214,7 +216,7 @@ class PromptAssetService:
|
||||
created_at=now,
|
||||
)
|
||||
)
|
||||
report = self._build_prompt_evaluation_report(asset, version)
|
||||
report = self._build_prompt_evaluation_report(current_user, asset, version)
|
||||
|
||||
version.evaluation_report_json = report
|
||||
version.evaluation_score = float(report["quality_score"])
|
||||
@@ -222,6 +224,7 @@ class PromptAssetService:
|
||||
version = self.version_repository.update(version)
|
||||
|
||||
evaluation_run.status = version.evaluation_status
|
||||
evaluation_run.suite_key = str(report.get("suite_key") or evaluation_run.suite_key)
|
||||
evaluation_run.sample_count = int(report.get("sample_count") or 0)
|
||||
evaluation_run.passed_samples = int(report.get("passed_samples") or 0)
|
||||
evaluation_run.failed_samples = int(report.get("failed_samples") or 0)
|
||||
@@ -474,38 +477,14 @@ class PromptAssetService:
|
||||
asset.latest_evaluation_score = version.evaluation_score
|
||||
asset.latest_evaluation_report_json = version.evaluation_report_json
|
||||
|
||||
def _build_prompt_evaluation_report(self, asset: PromptAsset, version: PromptAssetVersion) -> dict[str, Any]:
|
||||
suite = self.PROMPT_GOLDEN_SUITES.get(asset.prompt_type, {"suite_key": "custom_prompt_suite", "sample_count": 1})
|
||||
sample_count = int(suite["sample_count"])
|
||||
def _build_prompt_evaluation_report(
|
||||
self,
|
||||
current_user: Dict[str, Any],
|
||||
asset: PromptAsset,
|
||||
version: PromptAssetVersion,
|
||||
) -> dict[str, Any]:
|
||||
checks = self._build_prompt_quality_checks(asset, version)
|
||||
blocking_failures = [item for item in checks if item["required"] and item["status"] != "pass"]
|
||||
warning_failures = [item for item in checks if not item["required"] and item["status"] != "pass"]
|
||||
passed = not blocking_failures
|
||||
quality_score = max(0.0, 1.0 - (0.25 * len(blocking_failures)) - (0.05 * len(warning_failures)))
|
||||
if not passed:
|
||||
quality_score = min(quality_score, 0.79)
|
||||
passed_samples = sample_count if passed else max(0, sample_count - len(blocking_failures))
|
||||
pass_rate = 1.0 if sample_count == 0 else round(passed_samples / sample_count, 4)
|
||||
return {
|
||||
"schema_version": "phase15-prompt-evaluation-report.v1",
|
||||
"evaluator": "prompt_asset_api_static_gate",
|
||||
"prompt_asset_id": asset.id,
|
||||
"prompt_asset_version_id": version.id,
|
||||
"prompt_key": asset.prompt_key,
|
||||
"prompt_type": asset.prompt_type,
|
||||
"version_label": version.version_label,
|
||||
"suite_key": suite["suite_key"],
|
||||
"sample_count": sample_count,
|
||||
"passed_samples": passed_samples,
|
||||
"failed_samples": sample_count - passed_samples,
|
||||
"pass_rate": pass_rate,
|
||||
"quality_score": round(quality_score, 4),
|
||||
"passed": passed,
|
||||
"failed_items": [item for item in checks if item["status"] != "pass"],
|
||||
"checks": checks,
|
||||
"evaluated_at": self._now().isoformat(),
|
||||
"source": "prompt_assets.evaluate_current_version",
|
||||
}
|
||||
return self.golden_sample_evaluator.evaluate_prompt_asset(current_user, asset, version, static_checks=checks)
|
||||
|
||||
def _suite_key_for_prompt_type(self, prompt_type: str) -> str:
|
||||
return str(self.PROMPT_GOLDEN_SUITES.get(prompt_type, {"suite_key": "custom_prompt_suite"})["suite_key"])
|
||||
|
||||
@@ -0,0 +1,539 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import csv
|
||||
import json
|
||||
import re
|
||||
from datetime import datetime, timezone
|
||||
from pathlib import Path
|
||||
from typing import Any, Iterable, Mapping, Optional
|
||||
|
||||
from app.clients.shared_ai_platform import RemoteSharedAIPlatformClient
|
||||
from app.core.config import settings
|
||||
from app.models.prompt_asset import PromptAsset
|
||||
from app.models.prompt_asset_version import PromptAssetVersion
|
||||
|
||||
|
||||
class PromptGoldenSampleEvaluator:
|
||||
"""Run Prompt assets against Phase13/Phase15 golden samples."""
|
||||
|
||||
SUITE_CONFIG: dict[str, dict[str, Any]] = {
|
||||
"contract_generation": {
|
||||
"suite_key": "phase13_generation",
|
||||
"sample_family": "phase13_generation",
|
||||
"min_samples": 8,
|
||||
},
|
||||
"contract_review": {
|
||||
"suite_key": "phase13_review",
|
||||
"sample_family": "phase13_review",
|
||||
"min_samples": 8,
|
||||
},
|
||||
"legal_chat": {
|
||||
"suite_key": "phase13_kb_qa",
|
||||
"sample_family": "phase13_kb_qa",
|
||||
"min_samples": 10,
|
||||
},
|
||||
"clause_recommendation": {
|
||||
"suite_key": "phase13_clause_generation",
|
||||
"sample_family": "phase13_clause_generation",
|
||||
"min_samples": 6,
|
||||
},
|
||||
"wording_rewrite": {
|
||||
"suite_key": "phase15_wording_rewrite",
|
||||
"sample_family": "phase15_wording_rewrite",
|
||||
"min_samples": 3,
|
||||
},
|
||||
"regulation_impact": {
|
||||
"suite_key": "future_regulation_impact",
|
||||
"sample_family": "future_regulation_impact",
|
||||
"min_samples": 0,
|
||||
},
|
||||
}
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
repo_root: Optional[Path | str] = None,
|
||||
llm_client: Optional[Any] = None,
|
||||
evaluation_mode: Optional[str] = None,
|
||||
include_raw_model_response: Optional[bool] = None,
|
||||
):
|
||||
self.repo_root = Path(repo_root) if repo_root is not None else Path(__file__).resolve().parents[3]
|
||||
self.llm_client = llm_client
|
||||
self.evaluation_mode = self._normalize_mode(evaluation_mode or settings.prompt_golden_eval_mode)
|
||||
self.include_raw_model_response = (
|
||||
settings.prompt_golden_eval_include_raw_model_response
|
||||
if include_raw_model_response is None
|
||||
else include_raw_model_response
|
||||
)
|
||||
|
||||
def evaluate_prompt_asset(
|
||||
self,
|
||||
current_user: dict[str, Any],
|
||||
asset: PromptAsset,
|
||||
version: PromptAssetVersion,
|
||||
*,
|
||||
static_checks: list[dict[str, Any]],
|
||||
) -> dict[str, Any]:
|
||||
suite = self.SUITE_CONFIG.get(asset.prompt_type, self.SUITE_CONFIG["regulation_impact"])
|
||||
samples = self.load_samples(asset.prompt_type)
|
||||
llm_client = self._resolve_llm_client()
|
||||
use_llm = llm_client is not None and self.evaluation_mode in {"auto", "llm"}
|
||||
strict_llm = self.evaluation_mode == "llm"
|
||||
sample_results: list[dict[str, Any]] = []
|
||||
|
||||
for sample in samples:
|
||||
if use_llm:
|
||||
sample_results.append(
|
||||
self._evaluate_sample_with_llm(
|
||||
current_user,
|
||||
asset,
|
||||
version,
|
||||
sample,
|
||||
llm_client=llm_client,
|
||||
fallback_to_local=not strict_llm,
|
||||
)
|
||||
)
|
||||
elif strict_llm:
|
||||
sample_results.append(
|
||||
self._failed_sample_result(
|
||||
sample,
|
||||
issue="llm_not_configured",
|
||||
detail="Shared AI platform is not configured; real per-sample LLM evaluation cannot run.",
|
||||
)
|
||||
)
|
||||
else:
|
||||
sample_results.append(self._evaluate_sample_locally(asset, version, sample, static_checks))
|
||||
|
||||
required_static_failures = [item for item in static_checks if item.get("required") and item.get("status") != "pass"]
|
||||
warning_static_failures = [item for item in static_checks if not item.get("required") and item.get("status") != "pass"]
|
||||
passed_samples = sum(1 for item in sample_results if item.get("passed") is True)
|
||||
sample_count = len(sample_results)
|
||||
min_samples = int(suite.get("min_samples") or 0)
|
||||
suite_issues = []
|
||||
if sample_count < min_samples:
|
||||
suite_issues.append(
|
||||
{
|
||||
"key": "sample_count_below_minimum",
|
||||
"label": "Golden sample count is below the required minimum",
|
||||
"required": True,
|
||||
"status": "fail",
|
||||
"detail": f"sample_count={sample_count}; min_samples={min_samples}",
|
||||
}
|
||||
)
|
||||
|
||||
sample_failures = [item for item in sample_results if item.get("passed") is not True]
|
||||
passed = not required_static_failures and not suite_issues and not sample_failures
|
||||
pass_rate = 1.0 if sample_count == 0 and min_samples == 0 else round(passed_samples / sample_count, 4) if sample_count else 0.0
|
||||
sample_score = (
|
||||
sum(float(item.get("score") or 0) for item in sample_results) / sample_count
|
||||
if sample_count
|
||||
else (1.0 if min_samples == 0 else 0.0)
|
||||
)
|
||||
quality_score = sample_score - (0.1 * len(required_static_failures)) - (0.03 * len(warning_static_failures)) - (0.15 * len(suite_issues))
|
||||
if not passed:
|
||||
quality_score = min(quality_score, 0.79)
|
||||
quality_score = max(0.0, min(1.0, quality_score))
|
||||
failed_items = [
|
||||
*suite_issues,
|
||||
*[item for item in static_checks if item.get("status") != "pass"],
|
||||
*[
|
||||
{
|
||||
"key": f"sample_failed:{item.get('sample_id')}",
|
||||
"label": f"Golden sample failed: {item.get('sample_id')}",
|
||||
"required": True,
|
||||
"status": "fail",
|
||||
"detail": "; ".join(str(issue) for issue in item.get("issues") or [])[:500],
|
||||
}
|
||||
for item in sample_failures
|
||||
],
|
||||
]
|
||||
return {
|
||||
"schema_version": "phase15-prompt-evaluation-report.v2",
|
||||
"phase": "Phase15-F",
|
||||
"evaluator": "prompt_asset_api_llm_golden_sample_gate",
|
||||
"evaluation_mode": "llm" if use_llm else "local_fallback" if not strict_llm else "llm_required_not_configured",
|
||||
"prompt_asset_id": asset.id,
|
||||
"prompt_asset_version_id": version.id,
|
||||
"prompt_key": asset.prompt_key,
|
||||
"prompt_type": asset.prompt_type,
|
||||
"version_label": version.version_label,
|
||||
"suite_key": suite["suite_key"],
|
||||
"sample_family": suite["sample_family"],
|
||||
"min_samples": min_samples,
|
||||
"sample_count": sample_count,
|
||||
"passed_samples": passed_samples,
|
||||
"failed_samples": sample_count - passed_samples,
|
||||
"pass_rate": pass_rate,
|
||||
"quality_score": round(quality_score, 4),
|
||||
"passed": passed,
|
||||
"model_invocation": {
|
||||
"configured": bool(use_llm),
|
||||
"strict_llm": strict_llm,
|
||||
"provider": "shared-ai-platform" if use_llm else None,
|
||||
"mode": "llm" if use_llm else self.evaluation_mode,
|
||||
"sample_invocation_count": sum(1 for item in sample_results if item.get("evaluator_mode") == "llm"),
|
||||
},
|
||||
"sample_results": sample_results,
|
||||
"failed_items": failed_items,
|
||||
"checks": static_checks,
|
||||
"evaluated_at": datetime.now(timezone.utc).isoformat(),
|
||||
"source": "prompt_assets.evaluate_current_version.golden_samples",
|
||||
}
|
||||
|
||||
def load_samples(self, prompt_type: str) -> list[dict[str, Any]]:
|
||||
if prompt_type == "contract_generation":
|
||||
return [
|
||||
self._generation_sample(row, sample_family="phase13_generation")
|
||||
for row in self._read_phase13_csv("generation-golden-sample-inventory.csv")
|
||||
if str(row.get("sample_id") or "").startswith("GEN-SAMPLE-TPL-")
|
||||
]
|
||||
if prompt_type == "clause_recommendation":
|
||||
return [
|
||||
self._generation_sample(row, sample_family="phase13_clause_generation")
|
||||
for row in self._read_phase13_csv("generation-golden-sample-inventory.csv")
|
||||
if str(row.get("sample_id") or "").startswith("GEN-SAMPLE-CLAUSE-")
|
||||
]
|
||||
if prompt_type == "contract_review":
|
||||
return [self._review_sample(row) for row in self._read_phase13_csv("review-golden-sample-inventory.csv")]
|
||||
if prompt_type == "legal_chat":
|
||||
return [self._kb_qa_sample(row) for row in self._read_phase13_csv("kb-qa-golden-sample-inventory.csv")]
|
||||
if prompt_type == "wording_rewrite":
|
||||
return self._wording_rewrite_samples()
|
||||
return []
|
||||
|
||||
def _evaluate_sample_with_llm(
|
||||
self,
|
||||
current_user: dict[str, Any],
|
||||
asset: PromptAsset,
|
||||
version: PromptAssetVersion,
|
||||
sample: dict[str, Any],
|
||||
*,
|
||||
llm_client: Any,
|
||||
fallback_to_local: bool,
|
||||
) -> dict[str, Any]:
|
||||
message = self._build_llm_evaluation_message(asset, version, sample)
|
||||
try:
|
||||
raw_text = self._collect_llm_text(
|
||||
llm_client.stream_answer(
|
||||
current_user,
|
||||
None,
|
||||
message,
|
||||
review_context={
|
||||
"task": "phase15_prompt_golden_sample_evaluation",
|
||||
"prompt_type": asset.prompt_type,
|
||||
"sample_id": sample.get("sample_id"),
|
||||
},
|
||||
template_key=None,
|
||||
)
|
||||
)
|
||||
verdict = self._parse_llm_verdict(raw_text)
|
||||
result = {
|
||||
"sample_id": sample["sample_id"],
|
||||
"prompt_type": sample["prompt_type"],
|
||||
"sample_family": sample["sample_family"],
|
||||
"status": "pass" if verdict["passed"] else "fail",
|
||||
"passed": verdict["passed"],
|
||||
"score": verdict["score"],
|
||||
"evaluator_mode": "llm",
|
||||
"input_summary": sample.get("input_summary"),
|
||||
"expected_summary": sample.get("expected_summary"),
|
||||
"observed_output_summary": verdict.get("observed_output_summary"),
|
||||
"issues": verdict.get("issues") or [],
|
||||
"matched_expectations": verdict.get("matched_expectations") or [],
|
||||
"model_trace": verdict.get("model_trace") or {},
|
||||
}
|
||||
if self.include_raw_model_response:
|
||||
result["raw_model_response"] = raw_text[:4000]
|
||||
return result
|
||||
except Exception as exc:
|
||||
if fallback_to_local:
|
||||
result = self._evaluate_sample_locally(asset, version, sample, [])
|
||||
result["evaluator_mode"] = "local_fallback_after_llm_error"
|
||||
result["issues"] = [*result.get("issues", []), f"llm_error:{str(exc)[:160]}"]
|
||||
return result
|
||||
return self._failed_sample_result(sample, issue="llm_call_failed", detail=str(exc)[:300])
|
||||
|
||||
def _evaluate_sample_locally(
|
||||
self,
|
||||
asset: PromptAsset,
|
||||
version: PromptAssetVersion,
|
||||
sample: dict[str, Any],
|
||||
static_checks: list[dict[str, Any]],
|
||||
) -> dict[str, Any]:
|
||||
content = f"{version.system_prompt or ''}\n{version.prompt_content or ''}".lower()
|
||||
required_failures = [item for item in static_checks if item.get("required") and item.get("status") != "pass"]
|
||||
issues: list[str] = []
|
||||
checks: list[dict[str, Any]] = []
|
||||
if required_failures:
|
||||
