import json from datetime import datetime from decimal import Decimal from sqlalchemy.orm import Session from app.agents.orchestrator import MedicalConsultationOrchestrator from app.core.context import UserContext from app.core.exceptions import AppError from app.models.training_record import TrainingRecord, TrainingScoreDetail from app.repositories.case_repository import CaseRepository from app.repositories.evaluation_repository import EvaluationRepository from app.repositories.session_repository import SessionRepository from app.repositories.source_case_repository import SourceCaseRepository from app.schemas.evaluation import ( CreateEvaluationRequest, DimensionScore, EvaluationDetailResponse, EvaluationListItem, EvaluationListResponse, EvaluationResponse, ScoreDetailItem, ) from app.services.audit_service import AuditService from app.services.knowledge_service import KnowledgeService from app.services.runtime_memory import runtime_memory class EvaluationService: """评价服务：基于病例、评分规则和作答过程生成 training_record 与评分明细。""" def __init__(self, db: Session) -> None: self.db = db self.session_repo = SessionRepository(db) self.case_repo = CaseRepository(db) self.eval_repo = EvaluationRepository(db) self.source_repo = SourceCaseRepository(db) self.knowledge = KnowledgeService(db) self.audit = AuditService(db) self.orchestrator = MedicalConsultationOrchestrator() async def create_evaluation(self, ctx: UserContext, session_id: int, payload: CreateEvaluationRequest) -> EvaluationResponse: """评价生成：读取会话短期 memory、提交内容、评分规则和指南后写入 training_record。""" session = self.session_repo.get_owned_session(session_id, ctx.user_id) if not session: raise AppError("SESSION_NOT_FOUND", "session not found or not owned by current user", 404) if session.status not in {"evaluating", "completed"}: raise AppError("SESSION_STATUS_INVALID", "evaluation requires treatment submission", 400) existed = self.eval_repo.get_by_session(session.id, ctx.user_id) if existed: return self._to_response(existed) case = self.case_repo.get_active_case(session.case_id) if not case: raise AppError("CASE_NOT_FOUND", "case not found or inactive", 404) submission = self.session_repo.get_submission(session.id) if not submission or not submission.treatment_submitted_at: raise AppError("TREATMENT_REQUIRED", "treatment submission is required", 400) session.score_type = payload.score_type memory_messages = runtime_memory.get_messages(session.memory_key) keyword_seed = (case.key_symptoms or []) + (case.key_exams or []) + [case.diagnosis_primary or ""] guideline_result = self.knowledge.search_guidelines(case.department_id, session.training_type, keyword_seed) guideline_refs = guideline_result["source_refs"] scoring_rules = self.source_repo.get_scoring_rules(case.id) report = await self.orchestrator.evaluate( session=session, case=case, memory_messages=memory_messages, orders=session.orders, submission=submission, rubric=None, guideline_refs=guideline_refs, scoring_rules=scoring_rules, ) record = self._build_training_record(ctx, session, case, submission, report, scoring_rules, guideline_result) self.eval_repo.create_record(record) self.eval_repo.replace_score_details(record.id, self._build_score_details(record.id, report, scoring_rules)) self.session_repo.update_status(session, "completed") runtime_memory.release(session.memory_key) self.audit.log(ctx, "evaluation.generate", "training_record", str(record.id), session.id) return self._to_response(record) def _build_training_record( self, ctx: UserContext, session, case, submission, report: dict, scoring_rules: list, guideline_result: dict, ) -> TrainingRecord: """训练记录写入：完整流程结束后把评分结果沉淀到 training_record。""" end_time = datetime.utcnow() start_time = session.started_at or session.created_at or end_time duration_seconds = int((end_time - start_time).total_seconds()) if start_time else None total_score = float(report.get("total_score") or 0) structured = { "score_type": report.get("score_type", session.score_type), "total_score": total_score, "dimension_scores": report.get("dimension_scores") or [], "score_details": report.get("score_details") or [], "errors": report.get("errors") or [], "improvement_plan": report.get("improvement_plan") or [], "evidence_summary": report.get("evidence_summary") or [], "guideline_refs": report.get("guideline_refs") or [], "overall_comment": report.get("overall_comment") or "", "llm_model": report.get("_llm_model"), "latency_metrics": report.get("_latency_metrics") or {}, } return TrainingRecord( training_mode=session.mode, case_type=session.training_type, start_time=start_time, end_time=end_time, duration_seconds=duration_seconds, total_score=total_score, ai_score=total_score, teacher_score=None, evaluation_level=self._evaluation_level(total_score, report.get("score_type", session.score_type)), status="completed", feedback=structured["overall_comment"], thinking_chain=json.dumps( { "evidence_summary": structured["evidence_summary"], "guideline_refs": structured["guideline_refs"], "scoring_rule_count": len(scoring_rules), }, ensure_ascii=False, ), diagnosis_path=json.dumps( { "primary_diagnosis": submission.primary_diagnosis, "differential_diagnoses": submission.differential_diagnoses or [], "diagnosis_basis": submission.diagnosis_basis, "standard_diagnosis": case.diagnosis_primary, }, ensure_ascii=False, ), wrong_points=structured["errors"], missed_questions=[], recommendation_result={"improvement_plan": structured["improvement_plan"]}, ai_feedback_structured=structured, osce_station_score={}, interruption_count=0, emotion_analysis={}, prompt_version="v1", rag_context_version=self._rag_context_version(guideline_result), case_id=case.id, teacher_id=None, user_id=self._numeric_user_id(ctx.user_id), external_user_id=ctx.user_id, session_id=session.id, evaluation_record_id=None, score_type=structured["score_type"], pdf_file_path=None, ) def _build_score_details(self, record_id: int, report: dict, scoring_rules: list) -> list[TrainingScoreDetail]: """评分明细写入：把 LLM 结构化评分结果映射到 training_score_detail。""" raw_items = report.get("score_details") or report.get("dimension_scores") or [] rule_map = self._rule_map(scoring_rules) details: list[TrainingScoreDetail] = [] for item in raw_items: if not isinstance(item, dict): continue dimension = str(item.get("dimension") or "综合表现") matched_rule = self._match_rule(item, dimension, rule_map) deducted_reason = item.get("deducted_reason") if not deducted_reason: deducted_reason = "；".join(str(value) for value in (item.get("deductions") or []) if value) evidence = item.get("evidence_message_ids") if evidence is None: evidence = item.get("evidence") or [] details.append( TrainingScoreDetail( record_id=record_id, rule_id=int(item.get("rule_id") or matched_rule.id) if matched_rule else None, dimension=dimension[:50], score=self._decimal_or_none(item.get("score")), deducted_reason=deducted_reason or "", evidence_message_ids=evidence if isinstance(evidence, list) else [evidence], ai_confidence=self._decimal_or_none(item.get("ai_confidence") or 0.85), comment=item.get("comment") or item.get("improvement") or "", ) ) return details def _rule_map(self, scoring_rules: list) -> dict[str, object]: """评分规则映射：按维度和能力维度建立匹配索引。""" result = {} for rule in scoring_rules: for key in (getattr(rule, "dimension", ""), getattr(rule, "competency_dimension", "")): if key: result[str(key).strip()] = rule return result def _match_rule(self, item: dict, dimension: str, rule_map: dict[str, object]): """评分规则匹配：优先按 rule_id，其次按维度文本匹配 scoring_rule。""" rule_id = item.get("rule_id") if rule_id: for rule in rule_map.values(): if getattr(rule, "id", None) == rule_id: return rule return rule_map.get(dimension) or rule_map.get(str(item.get("competency_dimension") or "").strip()) def _decimal_or_none(self, value: object) -> Decimal | None: """分数转换：将 LLM 返回值转换为 Decimal，异常时置空。""" try: return Decimal(str(value)) except Exception: return None def _evaluation_level(self, score: float, score_type: str) -> str: """评价等级：根据百分制或五分制总分生成训练记录等级。""" normalized = score * 20 if score_type == "five_point" else score if normalized >= 90: return "excellent" if normalized >= 80: return "good" if normalized >= 60: return "pass" return "needs_improvement" def _rag_context_version(self, guideline_result: dict) -> str: """RAG 版本：记录评分时是否命中指南片段。""" matched = guideline_result.get("matched_chunks") or [] return f"knowledge_chunks:{len(matched)}" if matched else "none" def _numeric_user_id(self, user_id: str) -> int | None: """用户 ID 兼容：Django 返回的 id 写入 external_user_id，纯数字时同步写入源库 user_id。""" return int(user_id) if str(user_id).isdigit() else None def list_history(self, user_id: str) -> EvaluationListResponse: """历史评价：按 Django 用户中心 ID 查询完整训练后的 training_record。""" records = self.eval_repo.list_by_user(user_id) return EvaluationListResponse( items=[ EvaluationListItem( evaluation_id=record.id, case_title=self._case_title(record.case_id), score_type=record.score_type, total_score=float(record.total_score or 0), created_at=record.created_at, pdf_exported=bool(record.pdf_file_path), ) for record in records ] ) def get_detail(self, evaluation_id: int, user_id: str) -> EvaluationDetailResponse: """评价详情：按 Django 用户中心 ID 校验归属并返回完整报告。""" record = self.eval_repo.get_owned_record(evaluation_id, user_id) if not record: raise AppError("EVALUATION_NOT_FOUND", "evaluation not found or not owned by current user", 404) base = self._to_response(record) return EvaluationDetailResponse( **base.model_dump(), session_id=record.session_id or 0, case_id=record.case_id, case_title=self._case_title(record.case_id), created_at=record.created_at, pdf_file_path=record.pdf_file_path, ) def _to_response(self, record: TrainingRecord) -> EvaluationResponse: """评价转换：把 training_record 转换为接口响应结构。""" structured = record.ai_feedback_structured or {} dimension_scores = structured.get("dimension_scores") or [] return EvaluationResponse( evaluation_id=record.id, score_type=record.score_type, total_score=float(record.total_score or structured.get("total_score") or 0), dimension_scores=[DimensionScore(**item) for item in dimension_scores], score_details=self._score_detail_response(record), errors=structured.get("errors") or record.wrong_points or [], improvement_plan=structured.get("improvement_plan") or (record.recommendation_result or {}).get("improvement_plan") or [], evidence_summary=structured.get("evidence_summary") or [], guideline_refs=structured.get("guideline_refs") or [], overall_comment=structured.get("overall_comment") or record.feedback or "", ) def _score_detail_response(self, record: TrainingRecord) -> list[ScoreDetailItem]: """评分明细响应：优先读取 training_score_detail，旧记录回退到结构化维度评分。""" details = self.eval_repo.list_score_details(record.id) if details: return [ ScoreDetailItem( id=item.id, record_id=item.record_id, rule_id=item.rule_id, dimension=item.dimension, score=float(item.score) if item.score is not None else None, deducted_reason=item.deducted_reason, evidence_message_ids=item.evidence_message_ids or [], ai_confidence=float(item.ai_confidence) if item.ai_confidence is not None else None, comment=item.comment, ) for item in details ] structured = record.ai_feedback_structured or {} return [ ScoreDetailItem( record_id=record.id, dimension=item.get("dimension", "综合表现"), score=float(item.get("score") or 0), deducted_reason="；".join(str(value) for value in item.get("deductions", []) if value), evidence_message_ids=item.get("evidence") or [], ai_confidence=None, comment=item.get("comment") or "", ) for item in structured.get("dimension_scores") or [] if isinstance(item, dict) ] def _case_title(self, case_id: int | None) -> str: """病例标题：历史记录只保存 case_id，展示时按新病例主表读取标题。""" if not case_id: return "" case = self.case_repo.get_active_case(case_id) return case.title if case else ""