"""Extract configured structured fields from one completed conversation.""" from __future__ import annotations import json from datetime import UTC, datetime from typing import Any import httpx from db.models import ConversationMessage, ConversationSession, ModelResource from db.session import SessionLocal from sqlalchemy import select ANALYSIS_TIMEOUT_SECONDS = 60.0 MAX_TRANSCRIPT_CHARS = 120_000 def _endpoint(base_url: str, path: str) -> str: return f"{base_url.rstrip('/')}/{path.lstrip('/')}" def _json_schema(fields: list[dict[str, Any]]) -> dict[str, Any]: properties: dict[str, Any] = {} for field in fields: field_type = str(field.get("type") or "string") schema: dict[str, Any] = { "description": str(field.get("description") or ""), } if field_type == "enum": schema.update( { "type": ["string", "null"], "enum": [*(field.get("enum_values") or []), None], } ) else: schema["type"] = [field_type, "null"] properties[str(field["name"])] = schema return { "type": "object", "properties": properties, "required": list(properties), "additionalProperties": False, } def _transcript(messages: list[ConversationMessage]) -> str: lines = [ f"[{message.role}] {message.content.strip()}" for message in messages if message.content_type == "text" and message.content.strip() ] transcript = "\n".join(lines) if len(transcript) <= MAX_TRANSCRIPT_CHARS: return transcript return "[较早内容已截断]\n" + transcript[-MAX_TRANSCRIPT_CHARS:] def _parse_json_content(content: object) -> dict[str, Any]: text = str(content or "").strip() if text.startswith("```"): lines = text.splitlines() if lines and lines[0].startswith("```"): lines = lines[1:] if lines and lines[-1].strip() == "```": lines = lines[:-1] text = "\n".join(lines).strip() parsed = json.loads(text) if not isinstance(parsed, dict): raise ValueError("分析模型必须返回 JSON 对象") return parsed def _validated_result( raw: dict[str, Any], fields: list[dict[str, Any]] ) -> dict[str, Any]: result: dict[str, Any] = {} for field in fields: name = str(field["name"]) field_type = str(field.get("type") or "string") value = raw.get(name) valid = value is None if field_type == "string": valid = valid or isinstance(value, str) elif field_type == "boolean": valid = valid or isinstance(value, bool) elif field_type == "integer": valid = valid or (isinstance(value, int) and not isinstance(value, bool)) elif field_type == "number": valid = valid or ( isinstance(value, (int, float)) and not isinstance(value, bool) ) elif field_type == "enum": valid = valid or ( isinstance(value, str) and value in list(field.get("enum_values") or []) ) result[name] = value if valid else None return result async def _request_analysis( resource: ModelResource, fields: list[dict[str, Any]], transcript: str, ) -> dict[str, Any]: values = resource.values or {} secrets = resource.secrets or {} api_url = str(values.get("apiUrl") or "") api_key = str(secrets.get("apiKey") or "") model_id = str(values.get("modelId") or "") if resource.interface_type != "openai-llm": raise ValueError(f"分析暂不支持模型接口:{resource.interface_type}") if not api_url or not api_key or not model_id: raise ValueError("分析模型资源缺少 apiUrl、apiKey 或 modelId") schema = _json_schema(fields) system_prompt = ( "你是通话关键信息提取器。只能使用对话中明确出现的信息，禁止猜测、" "补全或编造。无法确定的字段必须返回 null。严格按照给定 JSON Schema " "返回一个 JSON 对象，不要输出解释或 Markdown。\n\nJSON Schema:\n" + json.dumps(schema, ensure_ascii=False) ) async with httpx.AsyncClient(timeout=ANALYSIS_TIMEOUT_SECONDS) as client: response = await client.post( _endpoint(api_url, "chat/completions"), headers={"Authorization": f"Bearer {api_key}"}, json={ "model": model_id, "messages": [ {"role": "system", "content": system_prompt}, {"role": "user", "content": transcript}, ], "temperature": 0, "stream": False, "response_format": {"type": "json_object"}, }, ) response.raise_for_status() payload = response.json() choices = payload.get("choices") if isinstance(payload, dict) else None if not isinstance(choices, list) or not choices: raise ValueError("分析模型没有返回 choices") message = choices[0].get("message") if isinstance(choices[0], dict) else None content = message.get("content") if isinstance(message, dict) else None return _validated_result(_parse_json_content(content), fields) async def analyze_conversation(conversation_id: str) -> None: """Analyze one claimed conversation and persist its terminal state.""" try: async with SessionLocal() as session: conversation = await session.get(ConversationSession, conversation_id) if not conversation or conversation.analysis_status != "processing": return data = dict(conversation.analysis_data or {}) plan = data.get("plan") if isinstance(data.get("plan"), dict) else {} fields = plan.get("fields") if isinstance(plan.get("fields"), list) else [] resource_id = str(plan.get("model_resource_id") or "") resource = await session.get(ModelResource, resource_id) if not resource or not resource.enabled or resource.capability != "LLM": raise ValueError("分析模型不存在、未启用或不是 LLM 资源") messages = ( await session.execute( select(ConversationMessage) .where(ConversationMessage.session_id == conversation_id) .order_by(ConversationMessage.sequence) ) ).scalars().all() transcript = _transcript(list(messages)) if not transcript: raise ValueError("会话没有可分析的文本转写") result = await _request_analysis(resource, list(fields), transcript) async with SessionLocal() as session: conversation = await session.get(ConversationSession, conversation_id) if not conversation: return data = dict(conversation.analysis_data or {}) data["result"] = result data["completedAt"] = datetime.now(UTC).isoformat() conversation.analysis_data = data conversation.analysis_status = "completed" conversation.analysis_error = "" await session.commit() except Exception as exc: async with SessionLocal() as session: conversation = await session.get(ConversationSession, conversation_id) if conversation: conversation.analysis_status = "failed" conversation.analysis_error = str(exc)[:2048] await session.commit()