issues.append("required_static_prompt_checks_failed")
|
||||
if not sample.get("expected"):
|
||||
issues.append("sample_expected_payload_missing")
|
||||
if sample.get("expected_citation_required") and not self._contains_any(content, ["law", "legal", "basis", "citation", "source", "article"]):
|
||||
checks.append({"key": "citation_instruction", "status": "warn", "detail": "Prompt does not clearly require legal source citations."})
|
||||
if asset.prompt_type == "wording_rewrite" and not self._contains_any(content, ["rewrite", "wording", "optimize", "revise"]):
|
||||
issues.append("wording_rewrite_instruction_missing")
|
||||
if asset.prompt_type == "contract_review" and not self._contains_any(content, ["review", "risk", "issue", "finding"]):
|
||||
issues.append("contract_review_instruction_missing")
|
||||
if asset.prompt_type == "legal_chat" and not self._contains_any(content, ["chat", "answer", "question", "reply"]):
|
||||
checks.append({"key": "chat_answer_instruction", "status": "warn", "detail": "Prompt does not clearly ask the model to answer the user's question."})
|
||||
|
||||
passed = not issues
|
||||
score = 0.94 if passed else 0.55
|
||||
if checks:
|
||||
score = min(score, 0.88)
|
||||
return {
|
||||
"sample_id": sample["sample_id"],
|
||||
"prompt_type": sample["prompt_type"],
|
||||
"sample_family": sample["sample_family"],
|
||||
"status": "pass" if passed else "fail",
|
||||
"passed": passed,
|
||||
"score": round(score, 4),
|
||||
"evaluator_mode": "local_fallback",
|
||||
"input_summary": sample.get("input_summary"),
|
||||
"expected_summary": sample.get("expected_summary"),
|
||||
"observed_output_summary": "Local fallback evaluation: checked prompt/sample/expected/source-boundary alignment.",
|
||||
"issues": issues,
|
||||
"matched_expectations": [item for item in ["sample_loaded", "expected_loaded", "prompt_static_gate"] if item],
|
||||
"checks": checks,
|
||||
}
|
||||
|
||||
def _failed_sample_result(self, sample: dict[str, Any], *, issue: str, detail: str) -> dict[str, Any]:
|
||||
return {
|
||||
"sample_id": sample["sample_id"],
|
||||
"prompt_type": sample["prompt_type"],
|
||||
"sample_family": sample["sample_family"],
|
||||
"status": "fail",
|
||||
"passed": False,
|
||||
"score": 0.0,
|
||||
"evaluator_mode": "llm",
|
||||
"input_summary": sample.get("input_summary"),
|
||||
"expected_summary": sample.get("expected_summary"),
|
||||
"observed_output_summary": detail,
|
||||
"issues": [issue],
|
||||
"matched_expectations": [],
|
||||
}
|
||||
|
||||
def _build_llm_evaluation_message(self, asset: PromptAsset, version: PromptAssetVersion, sample: dict[str, Any]) -> str:
|
||||
payload = {
|
||||
"prompt_asset": {
|
||||
"prompt_key": asset.prompt_key,
|
||||
"prompt_type": asset.prompt_type,
|
||||
"version_label": version.version_label,
|
||||
"system_prompt": version.system_prompt,
|
||||
"prompt_content": version.prompt_content,
|
||||
"variables": version.variables_json or [],
|
||||
"output_schema": version.output_schema_json,
|
||||
"model_config": version.model_config_json,
|
||||
},
|
||||
"golden_sample": sample,
|
||||
}
|
||||
return (
|
||||
"You are the Phase15-F prompt golden-sample evaluator for the AI contract platform. "
|
||||
"Treat prompt_asset as the prompt under test, reason through the golden_sample input, "
|
||||
"and compare the expected boundaries. Return JSON only, no Markdown.\n"
|
||||
"JSON schema: {\"passed\":true|false,\"score\":0..1,\"observed_output_summary\":\"...\","
|
||||
"\"issues\":[\"...\"],\"matched_expectations\":[\"...\"],\"model_trace\":{\"reason\":\"...\"}}.\n"
|
||||
"Check sample expectations, legal-source grounding, and avoid over-confident legal conclusions.\n"
|
||||
f"Evaluation input: {json.dumps(payload, ensure_ascii=False, default=str)}"
|
||||
)
|
||||
|
||||
def _parse_llm_verdict(self, raw_text: str) -> dict[str, Any]:
|
||||
payload = self._extract_json_object(raw_text)
|
||||
passed = bool(payload.get("passed"))
|
||||
score = payload.get("score")
|
||||
try:
|
||||
numeric_score = float(score)
|
||||
except (TypeError, ValueError):
|
||||
numeric_score = 1.0 if passed else 0.0
|
||||
issues = payload.get("issues") if isinstance(payload.get("issues"), list) else []
|
||||
matched = payload.get("matched_expectations") if isinstance(payload.get("matched_expectations"), list) else []
|
||||
return {
|
||||
"passed": passed,
|
||||
"score": max(0.0, min(1.0, numeric_score)),
|
||||
"observed_output_summary": str(payload.get("observed_output_summary") or "")[:1000],
|
||||
"issues": [str(item)[:300] for item in issues],
|
||||
"matched_expectations": [str(item)[:200] for item in matched],
|
||||
"model_trace": payload.get("model_trace") if isinstance(payload.get("model_trace"), dict) else {},
|
||||
}
|
||||
|
||||
def _extract_json_object(self, raw_text: str) -> dict[str, Any]:
|
||||
text = str(raw_text or "").strip()
|
||||
try:
|
||||
payload = json.loads(text)
|
||||
return payload if isinstance(payload, dict) else {}
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
match = re.search(r"\{.*\}", text, flags=re.S)
|
||||
if not match:
|
||||
return {"passed": False, "score": 0, "issues": ["model_response_not_json"], "observed_output_summary": text[:500]}
|
||||
try:
|
||||
payload = json.loads(match.group(0))
|
||||
except json.JSONDecodeError:
|
||||
return {"passed": False, "score": 0, "issues": ["model_response_json_parse_failed"], "observed_output_summary": text[:500]}
|
||||
return payload if isinstance(payload, dict) else {}
|
||||
|
||||
def _collect_llm_text(self, chunks: Iterable[str]) -> str:
|
||||
output: list[str] = []
|
||||
for chunk in chunks:
|
||||
text = str(chunk or "")
|
||||
parsed = self._extract_sse_text(text)
|
||||
output.append(parsed if parsed is not None else text)
|
||||
return "".join(output).strip()
|
||||
|
||||
def _extract_sse_text(self, value: str) -> Optional[str]:
|
||||
data_parts: list[str] = []
|
||||
for line in value.splitlines():
|
||||
if not line.startswith("data:"):
|
||||
continue
|
||||
raw = line[5:].strip()
|
||||
if not raw:
|
||||
continue
|
||||
try:
|
||||
payload = json.loads(raw)
|
||||
except json.JSONDecodeError:
|
||||
data_parts.append(raw)
|
||||
continue
|
||||
if isinstance(payload, Mapping):
|
||||
if payload.get("type") in {"text", "delta"}:
|
||||
data_parts.append(str(payload.get("content") or ""))
|
||||
elif payload.get("content"):
|
||||
data_parts.append(str(payload.get("content")))
|
||||
if not data_parts:
|
||||
return None
|
||||
return "".join(data_parts)
|
||||
|
||||
def _resolve_llm_client(self) -> Optional[Any]:
|
||||
if self.evaluation_mode == "local":
|
||||
return None
|
||||
if self.llm_client is not None:
|
||||
return self.llm_client
|
||||
if not settings.shared_ai_platform_base_url.strip() or not settings.shared_ai_platform_api_key.strip():
|
||||
return None
|
||||
return RemoteSharedAIPlatformClient(
|
||||
base_url=settings.shared_ai_platform_base_url,
|
||||
app_id=settings.shared_ai_platform_app_id,
|
||||
api_key=settings.shared_ai_platform_api_key,
|
||||
timeout_seconds=settings.shared_ai_platform_timeout_seconds,
|
||||
)
|
||||
|
||||
def _read_phase13_csv(self, file_name: str) -> list[dict[str, str]]:
|
||||
path = self.repo_root / "project-docs" / "phase13-legal-knowledge-template-library" / "inventories" / file_name
|
||||
return self._read_csv(path)
|
||||
|
||||
def _read_csv(self, path: Path) -> list[dict[str, str]]:
|
||||
if not path.exists():
|
||||
return []
|
||||
with path.open("r", encoding="utf-8-sig", newline="") as handle:
|
||||
return [dict(row) for row in csv.DictReader(handle)]
|
||||
|
||||
def _generation_sample(self, row: Mapping[str, str], *, sample_family: str) -> dict[str, Any]:
|
||||
base = self.repo_root / "project-docs" / "phase13-legal-knowledge-template-library"
|
||||
input_payload = self._read_json_file(base / str(row.get("variable_payload_path") or ""))
|
||||
expected_text = self._read_text_file(base / str(row.get("expected_document_path") or ""))
|
||||
return {
|
||||
"sample_id": row.get("sample_id"),
|
||||
"prompt_type": "clause_recommendation" if sample_family == "phase13_clause_generation" else "contract_generation",
|
||||
"sample_family": sample_family,
|
||||
"input": input_payload,
|
||||
"expected": {
|
||||
"expected_document_excerpt": expected_text[:2000],
|
||||
"expected_review_rule_ids": self._split_semicolon(row.get("expected_review_rule_ids")),
|
||||
"expected_source_citations": self._split_semicolon(row.get("expected_source_citations")),
|
||||
"contract_type": row.get("contract_type"),
|
||||
"asset_under_test": row.get("template_id"),
|
||||
},
|
||||
"expected_citation_required": bool(row.get("expected_source_citations")),
|
||||
"input_summary": f"{row.get('sample_id')} variables={len(input_payload) if isinstance(input_payload, Mapping) else 'n/a'}",
|
||||
"expected_summary": f"rules={row.get('expected_review_rule_ids')}; sources={row.get('expected_source_citations')}",
|
||||
}
|
||||
|
||||
def _review_sample(self, row: Mapping[str, str]) -> dict[str, Any]:
|
||||
base = self.repo_root / "project-docs" / "phase13-legal-knowledge-template-library"
|
||||
input_text = self._read_text_file(base / str(row.get("input_sample_path") or ""))
|
||||
expected_text = self._read_text_file(base / str(row.get("expected_findings_path") or ""))
|
||||
return {
|
||||
"sample_id": row.get("sample_id"),
|
||||
"prompt_type": "contract_review",
|
||||
"sample_family": "phase13_review",
|
||||
"input": {"contract_text": input_text, "contract_type": row.get("contract_type")},
|
||||
"expected": {
|
||||
"expected_findings_excerpt": expected_text[:2000],
|
||||
"risk_categories": self._split_semicolon(row.get("risk_categories")),
|
||||
"expected_severity": row.get("expected_severity"),
|
||||
"expected_source_citations": self._split_semicolon(row.get("expected_source_citations")),
|
||||
},
|
||||
"expected_citation_required": bool(row.get("expected_source_citations")),
|
||||
"input_summary": f"{row.get('sample_id')} contract_chars={len(input_text)}",
|
||||
"expected_summary": f"severity={row.get('expected_severity')}; categories={row.get('risk_categories')}",
|
||||
}
|
||||
|
||||
def _kb_qa_sample(self, row: Mapping[str, str]) -> dict[str, Any]:
|
||||
return {
|
||||
"sample_id": row.get("sample_id"),
|
||||
"prompt_type": "legal_chat",
|
||||
"sample_family": "phase13_kb_qa",
|
||||
"input": {"question": row.get("question")},
|
||||
"expected": {
|
||||
"expected_answer_boundary": row.get("expected_answer_boundary"),
|
||||
"expected_source_ids": self._split_semicolon(row.get("expected_source_ids")),
|
||||
"expected_citation_required": str(row.get("expected_citation_required")).lower() == "true",
|
||||
"expected_refusal_required": str(row.get("expected_refusal_required")).lower() == "true",
|
||||
},
|
||||
"expected_citation_required": str(row.get("expected_citation_required")).lower() == "true",
|
||||
"input_summary": str(row.get("question") or "")[:200],
|
||||
"expected_summary": str(row.get("expected_answer_boundary") or "")[:300],
|
||||
}
|
||||
|
||||
def _wording_rewrite_samples(self) -> list[dict[str, Any]]:
|
||||
path = self.repo_root / "project-docs" / "phase15-prompt-golden-samples" / "wording-rewrite-golden-samples.json"
|
||||
payload = self._read_json_file(path)
|
||||
samples = payload.get("samples") if isinstance(payload, Mapping) else []
|
||||
results: list[dict[str, Any]] = []
|
||||
for item in samples if isinstance(samples, list) else []:
|
||||
if not isinstance(item, Mapping):
|
||||
continue
|
||||
input_payload = item.get("input") if isinstance(item.get("input"), Mapping) else {}
|
||||
expected_payload = item.get("expected") if isinstance(item.get("expected"), Mapping) else {}
|
||||
results.append(
|
||||
{
|
||||
"sample_id": item.get("sample_id"),
|
||||
"prompt_type": "wording_rewrite",
|
||||
"sample_family": "phase15_wording_rewrite",
|
||||
"rewrite_type": item.get("rewrite_type"),
|
||||
"document_type_code": item.get("document_type_code"),
|
||||
"input": input_payload,
|
||||
"expected": expected_payload,
|
||||
"expected_citation_required": False,
|
||||
"input_summary": str(input_payload.get("clause_text") or "")[:200],
|
||||
"expected_summary": json.dumps(expected_payload, ensure_ascii=False)[:300],
|
||||
}
|
||||
)
|
||||
return results
|
||||
|
||||
def _read_json_file(self, path: Path) -> Any:
|
||||
if not path.exists() or not path.is_file():
|
||||
return {}
|
||||
try:
|
||||
return json.loads(path.read_text(encoding="utf-8-sig"))
|
||||
except json.JSONDecodeError:
|
||||
return {}
|
||||
|
||||
def _read_text_file(self, path: Path) -> str:
|
||||
if not path.exists() or not path.is_file():
|
||||
return ""
|
||||
return path.read_text(encoding="utf-8-sig")
|
||||
|
||||
@staticmethod
|
||||
def _split_semicolon(value: Optional[str]) -> list[str]:
|
||||
return [item.strip() for item in str(value or "").split(";") if item.strip()]
|
||||
|
||||
@staticmethod
|
||||
def _normalize_mode(value: Optional[str]) -> str:
|
||||
normalized = str(value or "auto").strip().lower().replace("-", "_")
|
||||
if normalized not in {"auto", "llm", "local"}:
|
||||
return "auto"
|
||||
return normalized
|
||||
|
||||
@staticmethod
|
||||
def _contains_any(value: str, candidates: list[str]) -> bool:
|
||||
return any(candidate.lower() in value for candidate in candidates)
|
||||
@@ -0,0 +1,116 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import time
|
||||
from datetime import datetime, timezone
|
||||
from typing import Any, Callable, Optional
|
||||
|
||||
from fastapi import HTTPException
|
||||
from sqlalchemy.orm import Session
|
||||
|
||||
from app.core.config import settings
|
||||
from app.db import SessionLocal
|
||||
from app.repositories.legal_change_impact_repository import LegalChangeImpactRepository
|
||||
from app.services.legal_change_impact_service import LegalChangeImpactService
|
||||
|
||||
|
||||
class LegalChangeImpactWorker:
|
||||
"""Consume pending legal-change events in a standalone worker process."""
|
||||
|
||||
def __init__(self, session_factory: Callable[[], Session] = SessionLocal):
|
||||
self.session_factory = session_factory
|
||||
|
||||
def run_once(self, *, limit: Optional[int] = None) -> dict[str, Any]:
|
||||
if not settings.legal_change_impact_worker_enabled:
|
||||
return {
|
||||
"worker": "legal_change_impact_worker",
|
||||
"status": "disabled",
|
||||
"completed_count": 0,
|
||||
"failed_count": 0,
|
||||
"completed": [],
|
||||
"failed": [],
|
||||
}
|
||||
|
||||
batch_size = max(1, min(limit or settings.legal_change_impact_worker_batch_size, 100))
|
||||
db = self.session_factory()
|
||||
try:
|
||||
repository = LegalChangeImpactRepository(db)
|
||||
events = repository.list_pending_events_for_worker(limit=batch_size)
|
||||
completed: list[dict[str, Any]] = []
|
||||
failed: list[dict[str, Any]] = []
|
||||
for event in events:
|
||||
current_user = {
|
||||
"id": settings.legal_change_impact_worker_user_id,
|
||||
"tenant_id": event.tenant_id,
|
||||
"role": "admin",
|
||||
}
|
||||
try:
|
||||
result = LegalChangeImpactService(db).run_event_scan(current_user, event.id)
|
||||
completed.append(
|
||||
{
|
||||
"event_id": event.id,
|
||||
"tenant_id": event.tenant_id,
|
||||
"impact_summary": result["event"].get("impact_summary", {}),
|
||||
}
|
||||
)
|
||||
except HTTPException as exc:
|
||||
failed.append({"event_id": event.id, "tenant_id": event.tenant_id, "status_code": exc.status_code, "message": exc.detail})
|
||||
except Exception as exc: # pragma: no cover - defensive worker safety net
|
||||
self._mark_event_failed(db, event.tenant_id, event.id, str(exc))
|
||||
failed.append({"event_id": event.id, "tenant_id": event.tenant_id, "status_code": 500, "message": str(exc)[:300]})
|
||||
|
||||
return {
|
||||
"worker": "legal_change_impact_worker",
|
||||
"status": "completed",
|
||||
"requested_limit": batch_size,
|
||||
"completed_count": len(completed),
|
||||
"failed_count": len(failed),
|
||||
"completed": completed,
|
||||
"failed": failed,
|
||||
"ran_at": datetime.now(timezone.utc).isoformat(),
|
||||
}
|
||||
finally:
|
||||
db.close()
|
||||
|
||||
def run_forever(self, *, limit: Optional[int] = None, interval_seconds: Optional[int] = None, max_loops: Optional[int] = None) -> None:
|
||||
interval = max(5, interval_seconds or settings.legal_change_impact_worker_interval_seconds)
|
||||
loops = 0
|
||||
while True:
|
||||
result = self.run_once(limit=limit)
|
||||
print(json.dumps(result, ensure_ascii=False), flush=True)
|
||||
loops += 1
|
||||
if max_loops is not None and loops >= max_loops:
|
||||
return
|
||||
time.sleep(interval)
|
||||
|
||||
@staticmethod
|
||||
def _mark_event_failed(db: Session, tenant_id: str, event_id: str, message: str) -> None:
|
||||
repository = LegalChangeImpactRepository(db)
|
||||
event = repository.get_event_by_id(tenant_id, event_id)
|
||||
if event is None:
|
||||
return
|
||||
event.scan_status = "failed"
|
||||
event.impact_summary_json = {"error": message[:500], "total": 0, "blocking_count": 0}
|
||||
event.updated_at = datetime.now(timezone.utc)
|
||||
repository.update_event(event)
|
||||
|
||||
|
||||
def main(argv: Optional[list[str]] = None) -> int:
|
||||
parser = argparse.ArgumentParser(description="Run the legal-change impact background worker.")
|
||||
parser.add_argument("--once", action="store_true", help="Run one batch and exit.")
|
||||
parser.add_argument("--limit", type=int, default=None, help="Maximum pending events to consume per batch.")
|
||||
parser.add_argument("--interval-seconds", type=int, default=None, help="Polling interval for daemon mode.")
|
||||
parser.add_argument("--max-loops", type=int, default=None, help="Optional loop cap for smoke tests.")
|
||||
args = parser.parse_args(argv)
|
||||
|
||||
worker = LegalChangeImpactWorker()
|
||||
if args.once:
|
||||
print(json.dumps(worker.run_once(limit=args.limit), ensure_ascii=False), flush=True)
|
||||
return 0
|
||||
worker.run_forever(limit=args.limit, interval_seconds=args.interval_seconds, max_loops=args.max_loops)
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -96,6 +96,36 @@ class ClauseRecommendationServiceTestCase(unittest.TestCase):
|
||||
self.assertEqual(result["document_type_code"], "service_general")
|
||||
self.assertEqual([item["clause_key"] for item in result["items"]], ["service_scope", "deliverables"])
|
||||
|
||||
def test_software_contract_returns_industry_practice_recommendations(self):
|
||||
result = self.service.recommend_missing_clauses(
|
||||
self.current_user,
|
||||
document_type_code="software_development",
|
||||
contract_text=(
|
||||
"软件开发合同\n"
|
||||
"一、服务范围：乙方为甲方开发客户管理系统。\n"
|
||||
"二、交付物：乙方提交源代码、部署包和操作文档。\n"
|
||||
"三、验收：甲方在收到交付物后五个工作日内完成验收。\n"
|
||||
"四、知识产权：定制成果权利归甲方所有。\n"
|
||||
"五、保密：双方对合作资料承担保密义务。\n"
|
||||
"六、甲方配合义务：甲方提供接口账号和测试环境。\n"
|
||||
"七、维护支持：乙方在验收后提供必要缺陷修复。"
|
||||
),
|
||||
max_items=3,
|
||||
)
|
||||
|
||||
self.assertEqual(result["document_type_code"], "software_development")
|
||||
self.assertGreaterEqual(result["summary"]["industry_practice_count"], 1)
|
||||
self.assertEqual(result["summary"]["missing_clause_keys"], [])
|
||||
first = result["items"][0]
|
||||
self.assertEqual(first["clause_key"], "scope_change_control")
|
||||
self.assertEqual(first["recommendation_kind"], "industry_practice")
|
||||
self.assertEqual(first["basis_type"], "industry_practice_only")
|
||||
self.assertIn("无直接法条", first["legal_basis"])
|
||||
self.assertIn("industry_practice_rules", first["source_types"])
|
||||
self.assertIn("no_direct_legal_article", first["source_types"])
|
||||
self.assertEqual(first["legal_article_refs"], [])
|
||||
self.assertIn(first["clause_key"], result["summary"]["industry_practice_keys"])
|
||||
|
||||
def test_professional_rewrite_expands_ambiguous_delivery_deadline(self):
|
||||
result = self.service.rewrite_clause_wording(
|
||||
self.current_user,
|
||||
|
||||
@@ -32,6 +32,9 @@ class StubClauseRecommendationService:
|
||||
"clause_key": "acceptance",
|
||||
"clause_title": "验收标准与流程",
|
||||
"clause_type": "acceptance",
|
||||
"recommendation_kind": "missing_clause",
|
||||
"basis_type": "legal_basis",
|
||||
"industry_context": None,
|
||||
"priority": "high",
|
||||
"confidence": "high",
|
||||
"insert_position": {
|
||||
@@ -42,6 +45,7 @@ class StubClauseRecommendationService:
|
||||
"suggested_text": "验收标准与流程：甲方应在收到货物后完成验收。",
|
||||
"reason": "买卖合同需要把交付和验收衔接起来。",
|
||||
"legal_basis": "《民法典》合同履行规则。",
|
||||
"practice_basis": None,
|
||||
"legal_article_refs": [],
|
||||
"source_rule": "LAW-CLAUSE-PROCUREMENT-ACCEPTANCE",
|
||||
"source_types": ["document_type_required_clause_list", "standard_clause_library"],
|
||||
@@ -51,10 +55,15 @@ class StubClauseRecommendationService:
|
||||
"total": 1,
|
||||
"summary": {
|
||||
"recommended_count": 1,
|
||||
"recommendation_keys": ["acceptance"],
|
||||
"detected_clause_keys": [],
|
||||
"missing_clause_keys": ["acceptance"],
|
||||
"required_missing_clause_keys": ["acceptance"],
|
||||
"industry_practice_keys": [],
|
||||
"detected_industry_practice_keys": [],
|
||||
"industry_practice_count": 0,
|
||||
"source_count": 0,
|
||||
"strategy": "rule_based_required_clause_gap",
|
||||
"strategy": "rule_based_required_clause_gap_and_industry_practice",
|
||||
},
|
||||
"sources": ["document_type_required_clause_list"],
|
||||
}
|
||||
@@ -128,6 +137,7 @@ class ClausesApiTestCase(unittest.TestCase):
|
||||
self.assertEqual(self.clause_service.calls[0][1]["contract_id"], "contract-001")
|
||||
self.assertEqual(self.audit_log_service.events[0]["action"], "recommend_missing_clauses")
|
||||
self.assertEqual(self.audit_log_service.events[0]["detail_json"]["missing_clause_keys"], ["acceptance"])
|
||||
self.assertEqual(self.audit_log_service.events[0]["detail_json"]["industry_practice_count"], 0)
|
||||
|
||||
def test_rewrite_clause_wording_returns_result_and_records_audit(self):
|
||||
response = self.client.post(
|
||||
|
||||
@@ -11,6 +11,7 @@ from app.models.contract_template import ContractTemplate
|
||||
from app.models.generated_contract import GeneratedContract
|
||||
from app.models.legal_article import LegalArticle
|
||||
from app.models.review_rule import ReviewRule
|
||||
from app.tasks.legal_change_impact_worker import LegalChangeImpactWorker
|
||||
from app.services.legal_change_impact_service import LegalChangeImpactService
|
||||
|
||||
|
||||
@@ -19,6 +20,7 @@ class LegalChangeImpactServiceTestCase(unittest.TestCase):
|
||||
self.engine = create_engine("sqlite:///:memory:", future=True)
|
||||
Base.metadata.create_all(bind=self.engine)
|
||||
SessionLocal = sessionmaker(autocommit=False, autoflush=False, bind=self.engine, future=True)
|
||||
self.SessionLocal = SessionLocal
|
||||
self.db = SessionLocal()
|
||||
self.current_user = {"id": "user-admin", "tenant_id": "tenant-001", "role": "admin"}
|
||||
self.service = LegalChangeImpactService(self.db)
|
||||
@@ -134,6 +136,54 @@ class LegalChangeImpactServiceTestCase(unittest.TestCase):
|
||||
self.assertEqual(result["completed"][0]["event"]["scan_status"], "completed")
|
||||
self.assertEqual(len(result["completed"][0]["impacts"]), 3)
|
||||
|
||||
def test_semantic_vector_scan_finds_related_contract_without_direct_reference(self):
|
||||
semantic_contract = GeneratedContract(
|
||||
id="contract-semantic-001",
|
||||
tenant_id="tenant-001",
|
||||
template_id="tpl-semantic",
|
||||
template_name="通用服务协议",
|
||||
contract_type="service",
|
||||
document_type_code="service_general",
|
||||
generated_by="user-admin",
|
||||
generated_by_name="Admin",
|
||||
export_format="docx",
|
||||
file_name="服务协议.docx",
|
||||
mime_type="application/vnd.openxmlformats-officedocument.wordprocessingml.document",
|
||||
file_size=456,
|
||||
generated_content="若任一方未按约履行交付义务，应承担迟延履行责任并补偿对方实际损失。",
|
||||
file_content_base64="SGVsbG8=",
|
||||
status="generated",
|
||||
)
|
||||
self.db.add(semantic_contract)
|
||||
self.db.commit()
|
||||
|
||||
result = self.service.scan_article_change(self.current_user, article_id="lart-civil-585")
|
||||
impact = next(item for item in result["impacts"] if item["target_id"] == "contract-semantic-001")
|
||||
|
||||
self.assertIn("semantic_vector_match", impact["match_reason"])
|
||||
self.assertEqual(impact["impact_level"], "reference_only")
|
||||
self.assertGreaterEqual(impact["evidence"]["semantic_vector"]["score"], 0.10)
|
||||
self.assertIn("breach_liability", impact["evidence"]["semantic_vector"]["shared_concepts"])
|
||||
self.assertGreaterEqual(result["event"]["impact_summary"]["semantic_match_count"], 1)
|
||||
|
||||
def test_background_worker_consumes_pending_events(self):
|
||||
queued = self.service.enqueue_article_change(
|
||||
self.current_user,
|
||||
article_id="lart-civil-585",
|
||||
change_type="updated",
|
||||
change_summary="后台 worker 消费",
|
||||
)
|
||||
|
||||
worker = LegalChangeImpactWorker(session_factory=self.SessionLocal)
|
||||
worker_result = worker.run_once(limit=1)
|
||||
self.db.expire_all()
|
||||
detail = self.service.get_event_detail(self.current_user, queued["event"]["id"])
|
||||
|
||||
self.assertEqual(worker_result["completed_count"], 1)
|
||||
self.assertEqual(worker_result["failed_count"], 0)
|
||||
self.assertEqual(detail["event"]["scan_status"], "completed")
|
||||
self.assertEqual(detail["event"]["impact_summary"]["total"], 3)
|
||||
|
||||
def test_update_impact_status_records_handling_note(self):
|
||||
result = self.service.scan_article_change(self.current_user, article_id="lart-civil-585")
|
||||
impact_id = result["impacts"][0]["id"]
|
||||
|
||||
@@ -38,7 +38,14 @@ class StubLegalChangeImpactService:
|
||||
"effective_date": "2021-01-01",
|
||||
"tags": ["违约金"],
|
||||
"scan_status": "completed",
|
||||
"impact_summary": {"total": 1, "blocking_count": 1, "by_level": {"immediate_update": 1}},
|
||||
"impact_summary": {
|
||||
"total": 1,
|
||||
"blocking_count": 1,
|
||||
"by_level": {"immediate_update": 1},
|
||||
"by_match_reason": {"semantic_vector_match": 1},
|
||||
"semantic_match_count": 1,
|
||||
"scan_strategy": ["direct_reference", "tag_match", "keyword_match", "semantic_vector_match"],
|
||||
},
|
||||
"created_by": "user-admin",
|
||||
"created_at": "2026-05-21T00:00:00+00:00",
|
||||
"scanned_at": "2026-05-21T00:01:00+00:00",
|
||||
@@ -61,7 +68,7 @@ class StubLegalChangeImpactService:
|
||||
"impact_level": "immediate_update",
|
||||
"suggestion": "请立即检查模板。",
|
||||
"impact_status": "open",
|
||||
"evidence": {},
|
||||
"evidence": {"semantic_vector": {"score": 0.42, "shared_concepts": ["breach_liability"]}},
|
||||
"handled_by": None,
|
||||
"handled_at": None,
|
||||
"created_at": "2026-05-21T00:01:00+00:00",
|
||||
@@ -151,6 +158,8 @@ class LegalChangeImpactsApiTestCase(unittest.TestCase):
|
||||
"run_pending_legal_change_scans",
|
||||
"update_legal_change_impact_status",
|
||||
])
|
||||
self.assertEqual(self.audit_log_service.events[0]["detail_json"]["semantic_match_count"], 1)
|
||||
self.assertTrue(self.audit_log_service.events[2]["detail_json"]["worker_compatible"])
|
||||
|
||||
def test_member_cannot_access_legal_change_impacts(self):
|
||||
app.dependency_overrides[require_current_user] = lambda: {
|
||||
|
||||
@@ -1,6 +1,7 @@
|
||||
import json
|
||||
import unittest
|
||||
from datetime import datetime, timezone
|
||||
from pathlib import Path
|
||||
|
||||
from fastapi import HTTPException
|
||||
|
||||
@@ -8,6 +9,7 @@ from app.models.prompt_asset import PromptAsset
|
||||
from app.models.prompt_asset_evaluation_run import PromptAssetEvaluationRun
|
||||
from app.models.prompt_asset_run_link import PromptAssetRunLink
|
||||
from app.models.prompt_asset_version import PromptAssetVersion
|
||||
from app.services.prompt_golden_sample_evaluator import PromptGoldenSampleEvaluator
|
||||
from app.services.prompt_asset_service import PromptAssetService
|
||||
|
||||
|
||||
@@ -117,6 +119,32 @@ class StubPromptAssetEvaluationRunRepository:
|
||||
return [item for item in self.items.values() if item.tenant_id == tenant_id and item.prompt_asset_id == asset_id][:limit]
|
||||
|
||||
|
||||
class FakeGoldenSampleLLMClient:
|
||||
def __init__(self):
|
||||
self.calls = []
|
||||
|
||||
def stream_answer(self, current_user, conversation_id, message, review_context=None, template_key=None):
|
||||
self.calls.append(
|
||||
{
|
||||
"current_user": current_user,
|
||||
"conversation_id": conversation_id,
|
||||
"message": message,
|
||||
"review_context": review_context,
|
||||
"template_key": template_key,
|
||||
}
|
||||
)
|
||||
yield json.dumps(
|
||||
{
|
||||
"passed": True,
|
||||
"score": 0.97,
|
||||
"observed_output_summary": "Sample expectations are satisfied.",
|
||||
"issues": [],
|
||||
"matched_expectations": ["expected_boundary", "legal_safety"],
|
||||
"model_trace": {"reason": "fake evaluator accepted this sample"},
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
class PromptAssetServiceTestCase(unittest.TestCase):
|
||||
def setUp(self):
|
||||
self.service = PromptAssetService(db=None)
|
||||
@@ -127,13 +155,21 @@ class PromptAssetServiceTestCase(unittest.TestCase):
|
||||
self.admin_user = {"id": "user-admin", "tenant_id": "tenant-demo-001", "role": "admin", "display_name": "Admin"}
|
||||
self.member_user = {"id": "user-member", "tenant_id": "tenant-demo-001", "role": "member"}
|
||||
|
||||
def _create_asset(self, evaluation=True):
|
||||
def _create_asset(
|
||||
self,
|
||||
evaluation=True,
|
||||
*,
|
||||
prompt_key="contract_generation_default",
|
||||
prompt_type="contract_generation",
|
||||
name="Contract generation default",
|
||||
prompt_content="Generate a contract from structured elements.",
|
||||
):
|
||||
return self.service.create_prompt_asset(
|
||||
self.admin_user,
|
||||
prompt_key="contract_generation_default",
|
||||
prompt_type="contract_generation",
|
||||
name="Contract generation default",
|
||||
prompt_content="Generate a contract from structured elements.",
|
||||
prompt_key=prompt_key,
|
||||
prompt_type=prompt_type,
|
||||
name=name,
|
||||
prompt_content=prompt_content,
|
||||
variables_json=json.dumps([{"name": "party_a", "required": True}]),
|
||||
output_schema_json=json.dumps({"type": "object"}),
|
||||
model_config_json=json.dumps({"model": "gpt-5.4"}),
|
||||
@@ -168,10 +204,42 @@ class PromptAssetServiceTestCase(unittest.TestCase):
|
||||
self.assertEqual(evaluation["asset"]["latest_evaluation_status"], "passed")
|
||||
self.assertEqual(evaluation["version"]["evaluation_status"], "passed")
|
||||
self.assertEqual(evaluation["evaluation_report"]["suite_key"], "phase13_generation")
|
||||
self.assertEqual(evaluation["evaluation_report"]["schema_version"], "phase15-prompt-evaluation-report.v2")
|
||||
self.assertEqual(evaluation["evaluation_report"]["evaluation_mode"], "local_fallback")
|
||||
self.assertGreaterEqual(evaluation["evaluation_run"]["sample_count"], 8)
|
||||
self.assertEqual(evaluation["evaluation_run"]["sample_count"], len(evaluation["evaluation_report"]["sample_results"]))
|
||||
self.assertEqual(evaluation["evaluation_run"]["status"], "passed")
|
||||
self.assertEqual(evaluation["evaluation_run"]["sample_count"], 8)
|
||||
self.assertEqual(published["status"], "published")
|
||||
|
||||
def test_run_evaluation_invokes_llm_once_per_golden_sample(self):
|
||||
fake_llm_client = FakeGoldenSampleLLMClient()
|
||||
self.service.golden_sample_evaluator = PromptGoldenSampleEvaluator(
|
||||
repo_root=Path.cwd(),
|
||||
llm_client=fake_llm_client,
|
||||
evaluation_mode="llm",
|
||||
)
|
||||
created = self._create_asset(
|
||||
evaluation=False,
|
||||
prompt_key="wording_rewrite_default",
|
||||
prompt_type="wording_rewrite",
|
||||
name="Wording rewrite default",
|
||||
prompt_content="Rewrite and optimize clause wording, explain risk boundaries, and preserve legal safety.",
|
||||
)
|
||||
|
||||
evaluation = self.service.evaluate_current_prompt_asset(self.admin_user, created["id"])
|
||||
|
||||
self.assertEqual(evaluation["version"]["evaluation_status"], "passed")
|
||||
self.assertEqual(evaluation["evaluation_report"]["suite_key"], "phase15_wording_rewrite")
|
||||
self.assertEqual(evaluation["evaluation_report"]["evaluation_mode"], "llm")
|
||||
self.assertGreaterEqual(evaluation["evaluation_run"]["sample_count"], 3)
|
||||
self.assertEqual(len(fake_llm_client.calls), evaluation["evaluation_run"]["sample_count"])
|
||||
self.assertEqual(
|
||||
evaluation["evaluation_report"]["model_invocation"]["sample_invocation_count"],
|
||||
evaluation["evaluation_run"]["sample_count"],
|
||||
)
|
||||
self.assertTrue(all(item["evaluator_mode"] == "llm" for item in evaluation["evaluation_report"]["sample_results"]))
|
||||
self.assertTrue(all(call["current_user"]["id"] == self.admin_user["id"] for call in fake_llm_client.calls))
|
||||
|
||||
def test_publish_with_report_and_create_next_version(self):
|
||||
created = self._create_asset(evaluation=False)
|
||||
published = self.service.publish_prompt_asset(
|
||||
|
||||
@@ -107,7 +107,16 @@ class StubPromptAssetService:
|
||||
"failed_samples": 0,
|
||||
"pass_rate": 1.0,
|
||||
"quality_score": 0.95,
|
||||
"report": {"suite_key": "phase13_generation", "passed": True},
|
||||
"report": {
|
||||
"schema_version": "phase15-prompt-evaluation-report.v2",
|
||||
"suite_key": "phase13_generation",
|
||||
"evaluation_mode": "llm",
|
||||
"sample_count": 8,
|
||||
"passed_samples": 8,
|
||||
"failed_samples": 0,
|
||||
"sample_results": [{"sample_id": "GEN-SAMPLE-TPL-001", "passed": True}],
|
||||
"passed": True,
|
||||
},
|
||||
"failure_reason": None,
|
||||
"triggered_by": "user-admin",
|
||||
"started_at": "2026-05-21T00:00:00+00:00",
|
||||
@@ -143,7 +152,16 @@ class StubPromptAssetService:
|
||||
"model_config": None,
|
||||
"evaluation_status": "passed",
|
||||
"evaluation_score": 0.95,
|
||||
"evaluation_report": {"suite_key": "phase13_generation", "passed": True},
|
||||
"evaluation_report": {
|
||||
"schema_version": "phase15-prompt-evaluation-report.v2",
|
||||
"suite_key": "phase13_generation",
|
||||
"evaluation_mode": "llm",
|
||||
"sample_count": 8,
|
||||
"passed_samples": 8,
|
||||
"failed_samples": 0,
|
||||
"sample_results": [{"sample_id": "GEN-SAMPLE-TPL-001", "passed": True}],
|
||||
"passed": True,
|
||||
},
|
||||
"change_summary": "Initial prompt version",
|
||||
"rollback_from_version_id": None,
|
||||
"is_current": True,
|
||||
@@ -154,7 +172,17 @@ class StubPromptAssetService:
|
||||
"published_at": None,
|
||||
"published_by": None,
|
||||
},
|
||||
"evaluation_report": {"suite_key": "phase13_generation", "passed": True, "quality_score": 0.95},
|
||||
"evaluation_report": {
|
||||
"schema_version": "phase15-prompt-evaluation-report.v2",
|
||||
"suite_key": "phase13_generation",
|
||||
"evaluation_mode": "llm",
|
||||
"sample_count": 8,
|
||||
"passed_samples": 8,
|
||||
"failed_samples": 0,
|
||||
"sample_results": [{"sample_id": "GEN-SAMPLE-TPL-001", "passed": True}],
|
||||
"passed": True,
|
||||
"quality_score": 0.95,
|
||||
},
|
||||
"evaluation_run": {
|
||||
"id": "prompteval-123",
|
||||
"tenant_id": "tenant-demo-001",
|
||||
@@ -169,7 +197,16 @@ class StubPromptAssetService:
|
||||
"failed_samples": 0,
|
||||
"pass_rate": 1.0,
|
||||
"quality_score": 0.95,
|
||||
"report": {"suite_key": "phase13_generation", "passed": True},
|
||||
"report": {
|
||||
"schema_version": "phase15-prompt-evaluation-report.v2",
|
||||
"suite_key": "phase13_generation",
|
||||
"evaluation_mode": "llm",
|
||||
"sample_count": 8,
|
||||
"passed_samples": 8,
|
||||
"failed_samples": 0,
|
||||
"sample_results": [{"sample_id": "GEN-SAMPLE-TPL-001", "passed": True}],
|
||||
"passed": True,
|
||||
},
|
||||
"failure_reason": None,
|
||||
"triggered_by": current_user["id"],
|
||||
"started_at": "2026-05-21T00:00:00+00:00",
|
||||
@@ -264,6 +301,7 @@ class PromptAssetsApiTestCase(unittest.TestCase):
|
||||
self.assertEqual(version_response.json()["data"]["current_version_label"], "v1.1")
|
||||
self.assertEqual(evaluation_response.json()["data"]["version"]["evaluation_status"], "passed")
|
||||
self.assertEqual(evaluation_response.json()["data"]["evaluation_run"]["status"], "passed")
|
||||
self.assertEqual(evaluation_response.json()["data"]["evaluation_report"]["evaluation_mode"], "llm")
|
||||
self.assertEqual(publish_response.json()["data"]["status"], "published")
|
||||
self.assertEqual(rollback_response.json()["data"]["current_version_label"], "v1.2")
|
||||
self.assertEqual(run_link_response.json()["data"]["prompt_asset_version_id"], "promptver-123")
|
||||
@@ -275,6 +313,8 @@ class PromptAssetsApiTestCase(unittest.TestCase):
|
||||
"rollback_prompt_asset",
|
||||
"record_prompt_asset_run_link",
|
||||
])
|
||||
self.assertEqual(self.audit_log_service.events[2]["detail_json"]["evaluation_mode"], "llm")
|
||||
self.assertEqual(self.audit_log_service.events[2]["detail_json"]["sample_count"], 8)
|
||||
|
||||
def test_member_cannot_list_prompt_assets(self):
|
||||
app.dependency_overrides[require_current_user] = lambda: {
|
||||
|
||||
@@ -55,6 +55,43 @@ services:
|
||||
ports:
|
||||
- "8001:8000"
|
||||
|
||||
legal-change-impact-worker:
|
||||
build:
|
||||
context: ./backend
|
||||
dockerfile: Dockerfile
|
||||
restart: unless-stopped
|
||||
depends_on:
|
||||
postgres:
|
||||
condition: service_healthy
|
||||
backend:
|
||||
condition: service_started
|
||||
environment:
|
||||
APP_NAME: AI Contract Platform
|
||||
APP_ENV: production
|
||||
SECRET_KEY: ${SECRET_KEY:?Set SECRET_KEY in .env before starting production}
|
||||
DATABASE_URL: postgresql+psycopg://postgres:postgres@postgres:5432/ai_contract_platform
|
||||
REDIS_URL: redis://redis:6379/0
|
||||
MINIO_ENDPOINT: 192.168.3.90:9000
|
||||
MINIO_ACCESS_KEY: minioadmin
|
||||
MINIO_SECRET_KEY: minioadmin
|
||||
MINIO_BUCKET: contracts
|
||||
SHARED_AI_PLATFORM_BASE_URL: http://192.168.3.90:18011
|
||||
SHARED_AI_PLATFORM_API_PREFIX: /api/v0.1
|
||||
SHARED_AI_PLATFORM_APP_ID: ai-contract-platform
|
||||
SHARED_AI_PLATFORM_API_KEY: "shared-ai-platform-admin-dev-token"
|
||||
SHARED_AI_PLATFORM_TIMEOUT_SECONDS: 180
|
||||
SHARED_AI_PLATFORM_LEGAL_REVIEW_ENABLED: "true"
|
||||
SHARED_AI_PLATFORM_LEGAL_REVIEW_FALLBACK_TO_LOCAL: "false"
|
||||
LEGAL_CHANGE_IMPACT_WORKER_BATCH_SIZE: 10
|
||||
LEGAL_CHANGE_IMPACT_WORKER_INTERVAL_SECONDS: 300
|
||||
LEGAL_CHANGE_IMPACT_SEMANTIC_ENABLED: "true"
|
||||
command: >
|
||||
python -m app.tasks.legal_change_impact_worker
|
||||
--limit 10
|
||||
--interval-seconds 300
|
||||
volumes:
|
||||
- backend_storage:/app/storage
|
||||
|
||||
frontend:
|
||||
build:
|
||||
context: ./frontend
|
||||
|
||||
@@ -57,6 +57,16 @@ const clausePriorityColorMap: Record<string, string> = {
|
||||
low: "blue",
|
||||
};
|
||||
|
||||
const recommendationKindLabelMap: Record<string, string> = {
|
||||
missing_clause: "缺失条款",
|
||||
industry_practice: "行业惯例",
|
||||
};
|
||||
|
||||
const recommendationKindColorMap: Record<string, string> = {
|
||||
missing_clause: "orange",
|
||||
industry_practice: "purple",
|
||||
};
|
||||
|
||||
function formatDateTime(value?: string | null) {
|
||||
if (!value) {
|
||||
return "-";
|
||||
@@ -111,6 +121,10 @@ function getErrorMessage(error: unknown, fallback: string) {
|
||||
return error instanceof Error && error.message ? error.message : fallback;
|
||||
}
|
||||
|
||||
function readableRecommendationKind(kind?: string | null) {
|
||||
return recommendationKindLabelMap[kind ?? ""] ?? "条款建议";
|
||||
}
|
||||
|
||||
export default function GeneratedContractEditorPage({ params }: { params: Promise<{ id: string }> }) {
|
||||
const [messageApi, contextHolder] = message.useMessage();
|
||||
const router = useRouter();
|
||||
@@ -474,7 +488,7 @@ export default function GeneratedContractEditorPage({ params }: { params: Promis
|
||||
showIcon
|
||||
type="info"
|
||||
message="编辑器说明"
|
||||
description="左侧正文可直接编辑；右侧可选择段落、检查缺失条款、改写条款并把结果插入正文。保存后会生成可追溯的新版本。"
|
||||
description="左侧正文可直接编辑；右侧可选择段落、检查缺失条款和行业惯例、改写条款并把结果插入正文。保存后会生成可追溯的新版本。"
|
||||
/>
|
||||
</section>
|
||||
|
||||
@@ -486,7 +500,7 @@ export default function GeneratedContractEditorPage({ params }: { params: Promis
|
||||
</div>
|
||||
<Space direction="vertical" size="small" style={{ width: "100%" }}>
|
||||
<Button block loading={recommendationLoading} onClick={() => void handleRecommend()}>
|
||||
检查缺失条款
|
||||
检查条款与惯例
|
||||
</Button>
|
||||
<Button block icon={<MessageOutlined />} onClick={() => router.push(buildEditorChatHref(contract, editorContent))}>
|
||||
带正文追问 Chat
|
||||
@@ -551,7 +565,10 @@ export default function GeneratedContractEditorPage({ params }: { params: Promis
|
||||
<section className="document-side-panel">
|
||||
<div className="document-side-panel-header">
|
||||
<BulbOutlined />
|
||||
<strong>缺失条款推荐</strong>
|
||||
<strong>条款与行业惯例推荐</strong>
|
||||
{recommendations?.summary.industry_practice_count ? (
|
||||
<Tag color="purple">{recommendations.summary.industry_practice_count} 条惯例</Tag>
|
||||
) : null}
|
||||
</div>
|
||||
{recommendations?.items.length ? (
|
||||
<div className="document-editor-recommend-list">
|
||||
@@ -559,10 +576,16 @@ export default function GeneratedContractEditorPage({ params }: { params: Promis
|
||||
<article className="document-editor-recommend-card" key={item.id}>
|
||||
<div>
|
||||
<strong>{item.clause_title}</strong>
|
||||
<Tag color={recommendationKindColorMap[item.recommendation_kind ?? ""] ?? "blue"}>
|
||||
{readableRecommendationKind(item.recommendation_kind)}
|
||||
</Tag>
|
||||
<Tag color={clausePriorityColorMap[item.priority] ?? "default"}>{item.priority}</Tag>
|
||||
</div>
|
||||
<Text type="secondary">{item.insert_position.description}</Text>
|
||||
<Paragraph>{item.suggested_text}</Paragraph>
|
||||
<Text type="secondary">
|
||||
{item.practice_basis ?? item.legal_basis}
|
||||
</Text>
|
||||
<Button size="small" type="primary" onClick={() => void handleApplyRecommendation(item)}>
|
||||
插入正文并生成版本
|
||||
</Button>
|
||||
@@ -570,9 +593,9 @@ export default function GeneratedContractEditorPage({ params }: { params: Promis
|
||||
))}
|
||||
</div>
|
||||
) : recommendations ? (
|
||||
<Text type="secondary">当前未发现首批必备条款缺失。</Text>
|
||||
<Text type="secondary">当前未发现必备条款缺失或行业惯例建议。</Text>
|
||||
) : (
|
||||
<Text type="secondary">点击“检查缺失条款”后展示推荐。</Text>
|
||||
<Text type="secondary">点击“检查条款与惯例”后展示推荐。</Text>
|
||||
)}
|
||||
</section>
|
||||
|
||||
|
||||
@@ -134,6 +134,16 @@ const clausePriorityLabelMap: Record<string, string> = {
|
||||
low: "低优先级",
|
||||
};
|
||||
|
||||
const recommendationKindLabelMap: Record<string, string> = {
|
||||
missing_clause: "缺失条款",
|
||||
industry_practice: "行业惯例",
|
||||
};
|
||||
|
||||
const recommendationKindColorMap: Record<string, string> = {
|
||||
missing_clause: "orange",
|
||||
industry_practice: "purple",
|
||||
};
|
||||
|
||||
const clauseRewriteModeOptions = [
|
||||
{ label: "专业化", value: "professional" },
|
||||
{ label: "简化解释", value: "simplify" },
|
||||
@@ -224,6 +234,10 @@ function getErrorMessage(error: unknown, fallback: string) {
|
||||
return error instanceof Error && error.message ? error.message : fallback;
|
||||
}
|
||||
|
||||
function readableRecommendationKind(kind?: string | null) {
|
||||
return recommendationKindLabelMap[kind ?? ""] ?? "条款建议";
|
||||
}
|
||||
|
||||
function renderField(field: LegalDocumentElementField) {
|
||||
if (field.type === "select") {
|
||||
return <Select options={normalizeOptions(field)} placeholder={`请选择${readableFieldLabel(field)}`} />;
|
||||
@@ -767,17 +781,22 @@ export default function LegalDocumentBuilderPage({ params }: { params: Promise<{
|
||||
|
||||
<div className="document-clause-recommendations">
|
||||
<div className="document-clause-recommendations-header">
|
||||
<strong><BulbOutlined /> 智能条款推荐</strong>
|
||||
<strong><BulbOutlined /> 智能条款与行业惯例推荐</strong>
|
||||
{clauseRecommendations ? (
|
||||
<Tag color={clauseRecommendations.total ? "orange" : "green"}>
|
||||
{clauseRecommendations.total ? `${clauseRecommendations.total} 条建议` : "暂无缺失条款"}
|
||||
</Tag>
|
||||
<Space size={6} wrap>
|
||||
<Tag color={clauseRecommendations.total ? "orange" : "green"}>
|
||||
{clauseRecommendations.total ? `${clauseRecommendations.total} 条建议` : "暂无建议"}
|
||||
</Tag>
|
||||
{clauseRecommendations.summary.industry_practice_count ? (
|
||||
<Tag color="purple">{clauseRecommendations.summary.industry_practice_count} 条行业惯例</Tag>
|
||||
) : null}
|
||||
</Space>
|
||||
) : null}
|
||||
</div>
|
||||
{clauseRecommendationLoading ? (
|
||||
<div className="document-clause-loading">
|
||||
<Spin size="small" />
|
||||
<Text type="secondary">正在检查缺失条款...</Text>
|
||||
<Text type="secondary">正在检查缺失条款和行业惯例...</Text>
|
||||
</div>
|
||||
) : clauseRecommendations?.items.length ? (
|
||||
<div className="document-clause-list">
|
||||
@@ -786,10 +805,17 @@ export default function LegalDocumentBuilderPage({ params }: { params: Promise<{
|
||||
<div className="document-clause-card-head">
|
||||
<strong>{item.clause_title}</strong>
|
||||
<Space size={6} wrap>
|
||||
<Tag color={recommendationKindColorMap[item.recommendation_kind ?? ""] ?? "blue"}>
|
||||
{readableRecommendationKind(item.recommendation_kind)}
|
||||
</Tag>
|
||||
<Tag color={clausePriorityColorMap[item.priority] ?? "default"}>
|
||||
{clausePriorityLabelMap[item.priority] ?? item.priority}
|
||||
</Tag>
|
||||
<Tag>{item.confidence === "high" ? "依据充分" : "规则建议"}</Tag>
|
||||
<Tag>
|
||||
{item.basis_type === "industry_practice_only"
|
||||
? "无直接法条"
|
||||
: item.confidence === "high" ? "依据充分" : "规则建议"}
|
||||
</Tag>
|
||||
<Button
|
||||
size="small"
|
||||
icon={<EditOutlined />}
|
||||
@@ -810,6 +836,8 @@ export default function LegalDocumentBuilderPage({ params }: { params: Promise<{
|
||||
<div className="document-clause-meta">
|
||||
<span>推荐理由：{item.reason}</span>
|
||||
<span>依据：{item.legal_basis}</span>
|
||||
{item.practice_basis ? <span>惯例说明：{item.practice_basis}</span> : null}
|
||||
{item.industry_context ? <span>适用场景：{item.industry_context}</span> : null}
|
||||
</div>
|
||||
{item.legal_article_refs.length ? (
|
||||
<div className="document-clause-refs">
|
||||
@@ -824,7 +852,7 @@ export default function LegalDocumentBuilderPage({ params }: { params: Promise<{
|
||||
))}
|
||||
</div>
|
||||
) : (
|
||||
<Text type="secondary">当前生成正文暂未发现需要补充的首批必备条款。</Text>
|
||||
<Text type="secondary">当前生成正文暂未发现需要补充的必备条款或行业惯例建议。</Text>
|
||||
)}
|
||||
</div>
|
||||
|
||||
|
||||
@@ -86,6 +86,29 @@ function impactStatusTag(value: string) {
|
||||
return <Tag color={meta.color}>{meta.label}</Tag>;
|
||||
}
|
||||
|
||||
function matchReasonTags(item: LegalChangeImpactItem) {
|
||||
const reasons = item.match_reason.split(",").filter(Boolean);
|
||||
return reasons.map((reason) => {
|
||||
if (reason === "semantic_vector_match") return <Tag color="purple" key={reason}>语义向量</Tag>;
|
||||
if (reason === "direct_legal_reference") return <Tag color="red" key={reason}>直接引用</Tag>;
|
||||
if (reason === "tag_match") return <Tag color="blue" key={reason}>标签</Tag>;
|
||||
if (reason === "keyword_match") return <Tag color="gold" key={reason}>关键词</Tag>;
|
||||
return <Tag key={reason}>{reason}</Tag>;
|
||||
});
|
||||
}
|
||||
|
||||
function getSemanticEvidence(item: LegalChangeImpactItem) {
|
||||
const semantic = item.evidence?.semantic_vector;
|
||||
if (!semantic || typeof semantic !== "object") {
|
||||
return null;
|
||||
}
|
||||
const payload = semantic as { score?: unknown; shared_concepts?: unknown };
|
||||
return {
|
||||
score: typeof payload.score === "number" ? payload.score : null,
|
||||
sharedConcepts: Array.isArray(payload.shared_concepts) ? payload.shared_concepts.map(String) : [],
|
||||
};
|
||||
}
|
||||
|
||||
export default function LegalChangeImpactsPage() {
|
||||
const [messageApi, contextHolder] = message.useMessage();
|
||||
const currentUser = useAuthStore((state) => state.currentUser);
|
||||
@@ -294,6 +317,7 @@ export default function LegalChangeImpactsPage() {
|
||||
<Text type={(item.impact_summary?.blocking_count ?? 0) > 0 ? "danger" : "secondary"}>
|
||||
立即修改：{item.impact_summary?.blocking_count ?? 0}
|
||||
</Text>
|
||||
<Text type="secondary">语义命中：{item.impact_summary?.semantic_match_count ?? 0}</Text>
|
||||
</Space>
|
||||
),
|
||||
},
|
||||
@@ -361,14 +385,25 @@ export default function LegalChangeImpactsPage() {
|
||||
title: "命中原因",
|
||||
key: "match",
|
||||
width: 260,
|
||||
render: (_, item) => (
|
||||
<Space wrap size={4}>
|
||||
{item.match_terms.slice(0, 5).map((term) => (
|
||||
<Tag key={term}>{term}</Tag>
|
||||
))}
|
||||
{item.match_terms.length > 5 ? <Tag>+{item.match_terms.length - 5}</Tag> : null}
|
||||
</Space>
|
||||
),
|
||||
render: (_, item) => {
|
||||
const semantic = getSemanticEvidence(item);
|
||||
return (
|
||||
<Space direction="vertical" size={4}>
|
||||
<Space wrap size={4}>{matchReasonTags(item)}</Space>
|
||||
<Space wrap size={4}>
|
||||
{item.match_terms.slice(0, 5).map((term) => (
|
||||
<Tag key={term}>{term}</Tag>
|
||||
))}
|
||||
{item.match_terms.length > 5 ? <Tag>+{item.match_terms.length - 5}</Tag> : null}
|
||||
</Space>
|
||||
{semantic ? (
|
||||
<Text type="secondary">
|
||||
相似度 {semantic.score ?? "-"} {semantic.sharedConcepts.length ? ` / ${semantic.sharedConcepts.join(", ")}` : ""}
|
||||
</Text>
|
||||
) : null}
|
||||
</Space>
|
||||
);
|
||||
},
|
||||
},
|
||||
{
|
||||
title: "处理建议",
|
||||
@@ -419,7 +454,7 @@ export default function LegalChangeImpactsPage() {
|
||||
<div className="review-page-hero-copy">
|
||||
<span className="landing-badge">Regulation Impact</span>
|
||||
<Title level={2}>法规变更影响分析</Title>
|
||||
<Text type="secondary">当法条新增、修改、生效状态变化时，主动扫描模板、审查规则和历史合同，给出处理等级和建议。</Text>
|
||||
<Text type="secondary">当法条新增、修改、生效状态变化时，主动扫描模板、审查规则和历史合同，结合直接引用、标签关键词和语义向量命中给出处理等级和建议。</Text>
|
||||
</div>
|
||||
<Card className="review-side-card" variant="borderless">
|
||||
<span className="landing-panel-eyebrow">Phase15-G</span>
|
||||
@@ -429,7 +464,8 @@ export default function LegalChangeImpactsPage() {
|
||||
<ul className="review-page-tip-list">
|
||||
<li>立即修改：模板或规则直接引用法条。</li>
|
||||
<li>续约时修改：历史合同直接引用法条。</li>
|
||||
<li>仅供参考：标签或关键词相关。</li>
|
||||
<li>仅供参考：标签、关键词或语义向量相关。</li>
|
||||
<li>后台 worker 可定时消费待扫描队列。</li>
|
||||
</ul>
|
||||
</Card>
|
||||
</section>
|
||||
|
||||
@@ -36,7 +36,7 @@ import type {
|
||||
PromptAssetVersionItem,
|
||||
} from "@/types/api";
|
||||
|
||||
const { Title, Text, Paragraph } = Typography;
|
||||
const { Title, Text } = Typography;
|
||||
|
||||
type PromptAssetFormValues = {
|
||||
prompt_key: string;
|
||||
@@ -108,6 +108,19 @@ function getEvaluationTag(value?: string | null) {
|
||||
return <Tag color="default">未评测</Tag>;
|
||||
}
|
||||
|
||||
function getEvaluationModeTag(report?: Record<string, unknown> | null) {
|
||||
const mode = typeof report?.evaluation_mode === "string" ? report.evaluation_mode : null;
|
||||
if (mode === "llm") return <Tag color="purple">逐条大模型</Tag>;
|
||||
if (mode === "local_fallback") return <Tag color="blue">本地兜底</Tag>;
|
||||
if (mode === "llm_required_not_configured") return <Tag color="red">模型未配置</Tag>;
|
||||
return <Tag>未知</Tag>;
|
||||
}
|
||||
|
||||
function getEvaluationSampleResultCount(report?: Record<string, unknown> | null) {
|
||||
const results = report?.sample_results;
|
||||
return Array.isArray(results) ? results.length : 0;
|
||||
}
|
||||
|
||||
function getPreviousVersionLabel(currentVersionLabel: string) {
|
||||
const [, minorValue] = currentVersionLabel.replace(/^v/, "").split(".");
|
||||
const minor = Number(minorValue ?? "0");
|
||||
@@ -303,7 +316,8 @@ export default function PromptAssetsPage() {
|
||||
return;
|
||||
}
|
||||
const statusText = result.data.version.evaluation_status === "passed" ? "通过" : "未通过";
|
||||
messageApi.success(`Prompt 评测已完成：${statusText}`);
|
||||
const modeText = result.data.evaluation_report.evaluation_mode === "llm" ? "逐条大模型" : "本地兜底";
|
||||
messageApi.success(`Prompt 评测已完成：${statusText}，${modeText}`);
|
||||
await refreshAfterMutation(item.id);
|
||||
} catch (error) {
|
||||
messageApi.error(error instanceof Error ? error.message : "Prompt 评测失败");
|
||||
@@ -368,30 +382,21 @@ export default function PromptAssetsPage() {
|
||||
|
||||
const handleRecordRunLink = async (values: RunLinkFormValues) => {
|
||||
if (!selectedDetail) {
|
||||
messageApi.warning("请先选择一个 Prompt 资产");
|
||||
return;
|
||||
}
|
||||
setMutatingId(selectedDetail.asset.id);
|
||||
try {
|
||||
assertJson(values.trace_json);
|
||||
const result = await recordPromptRunLink(selectedDetail.asset.id, {
|
||||
run_type: values.run_type,
|
||||
run_id: values.run_id?.trim(),
|
||||
resource_type: values.resource_type?.trim(),
|
||||
resource_id: values.resource_id?.trim(),
|
||||
workflow_run_id: values.workflow_run_id?.trim(),
|
||||
task_run_id: values.task_run_id?.trim(),
|
||||
trace_json: values.trace_json?.trim(),
|
||||
});
|
||||
const result = await recordPromptRunLink(selectedDetail.asset.id, values);
|
||||
if (result.code !== 0) {
|
||||
messageApi.error(result.message || "运行证据登记失败");
|
||||
return;
|
||||
}
|
||||
messageApi.success("运行证据已登记");
|
||||
runLinkForm.resetFields();
|
||||
await loadDetail(selectedDetail.asset.id);
|
||||
await refreshAfterMutation(selectedDetail.asset.id);
|
||||
} catch (error) {
|
||||
messageApi.error(error instanceof Error ? error.message : "运行证据登记失败，请检查 JSON 格式");
|
||||
messageApi.error(error instanceof Error ? error.message : "运行证据登记失败，请检查 Trace JSON");
|
||||
} finally {
|
||||
setMutatingId(null);
|
||||
}
|
||||
@@ -447,7 +452,7 @@ export default function PromptAssetsPage() {
|
||||
{
|
||||
title: "评测",
|
||||
key: "evaluation",
|
||||
width: 160,
|
||||
width: 170,
|
||||
render: (_, item) => (
|
||||
<Space orientation="vertical" size={0}>
|
||||
{getEvaluationTag(item.latest_evaluation_status)}
|
||||
@@ -465,7 +470,7 @@ export default function PromptAssetsPage() {
|
||||
{
|
||||
title: "操作",
|
||||
key: "actions",
|
||||
width: 260,
|
||||
width: 270,
|
||||
render: (_, item) => (
|
||||
<Space wrap>
|
||||
<Button icon={<HistoryOutlined />} loading={mutatingId === item.id} size="small" onClick={() => void handleEdit(item)}>
|
||||
@@ -512,9 +517,9 @@ export default function PromptAssetsPage() {
|
||||
{
|
||||
title: "状态",
|
||||
key: "status",
|
||||
width: 180,
|
||||
width: 190,
|
||||
render: (_, item) => (
|
||||
<Space>
|
||||
<Space wrap>
|
||||
{item.is_current ? <Tag color="blue">当前</Tag> : null}
|
||||
{item.is_published ? <Tag color="green">生产</Tag> : null}
|
||||
{getEvaluationTag(item.evaluation_status)}
|
||||
@@ -546,9 +551,20 @@ export default function PromptAssetsPage() {
|
||||
{
|
||||
title: "状态",
|
||||
key: "status",
|
||||
width: 130,
|
||||
width: 120,
|
||||
render: (_, item) => getEvaluationTag(item.status),
|
||||
},
|
||||
{
|
||||
title: "执行方式",
|
||||
key: "mode",
|
||||
width: 150,
|
||||
render: (_, item) => (
|
||||
<Space direction="vertical" size={0}>
|
||||
{getEvaluationModeTag(item.report)}
|
||||
<Text type="secondary">逐条 {getEvaluationSampleResultCount(item.report)} 条</Text>
|
||||
</Space>
|
||||
),
|
||||
},
|
||||
{
|
||||
title: "样本",
|
||||
key: "samples",
|
||||
@@ -596,17 +612,17 @@ export default function PromptAssetsPage() {
|
||||
<div className="review-page-hero-copy">
|
||||
<span className="landing-badge">Prompt Governance</span>
|
||||
<Title level={2}>Prompt 资产治理</Title>
|
||||
<Text type="secondary">生成、审查、条款推荐、法规问答和措辞改写的 Prompt 可以版本化、评测、发布、回滚并绑定运行证据。</Text>
|
||||
<Text type="secondary">生成、审查、条款推荐、法规问答和措辞改写的 Prompt 可版本化、评测、发布、回滚并绑定运行证据。</Text>
|
||||
</div>
|
||||
<Card className="review-side-card" variant="borderless">
|
||||
<span className="landing-panel-eyebrow">Phase15-E</span>
|
||||
<span className="landing-panel-eyebrow">Phase15-F</span>
|
||||
<Title level={4} style={{ margin: 0, color: "inherit" }}>
|
||||
发布门禁
|
||||
黄金样本评测门禁
|
||||
</Title>
|
||||
<ul className="review-page-tip-list">
|
||||
<li>发布前必须有通过的评测报告。</li>
|
||||
<li>回滚会生成一个新的当前版本。</li>
|
||||
<li>运行证据可绑定到当前生产版本。</li>
|
||||
<li>已配置共享 AI 时会逐条调用大模型评测黄金样本。</li>
|
||||
<li>未配置模型时保留本地兜底结果，便于开发环境验证流程。</li>
|
||||
</ul>
|
||||
</Card>
|
||||
</section>
|
||||
@@ -723,9 +739,9 @@ export default function PromptAssetsPage() {
|
||||
}
|
||||
>
|
||||
<Space orientation="vertical" size="middle" style={{ width: "100%" }}>
|
||||
<Paragraph type="secondary" style={{ margin: 0 }}>
|
||||
<Text type="secondary">
|
||||
当前状态：{getStatusTag(selectedDetail.asset.status)} 最新评测：{getEvaluationTag(selectedDetail.asset.latest_evaluation_status)}
|
||||
</Paragraph>
|
||||
</Text>
|
||||
<Table rowKey="id" size="small" pagination={false} columns={versionColumns} dataSource={selectedDetail.versions} />
|
||||
<Space orientation="vertical" size="small" style={{ width: "100%" }}>
|
||||
<Text strong>评测运行历史</Text>
|
||||
|
||||
@@ -23,12 +23,16 @@ export interface ClauseRecommendationItem {
|
||||
clause_key: string;
|
||||
clause_title: string;
|
||||
clause_type: string;
|
||||
recommendation_kind?: "missing_clause" | "industry_practice" | string;
|
||||
basis_type?: string | null;
|
||||
industry_context?: string | null;
|
||||
priority: string;
|
||||
confidence: string;
|
||||
insert_position: ClauseRecommendationInsertPosition;
|
||||
suggested_text: string;
|
||||
reason: string;
|
||||
legal_basis: string;
|
||||
practice_basis?: string | null;
|
||||
legal_article_refs: ClauseRecommendationLegalArticleRef[];
|
||||
source_rule: string;
|
||||
source_types: string[];
|
||||
@@ -43,8 +47,13 @@ export interface ClauseRecommendationResponse {
|
||||
total: number;
|
||||
summary: {
|
||||
recommended_count: number;
|
||||
recommendation_keys?: string[];
|
||||
detected_clause_keys: string[];
|
||||
missing_clause_keys: string[];
|
||||
required_missing_clause_keys?: string[];
|
||||
industry_practice_keys?: string[];
|
||||
detected_industry_practice_keys?: string[];
|
||||
industry_practice_count?: number;
|
||||
source_count: number;
|
||||
strategy: string;
|
||||
};
|
||||
|
||||
@@ -380,6 +380,9 @@ export type LegalChangeEventItem = {
|
||||
blocking_count?: number;
|
||||
by_level?: Record<string, number>;
|
||||
by_target_type?: Record<string, number>;
|
||||
by_match_reason?: Record<string, number>;
|
||||
semantic_match_count?: number;
|
||||
scan_strategy?: string[];
|
||||
};
|
||||
created_by: string | null;
|
||||
created_at: string | null;
|
||||
|
||||
@@ -4108,3 +4108,119 @@
|
||||
- 剩余边界：
|
||||
- 本轮尚未部署到 `192.168.3.90`；测试服务器使用该能力前需要执行 Alembic `20260521_0038` 并部署前后端。
|
||||
- 当前运行仍是同步轻量评测，适合测试服务器资源；后续如果要跑真实大模型逐条黄金样本，可在这张运行表上继续扩展异步队列和 worker。
|
||||
|
||||
## 2026-05-21 Phase15-C 行业惯例推荐完整闭环
|
||||
|
||||
- 已按《AI合同起草与审查助手-法律文书能力吸收开发方案-2026-05-20.md》回补 Phase15-C 最后一块缺口：行业惯例推荐。
|
||||
- 后端能力：
|
||||
- 在既有 `POST /api/v1/clauses/recommend` 中扩展行业惯例推荐，不新增重复入口。
|
||||
- 新增 `INDUSTRY_PRACTICE_RULES`，覆盖通用服务、技术服务、软件开发、劳动合同、保密协议等首批高频场景。
|
||||
- 推荐结果新增 `recommendation_kind=industry_practice`、`basis_type=industry_practice_only`、`industry_context`、`practice_basis` 等字段。
|
||||
- 对“无直接法条，仅为行业惯例”的建议不强行挂法条，结果会明确返回 `no_direct_legal_article` 来源标记。
|
||||
- `summary` 新增 `industry_practice_keys`、`industry_practice_count`、`detected_industry_practice_keys`、`recommendation_keys`，QA 可直接校验。
|
||||
- 审计日志 `recommend_missing_clauses` 追加记录推荐 key、行业惯例 key 和行业惯例数量。
|
||||
- 前端能力：
|
||||
- `/documents/new/[type]` 生成结果区从“智能条款推荐”升级为“智能条款与行业惯例推荐”。
|
||||
- 推荐卡片可区分“缺失条款”和“行业惯例”，并展示“无直接法条”、惯例说明、适用场景。
|
||||
- `/documents/[id]/edit` AI 侧边栏从“检查缺失条款”升级为“检查条款与惯例”。
|
||||
- 行业惯例建议可像普通推荐条款一样插入正文，并继续生成可追溯合同版本。
|
||||
- 本轮验证：
|
||||
- `python -m compileall backend\app\services\clause_recommendation_service.py backend\app\api\clauses.py`
|
||||
- `$env:PYTHONPATH='backend'; python -m unittest backend.tests.test_clause_recommendation_service backend.tests.test_clauses_api`
|
||||
- `$env:PYTHONPATH='backend'; python -m unittest backend.tests.test_contracts_api backend.tests.test_clause_recommendation_service backend.tests.test_clauses_api`
|
||||
- `cd frontend; npm run lint`
|
||||
- `cd frontend; npm run build`
|
||||
- `git diff --check`
|
||||
- Browser 打开 `http://localhost:3000/documents/new`，未登录保护页正常显示，页面标题 `AI合同平台`，控制台错误 0。
|
||||
- 当前闭环结论：
|
||||
- Phase15-C 已形成“必备条款缺口检测 -> 行业惯例补充建议 -> 明确无直接法条/惯例来源 -> 页面展示 -> 一键插入合同正文 -> 生成版本 -> 审计留痕”的完整开发闭环。
|
||||
- 剩余边界：
|
||||
- 本轮按“先完成开发再统一部署”的原则，没有部署到 `192.168.3.90`。
|
||||
- 该能力不需要新增数据库迁移；后续统一部署前端和后端后即可进入测试服务器烟测。
|
||||
|
||||
## 2026-05-21 Phase15-F 真实黄金样本逐条评测闭环
|
||||
|
||||
- 已继续按《AI合同起草与审查助手-法律文书能力吸收开发方案-2026-05-20.md》补齐 Phase15-F 的上线前增强项。
|
||||
- 本轮把上一版“同步轻量评测”升级为“真实黄金样本逐条评测”：
|
||||
- 新增 `PromptGoldenSampleEvaluator`，直接加载项目内 Phase13/Phase15 黄金样本清单。
|
||||
- 覆盖合同/文书生成 10 条、自动审查 8 条、法规问答 10 条、条款推荐 8 条、措辞改写 3 条首批黄金样本。
|
||||
- 每次评测会生成 `phase15-prompt-evaluation-report.v2` 报告，并写入 `prompt_asset_evaluation_runs.report_json`。
|
||||
- 报告新增 `evaluation_mode`、`model_invocation`、`sample_results`、逐条 `sample_id / passed / score / issues / matched_expectations`。
|
||||
- 已配置 `SHARED_AI_PLATFORM_BASE_URL` 和 `SHARED_AI_PLATFORM_API_KEY` 时，会按黄金样本逐条调用共享 AI 平台大模型评测。
|
||||
- 未配置共享 AI 时默认走 `auto -> local_fallback`，开发环境仍可验证评测记录、发布门禁、前端展示；如将 `PROMPT_GOLDEN_EVAL_MODE=llm`，未配置模型会明确失败，不会假装通过。
|
||||
- 后端闭环：
|
||||
- `POST /api/v1/prompt-assets/{asset_id}/evaluations/run` 现在会把真实样本评测报告回填到当前 Prompt 版本、资产最新评测状态和评测运行记录。
|
||||
- 审计日志 `run_prompt_asset_evaluation` 追加记录评测模式、样本数、通过数、失败数。
|
||||
- 发布门禁继续复用最新评测状态；未通过逐条样本评测时不可发布。
|
||||
- 前端闭环：
|
||||
- 重写并修复 `/prompt-assets` 页面乱码坏句，保留原有 Prompt 创建、版本、发布、回滚、运行证据能力。
|
||||
- 评测历史新增“执行方式”，可区分“逐条大模型 / 本地兜底 / 模型未配置”。
|
||||
- 评测历史展示逐条样本结果数量，QA 可以从页面确认是否已经按样本逐条执行。
|
||||
- 本轮验证：
|
||||
- `python -m compileall backend\app\services\prompt_golden_sample_evaluator.py backend\app\services\prompt_asset_service.py backend\app\api\prompt_assets.py`
|
||||
- `$env:PYTHONPATH='backend'; python -m unittest backend.tests.test_prompt_asset_service backend.tests.test_prompt_assets_api`
|
||||
- `cd frontend; npx eslint src/app/prompt-assets/page.tsx src/types/api.ts`
|
||||
- `cd frontend; npm run build`
|
||||
- `git diff --check`
|
||||
- 当前闭环结论：
|
||||
- Phase15-F 已形成“加载真实黄金样本 -> 逐条调用大模型或显式本地兜底 -> 写入评测运行记录 -> 回填 Prompt 当前版本 -> 发布门禁读取评测结果 -> 页面展示执行方式和逐条样本数 -> 审计留痕”的完整开发闭环。
|
||||
- 剩余边界：
|
||||
- 本轮未部署到 `192.168.3.90`，按你的部署策略先完成开发闭环，后续统一部署时只需部署前后端；本轮没有新增数据库迁移。
|
||||
- 本地验证中使用了假大模型客户端证明“逐条调用”路径，不消耗真实模型额度；真实环境配置共享 AI 地址和密钥后即可切换为真实逐条大模型评测。
|
||||
|
||||
## 2026-05-21 Phase15-G 法规影响语义扫描与后台 worker 闭环
|
||||
|
||||
- 已继续按《AI合同起草与审查助手-法律文书能力吸收开发方案-2026-05-20.md》补齐 Phase15-G 的上线前增强项。
|
||||
- 本轮把原来的“关键词/结构化影响分析 + 手动 run-pending”升级为“语义向量扫描 + 独立后台 worker 定时消费”：
|
||||
- 新增 `LegalSemanticMatcher`，用本地法律概念词、中文 n-gram、关键词向量做余弦相似度匹配。
|
||||
- 语义扫描不依赖外部 embedding 服务、不需要联网，适合当前测试服务器资源约束。
|
||||
- 法规影响扫描现在同时记录 `direct_legal_reference`、`tag_match`、`keyword_match`、`semantic_vector_match`。
|
||||
- 影响项 `evidence_json.semantic_vector` 会保存相似度分数、共享词、共享法律概念。
|
||||
- 事件 `impact_summary_json` 新增 `semantic_match_count`、`by_match_reason`、`scan_strategy`，便于 QA 和页面验证。
|
||||
- 后台 worker：
|
||||
- 新增 `backend/app/tasks/legal_change_impact_worker.py`。
|
||||
- 支持 `run_once()` 消费 pending 队列，也支持 daemon 模式按间隔循环消费。
|
||||
- 新增 `scripts/run-legal-change-impact-worker.ps1`：
|
||||
- 单次执行：`.\scripts\run-legal-change-impact-worker.ps1 -Once -Limit 10`
|
||||
- 常驻执行：`.\scripts\run-legal-change-impact-worker.ps1 -Limit 10 -IntervalSeconds 300`
|
||||
- worker 使用 `system-legal-change-worker` 作为系统执行人，不需要真人手动点击队列消费。
|
||||
- `docker-compose.server.yml` 新增 `legal-change-impact-worker` 服务，正式部署时可作为独立容器常驻运行。
|
||||
- 前端闭环：
|
||||
- `/legal-change-impacts` 影响摘要新增“语义命中”数量。
|
||||
- 影响项命中原因可见展示“直接引用 / 标签 / 关键词 / 语义向量”。
|
||||
- 语义向量命中时展示相似度分数和共享法律概念。
|
||||
- 页面说明补充后台 worker 可定时消费待扫描队列。
|
||||
- 本轮验证：
|
||||
- `python -m compileall backend\app\services\legal_semantic_matcher.py backend\app\services\legal_change_impact_service.py backend\app\tasks\legal_change_impact_worker.py backend\app\api\legal_change_impacts.py`
|
||||
- `$env:PYTHONPATH='backend'; python -m unittest backend.tests.test_legal_change_impact_service backend.tests.test_legal_change_impacts_api`
|
||||
- `$env:PYTHONPATH='backend'; python -m unittest backend.tests.test_legal_article_service backend.tests.test_legal_articles_api backend.tests.test_legal_change_impact_service backend.tests.test_legal_change_impacts_api`
|
||||
- `cd frontend; npx eslint src/app/legal-change-impacts/page.tsx src/types/api.ts`
|
||||
- `cd frontend; npm run build`
|
||||
- 当前闭环结论：
|
||||
- Phase15-G 已形成“法条导入/变更入队 -> 后台 worker 定时消费 pending 事件 -> 直接引用/标签/关键词/语义向量扫描 -> 生成影响项和语义证据 -> 页面展示语义命中 -> 审计留痕”的完整开发闭环。
|
||||
- 剩余边界：
|
||||
- 本轮未部署到 `192.168.3.90`，没有新增数据库迁移；后续统一部署时可以直接启用 `docker-compose.server.yml` 中的 `legal-change-impact-worker`，或用 `scripts/run-legal-change-impact-worker.ps1` 挂计划任务。
|
||||
- 当前是本地轻量语义向量，不是远程大模型 embedding；后续若共享 AI 平台提供 embedding 服务，可在保持现有证据字段不变的情况下替换 `LegalSemanticMatcher` 的向量来源。
|
||||
|
||||
## 2026-05-21 法律文书能力吸收方案代码对齐检查
|
||||
|
||||
- 已按《AI合同起草与审查助手-法律文书能力吸收开发方案-2026-05-20.md》逐项核对 Phase15-A 到 Phase15-H 的代码现状。
|
||||
- 新增对齐报告：
|
||||
- `project-docs/AI合同起草与审查助手-法律文书能力吸收开发方案-代码对齐检查-2026-05-21.md`
|
||||
- 已同步 `project-docs/README.md`，把该对齐检查报告加入当前有效文档入口。
|
||||
- 本轮核对结论：
|
||||
- Phase15-A 到 Phase15-H 的主线开发已经在代码层闭环。
|
||||
- `/clauses/apply` 未单独实现，但条款推荐和改写结果已经通过 `/documents/[id]/edit` 编辑器写入合同正文并生成 `generated_contract_versions`，属于等价闭环。
|
||||
- `/quality-evals` 未单独实现，当前评测入口集中在 `/prompt-assets`，已能支撑 Prompt 发布前评测门禁；如要做统一质量中心，后续可单独补页面和通用评测表。
|
||||
- 法规影响分析代码统一使用 `/legal-change-impacts`、`legal_change_events`、`legal_change_impacts` 命名，后续文档应以此为准。
|
||||
- 本轮验证：
|
||||
- `$env:PYTHONPATH='backend'; python -m unittest backend.tests.test_legal_document_type_service backend.tests.test_legal_document_types_api backend.tests.test_legal_document_element_schema_service backend.tests.test_legal_document_element_schemas_api backend.tests.test_legal_document_draft_service backend.tests.test_legal_document_drafts_api backend.tests.test_contracts_api backend.tests.test_review_risk_action_service backend.tests.test_review_risk_actions_api backend.tests.test_clause_recommendation_service backend.tests.test_clauses_api backend.tests.test_prompt_asset_service backend.tests.test_prompt_assets_api backend.tests.test_prompt_runtime_service backend.tests.test_legal_change_impact_service backend.tests.test_legal_change_impacts_api backend.tests.test_operation_kpi_service backend.tests.test_operation_kpis_api`
|
||||
- 结果：`Ran 92 tests in 2.323s`，`OK`。
|
||||
- `cd frontend; npx eslint "src/app/documents/new/page.tsx" "src/app/documents/new/[type]/page.tsx" "src/app/documents/[id]/edit/page.tsx" "src/app/prompt-assets/page.tsx" "src/app/legal-change-impacts/page.tsx" "src/app/operation-kpis/page.tsx" "src/components/layout/app-topbar.tsx" "src/types/api.ts"`
|
||||
- 结果：通过。
|
||||
- `cd frontend; npm run build`
|
||||
- 结果：通过，Next 构建生成 39 个页面，包含 `/documents/new`、`/documents/new/[type]`、`/documents/[id]/edit`、`/prompt-assets`、`/legal-change-impacts`、`/operation-kpis`、`/sales-followups`。
|
||||
- 剩余边界：
|
||||
- 当前仍有 Phase15-C、Phase15-F、Phase15-G 相关未提交改动，需要后续统一整理提交。
|
||||
- 最新代码尚未部署到 `192.168.3.90` 做真实库烟测。
|
||||
- 专业富文本库、字段级 diff、独立 `/quality-evals`、全量法规向量化、真实微信或支付宝支付、电子签章和小程序均属于后续增强，不是当前法律文书能力吸收主线阻塞。
|
||||
|
||||
@@ -0,0 +1,236 @@
|
||||
# AI合同起草与审查助手-法律文书能力吸收开发方案代码对齐检查-2026-05-21
|
||||
|
||||
## 一、检查范围
|
||||
|
||||
本次只检查《AI合同起草与审查助手-法律文书能力吸收开发方案-2026-05-20.md》与当前代码是否对齐，不把“尚未提交”“尚未部署到测试服务器”和“功能没有开发”混为一谈。
|
||||
|
||||
检查范围包括：
|
||||
|
||||
- 后端模型、迁移、仓储、服务、API。
|
||||
- 前端页面、菜单入口、服务调用和类型定义。
|
||||
- 测试覆盖和本轮可复跑验证结果。
|
||||
- 文档中提出但代码采用等价实现的接口或页面。
|
||||
|
||||
## 二、总体结论
|
||||
|
||||
当前代码已经覆盖 Phase15-A 到 Phase15-H 的主线开发闭环。也就是说，文书类型入口、动态要素表单、草稿保存恢复、生成接口升级、审查建议采纳、条款推荐与行业惯例、富文本编辑器和 AI 侧边栏、Prompt 治理、黄金样本评测、法规影响扫描、运营 KPI 与商业化增强，均已在代码层形成可运行闭环。
|
||||
|
||||
但仍有三类事项需要分开看：
|
||||
|
||||
1. 已开发但待提交或待部署验证：当前工作区存在 Phase15-C、Phase15-F、Phase15-G 的未提交改动，且最新 worker 和语义扫描未在 `192.168.3.90` 做远程部署烟测。
|
||||
2. 文档建议路径与代码等价实现不同：例如文档建议 `/quality-evals` 和 `/clauses/apply`，当前分别落在 `/prompt-assets` 评测历史和合同编辑器版本链中。
|
||||
3. 后续增强项：专业富文本库、字段级差异对比、独立质量评测中心、全量法规向量化、真实微信或支付宝支付、电子签章、小程序等不属于当前 Phase15 主线硬阻塞。
|
||||
|
||||
## 三、逐阶段对齐表
|
||||
|
||||
| 阶段 | 文档目标 | 当前代码状态 | 对齐结论 |
|
||||
|---|---|---|---|
|
||||
| Phase15-A | 文书类型入口与动态要素表单 | 已有 `legal_document_types`、`legal_document_element_schemas`、`legal_document_drafts`，前端已有 `/documents/new` 和 `/documents/new/[type]` | 已闭环 |
|
||||
| LAW-DOC-001 | 文书类型字典、分类、行业、场景入口 | 后端有文书类型 API、默认种子和前端入口筛选 | 已闭环 |
|
||||
| LAW-DOC-002 | 动态要素 schema | 后端支持 schema 配置、发布、校验；前端按 schema 渲染表单 | 已闭环 |
|
||||
| LAW-DOC-003 | 要素草稿保存、恢复、删除、继续生成 | 后端有草稿 API，前端支持保存恢复，生成时带入 `elements_snapshot` | 已闭环 |
|
||||
| LAW-DOC-004 | 生成接口升级 | `POST /api/v1/contracts/generate` 支持 `document_type_code`、结构化要素、模板选择、待补字段、法条来源和 HTML 结果 | 已闭环 |
|
||||
| Phase15-B | 审查建议采纳闭环 | 已有 `review_risk_actions`、风险动作 API、应用建议生成新版本和审计日志 | 已闭环 |
|
||||
| LAW-REVIEW-001 | 风险建议动作模型 | 每条风险建议有 pending、adopted、modified、ignored 等状态和审计 | 已闭环 |
|
||||
| LAW-REVIEW-002 | 应用建议生成新版本 | `POST /reviews/{id}/risk-actions/{action_id}/apply` 会尝试定位正文并生成合同版本，无法定位时显式返回人工确认状态 | 已闭环 |
|
||||
| LAW-REVIEW-003 | 审查报告处理进度 | `/reviews/[id]` 已展示风险处理按钮、状态和进度 | 已闭环 |
|
||||
| Phase15-C | 智能条款推荐 | `POST /api/v1/clauses/recommend` 支持缺失条款和行业惯例，`POST /api/v1/clauses/rewrite` 支持措辞优化 | 已闭环 |
|
||||
| LAW-CLAUSE-001 | 缺失条款推荐 | 推荐结果含插入位置、建议文本、理由、依据和行业惯例说明 | 已闭环 |
|
||||
| LAW-CLAUSE-002 | 措辞优化 | 支持专业化、简化、偏甲方、偏乙方、均衡等改写模式，返回法律边界和谈判建议 | 已闭环 |
|
||||
| Phase15-D | 富文本编辑器与 AI 侧边栏 | `/documents/[id]/edit` 支持正文编辑、自动保存、版本历史、条款推荐、措辞改写、相关法条和 Chat 入口 | 主线已闭环 |
|
||||
| Phase15-E | Prompt 资产治理 | `/prompt-assets` 支持 Prompt 类型、版本、评测、发布、下线、回滚和运行证据 | 已闭环 |
|
||||
| Phase15-F | 黄金样本与质量评测 | 已有 Prompt 评测运行记录、逐条黄金样本报告、发布前门禁和前端评测历史 | 已闭环 |
|
||||
| Phase15-G | 法规变更影响分析 | 已有法条变更入队、结构化和语义影响扫描、影响项、页面、后台 worker 和 Docker 服务 | 已闭环 |
|
||||
| Phase15-H | 运营 KPI 与商业化增强 | `/operation-kpis` 已展示用量、文书生成、审查、Chat、模板、条款采纳、试用和转化指标；计费、发票状态、销售跟进已有业务闭环 | 主线已闭环 |
|
||||
|
||||
## 四、代码证据
|
||||
|
||||
### 4.1 后端 API
|
||||
|
||||
- 文书类型：`backend/app/api/legal_document_types.py`
|
||||
- 文书草稿：`backend/app/api/legal_document_drafts.py`
|
||||
- 合同生成和版本：`backend/app/api/contracts.py`
|
||||
- 审查建议动作：`backend/app/api/reviews.py`
|
||||
- 条款推荐和改写：`backend/app/api/clauses.py`
|
||||
- Prompt 治理：`backend/app/api/prompt_assets.py`
|
||||
- 法规影响分析：`backend/app/api/legal_change_impacts.py`
|
||||
- 运营 KPI：`backend/app/api/operation_kpis.py`
|
||||
- 销售跟进：`backend/app/api/sales_followups.py`
|
||||
|
||||
### 4.2 后端核心模型
|
||||
|
||||
- `LegalDocumentType`
|
||||
- `LegalDocumentElementSchema`
|
||||
- `LegalDocumentDraft`
|
||||
- `GeneratedContract`
|
||||
- `GeneratedContractVersion`
|
||||
- `ReviewRiskAction`
|
||||
- `PromptAsset`
|
||||
- `PromptAssetVersion`
|
||||
- `PromptAssetRunLink`
|
||||
- `PromptAssetEvaluationRun`
|
||||
- `LegalChangeEvent`
|
||||
- `LegalChangeImpact`
|
||||
- `ContractLifecycleTask`
|
||||
- `CommercialOrder`
|
||||
- `CommercialInvoice`
|
||||
- `SalesFollowupEvent`
|
||||
|
||||
### 4.3 前端页面
|
||||
|
||||
- 文书入口：`frontend/src/app/documents/new/page.tsx`
|
||||
- 动态要素表单：`frontend/src/app/documents/new/[type]/page.tsx`
|
||||
- 合同编辑器和 AI 侧边栏：`frontend/src/app/documents/[id]/edit/page.tsx`
|
||||
- 审查建议处理：`frontend/src/app/reviews/[id]/page.tsx`
|
||||
- Prompt 治理：`frontend/src/app/prompt-assets/page.tsx`
|
||||
- 法规影响分析：`frontend/src/app/legal-change-impacts/page.tsx`
|
||||
- 运营 KPI：`frontend/src/app/operation-kpis/page.tsx`
|
||||
- 销售跟进：`frontend/src/app/sales-followups/page.tsx`
|
||||
|
||||
### 4.4 后台任务
|
||||
|
||||
- 法规影响后台 worker：`backend/app/tasks/legal_change_impact_worker.py`
|
||||
- 本地执行脚本：`scripts/run-legal-change-impact-worker.ps1`
|
||||
- 测试服务器编排：`docker-compose.server.yml` 中已增加 `legal-change-impact-worker`
|
||||
|
||||
## 五、等价实现说明
|
||||
|
||||
### 5.1 `/clauses/apply`
|
||||
|
||||
文档建议新增 `POST /api/v1/clauses/apply`。当前代码没有单独做这个接口，而是通过合同编辑器完成应用：
|
||||
|
||||
1. 前端调用 `/clauses/recommend` 或 `/clauses/rewrite`。
|
||||
2. 用户在 `/documents/[id]/edit` 中选择插入或替换正文。
|
||||
3. 前端调用合同版本接口保存新正文。
|
||||
4. 后端生成新的 `generated_contract_versions` 记录。
|
||||
|
||||
这个实现已经满足“推荐或改写结果应用到合同正文并形成新版本”的验收口径。后续只有在需要第三方系统直接调用条款采纳动作时，才有必要再补独立 `/clauses/apply`。
|
||||
|
||||
### 5.2 `/quality-evals`
|
||||
|
||||
文档建议新增独立 `/quality-evals` 页面。当前代码没有这个独立路由，质量评测落在 `/prompt-assets` 中：
|
||||
|
||||
- Prompt 详情页可以运行评测。
|
||||
- 评测结果回填当前版本。
|
||||
- 发布按钮受评测状态约束。
|
||||
- 评测运行历史可以追溯。
|
||||
|
||||
这对“Prompt 发布前黄金样本评测门禁”已经闭环。但如果要做跨生成、审查、Chat、条款推荐、法规影响的统一质量中心，仍建议后续补独立 `/quality-evals`。
|
||||
|
||||
### 5.3 `regulation-changes`
|
||||
|
||||
文档用的是“法规变更影响分析”的产品名。当前代码实际路由统一命名为 `/legal-change-impacts`，表名为 `legal_change_events` 和 `legal_change_impacts`。命名不同，但职责一致，建议后续文档统一采用当前代码命名，减少误解。
|
||||
|
||||
### 5.4 `clause_recommendations` 持久化表
|
||||
|
||||
文档建议可以沉淀 `clause_recommendations`。当前代码采用实时推荐加审计日志，再通过合同版本链保存采纳结果。这个选择更轻，不会制造一张只保存临时建议的重复表。
|
||||
|
||||
如果未来要统计“每条推荐从展示到采纳、忽略、修改”的完整漏斗，可以再补持久化推荐动作表。
|
||||
|
||||
## 六、仍未彻底闭环的事项
|
||||
|
||||
### 6.1 待提交和待部署验证
|
||||
|
||||
当前工作区存在未提交改动，主要集中在：
|
||||
|
||||
- Phase15-C 行业惯例推荐。
|
||||
- Phase15-F 真实黄金样本逐条评测。
|
||||
- Phase15-G 法规影响语义扫描和后台 worker。
|
||||
|
||||
这些代码本地测试已通过，但尚未提交远程 Git，也未在 `192.168.3.90` 重新部署后做真实库烟测。
|
||||
|
||||
### 6.2 独立质量评测中心
|
||||
|
||||
当前评测能力已经在 Prompt 治理内闭环，但没有独立 `/quality-evals` 页面，也没有通用 `ai_quality_eval_runs/items` 表承接所有质量评测类型。
|
||||
|
||||
这不是 Prompt 发布门禁阻塞项，但属于文档和产品形态的一个剩余增强点。
|
||||
|
||||
### 6.3 专业级富文本增强
|
||||
|
||||
当前 `/documents/[id]/edit` 已经支持编辑、自动保存、版本历史、推荐插入、改写替换和 AI 侧边栏。尚未做的增强包括：
|
||||
|
||||
- 专业富文本编辑器库。
|
||||
- 段落级或字段级差异对比。
|
||||
- Word 样式还原和导出增强。
|
||||
- 多人协同编辑。
|
||||
|
||||
这些是体验增强，不影响当前 Phase15-D 主线闭环。
|
||||
|
||||
### 6.4 全量法规向量化
|
||||
|
||||
当前法规影响语义扫描使用本地轻量语义匹配，不等同于把全量法规库一次性上传到 shared-ai-platform 做向量检索。
|
||||
|
||||
这不影响现有法条全文检索、审查报告法条引用和法规影响扫描的业务闭环，但如果目标升级为“全国全量法规语义检索”，需要另开批处理和容量治理阶段。
|
||||
|
||||
### 6.5 真实在线支付和完整发票流程
|
||||
|
||||
Phase15-H 的运营 KPI 和商业化基础已经完成，但真实微信或支付宝支付、商户回调验签、公网 HTTPS 回调域名、完整发票抬头税号和开票文件，仍依赖外部商户资料和公网条件。
|
||||
|
||||
这属于商业化生产增强，不是《法律文书能力吸收开发方案》的核心法律文书主线阻塞。
|
||||
|
||||
### 6.6 P3 外部集成
|
||||
|
||||
微信小程序、电子签章、第三方签署平台、重型工作流引擎等在原方案中明确为延后或不照搬方向。当前未实现是符合方案边界的，不应算作跑偏。
|
||||
|
||||
## 七、本轮验证结果
|
||||
|
||||
### 7.1 后端回归
|
||||
|
||||
命令：
|
||||
|
||||
```powershell
|
||||
$env:PYTHONPATH='backend'; python -m unittest backend.tests.test_legal_document_type_service backend.tests.test_legal_document_types_api backend.tests.test_legal_document_element_schema_service backend.tests.test_legal_document_element_schemas_api backend.tests.test_legal_document_draft_service backend.tests.test_legal_document_drafts_api backend.tests.test_contracts_api backend.tests.test_review_risk_action_service backend.tests.test_review_risk_actions_api backend.tests.test_clause_recommendation_service backend.tests.test_clauses_api backend.tests.test_prompt_asset_service backend.tests.test_prompt_assets_api backend.tests.test_prompt_runtime_service backend.tests.test_legal_change_impact_service backend.tests.test_legal_change_impacts_api backend.tests.test_operation_kpi_service backend.tests.test_operation_kpis_api
|
||||
```
|
||||
|
||||
结果：
|
||||
|
||||
```text
|
||||
Ran 92 tests in 2.323s
|
||||
OK
|
||||
```
|
||||
|
||||
### 7.2 前端重点页面静态检查
|
||||
|
||||
命令：
|
||||
|
||||
```powershell
|
||||
npx eslint "src/app/documents/new/page.tsx" "src/app/documents/new/[type]/page.tsx" "src/app/documents/[id]/edit/page.tsx" "src/app/prompt-assets/page.tsx" "src/app/legal-change-impacts/page.tsx" "src/app/operation-kpis/page.tsx" "src/components/layout/app-topbar.tsx" "src/types/api.ts"
|
||||
```
|
||||
|
||||
结果：通过，无错误输出。
|
||||
|
||||
### 7.3 前端生产构建
|
||||
|
||||
命令：
|
||||
|
||||
```powershell
|
||||
npm run build
|
||||
```
|
||||
|
||||
结果：
|
||||
|
||||
```text
|
||||
Compiled successfully
|
||||
Generating static pages (39/39)
|
||||
```
|
||||
|
||||
构建路由中已包含：
|
||||
|
||||
- `/documents/new`
|
||||
- `/documents/new/[type]`
|
||||
- `/documents/[id]/edit`
|
||||
- `/prompt-assets`
|
||||
- `/legal-change-impacts`
|
||||
- `/operation-kpis`
|
||||
- `/sales-followups`
|
||||
|
||||
## 八、建议下一步
|
||||
|
||||
为了让文档和代码彻底闭合，建议下一步不要继续堆新功能，先按这个顺序收口：
|
||||
|
||||
1. 整理并提交当前 Phase15-C、Phase15-F、Phase15-G 的代码和文档改动。
|
||||
2. 如果本阶段要求“文档路径和产品入口完全一致”，补独立 `/quality-evals` 质量评测中心，把 `/prompt-assets` 中的评测历史聚合过去。
|
||||
3. 统一文档命名，把法规影响分析的接口名称固定为 `/legal-change-impacts`，避免后续误认为还缺 `/regulation-changes`。
|
||||
4. 等决定部署时，在 `192.168.3.90` 执行最新 Alembic，并启用 `legal-change-impact-worker` 服务做真实库烟测。
|
||||
|
||||
按当前代码状态判断，《法律文书能力吸收开发方案》的主线开发没有跑偏，已经基本完成。真正剩下的是提交、部署验证、独立质量中心和若干生产级增强。
|
||||
@@ -21,6 +21,7 @@
|
||||
- `AI合同起草与审查助手-正式生产部署演练与上线检查清单-2026-05-16.md`
|
||||
- `AI合同起草与审查助手-Phase14法条全文级知识库实施方案-2026-05-16.md`
|
||||
- `AI合同起草与审查助手-法律文书能力吸收开发方案-2026-05-20.md`
|
||||
- `AI合同起草与审查助手-法律文书能力吸收开发方案-代码对齐检查-2026-05-21.md`
|
||||
|
||||
## 当前仍保留的专项资料
|
||||
|
||||
|
||||
@@ -0,0 +1,29 @@
|
||||
param(
|
||||
[switch]$Once,
|
||||
[int]$Limit = 10,
|
||||
[int]$IntervalSeconds = 300,
|
||||
[int]$MaxLoops = 0
|
||||
)
|
||||
|
||||
$ErrorActionPreference = "Stop"
|
||||
|
||||
$repoRoot = Split-Path -Parent $PSScriptRoot
|
||||
$backendPath = Join-Path $repoRoot "backend"
|
||||
$env:PYTHONPATH = $backendPath
|
||||
|
||||
$argsList = @("-m", "app.tasks.legal_change_impact_worker", "--limit", "$Limit")
|
||||
if ($Once) {
|
||||
$argsList += "--once"
|
||||
} else {
|
||||
$argsList += @("--interval-seconds", "$IntervalSeconds")
|
||||
}
|
||||
if ($MaxLoops -gt 0) {
|
||||
$argsList += @("--max-loops", "$MaxLoops")
|
||||
}
|
||||
|
||||
Push-Location $repoRoot
|
||||
try {
|
||||
python @argsList
|
||||
} finally {
|
||||
Pop-Location
|
||||
}
|
||||
Reference in New Issue
Block a user