e124e4bbcb
- 重构报销状态注册表、审批流路由与平台风险标记 - 完善管家意图规划器与模型计划构建器全链路 - 新增 OCR Worker 脚本、数据库会话管理与通知状态 - 优化文档中心、日志视图、预算中心与员工管理交互 - 增强工作台摘要、图标资源与全局主题样式 - 补充审批路由、状态注册、OCR 服务与管家规划器测试覆盖
302 lines
15 KiB
Python
302 lines
15 KiB
Python
from __future__ import annotations
|
||
|
||
import json
|
||
from dataclasses import dataclass
|
||
from typing import Any
|
||
|
||
from app.schemas.steward import (
|
||
StewardSlotDecisionRequest,
|
||
StewardSlotDecisionResponse,
|
||
StewardSlotOption,
|
||
)
|
||
from app.services.ontology_field_registry import normalize_ontology_form_values
|
||
from app.services.runtime_chat import RuntimeChatService
|
||
from app.services.steward_constants import BUSINESS_CANONICAL_FIELD_ORDER, BUSINESS_CANONICAL_FIELDS
|
||
|
||
|
||
STEWARD_SLOT_DECISION_FUNCTION_NAME = "submit_steward_slot_decision"
|
||
|
||
|
||
FIELD_CATALOG: dict[str, dict[str, str]] = {
|
||
"expense_type": {"label": "费用类型", "description": "申请或报销所属费用场景，如差旅、交通、住宿、业务招待。"},
|
||
"time_range": {"label": "时间", "description": "申请时为出差起止日期，报销时为费用发生日期。"},
|
||
"location": {"label": "地点", "description": "出差目的地、费用发生地或业务活动地点。"},
|
||
"reason": {"label": "事由", "description": "出差、报销或业务活动的业务原因。"},
|
||
"amount": {"label": "金额", "description": "报销时为实际金额；申请时金额可由系统估算，不应默认要求用户填写。"},
|
||
"transport_mode": {"label": "出行方式", "description": "差旅申请交通费用测算所需字段，由用户明确选择或表达。"},
|
||
"attachments": {"label": "附件/凭证", "description": "发票、行程单、付款截图或其他证明材料。"},
|
||
"customer_name": {"label": "客户或项目对象", "description": "业务招待、客户拜访或项目支撑涉及的对象。"},
|
||
"merchant_name": {"label": "商户/开票方", "description": "报销票据上的商户或开票方。"},
|
||
"department_name": {"label": "所属部门", "description": "申请人或费用归属部门。"},
|
||
"employee_name": {"label": "申请人", "description": "发起申请或报销的员工。"},
|
||
"employee_no": {"label": "员工编号", "description": "公司内部员工编号。"},
|
||
}
|
||
|
||
APPLICATION_NON_BLOCKING_FIELDS = {"amount", "attachments", "employee_no", "department_name", "employee_name"}
|
||
|
||
|
||
@dataclass(frozen=True, slots=True)
|
||
class StewardSlotDecisionAgentResult:
|
||
payload: dict[str, Any]
|
||
model_call_traces: list[dict[str, Any]]
|
||
|
||
|
||
class StewardSlotDecisionAgent:
|
||
"""用大模型 function calling 判断当前任务缺什么，以及下一步是否应先追问。"""
|
||
|
||
def __init__(self, runtime_chat_service: RuntimeChatService) -> None:
|
||
self.runtime_chat_service = runtime_chat_service
|
||
|
||
def decide(self, request: StewardSlotDecisionRequest) -> StewardSlotDecisionResponse:
|
||
normalized_request = self._normalize_request(request)
|
||
result = self.runtime_chat_service.complete_with_tool_call(
|
||
self._build_messages(normalized_request),
|
||
tools=[self._build_tool_schema()],
|
||
tool_choice={
|
||
"type": "function",
|
||
"function": {"name": STEWARD_SLOT_DECISION_FUNCTION_NAME},
|
||
},
|
||
max_tokens=1200,
|
||
temperature=0.05,
|
||
timeout_seconds=30,
|
||
max_attempts=1,
|
||
)
|
||
if result.tool_call is not None and result.tool_call.name == STEWARD_SLOT_DECISION_FUNCTION_NAME:
|
||
response = self._build_response_from_model_payload(
|
||
result.tool_call.arguments,
|
||
normalized_request,
|
||
result.calls_as_dicts(),
|
||
)
|
||
if response is not None:
|
||
return response
|
||
return self._build_rule_fallback(normalized_request, result.calls_as_dicts())
|
||
|
||
@staticmethod
|
||
def _normalize_request(request: StewardSlotDecisionRequest) -> StewardSlotDecisionRequest:
|
||
normalized_fields = {
|
||
key: value
|
||
for key, value in normalize_ontology_form_values(request.ontology_fields).items()
|
||
if key in BUSINESS_CANONICAL_FIELDS and str(value or "").strip()
|
||
}
|
||
missing_fields: list[str] = []
|
||
for item in request.missing_fields:
|
||
key = str(item or "").strip()
|
||
if request.task_type == "expense_application" and key in APPLICATION_NON_BLOCKING_FIELDS:
|
||
continue
|
||
if key in BUSINESS_CANONICAL_FIELDS and key not in missing_fields and not normalized_fields.get(key):
|
||
missing_fields.append(key)
|
||
return StewardSlotDecisionRequest(
|
||
task_type=request.task_type,
|
||
user_message=str(request.user_message or "").strip(),
|
||
ontology_fields=normalized_fields,
|
||
missing_fields=missing_fields,
|
||
task_context=request.task_context if isinstance(request.task_context, dict) else {},
|
||
)
|
||
|
||
@staticmethod
|
||
def _build_messages(request: StewardSlotDecisionRequest) -> list[dict[str, Any]]:
|
||
context_payload = {
|
||
"task_type": request.task_type,
|
||
"user_message": request.user_message,
|
||
"ontology_fields": request.ontology_fields,
|
||
"missing_fields_from_intent_agent": request.missing_fields,
|
||
"field_catalog": {
|
||
key: FIELD_CATALOG[key]
|
||
for key in BUSINESS_CANONICAL_FIELD_ORDER
|
||
if key in FIELD_CATALOG
|
||
},
|
||
"task_context": request.task_context,
|
||
}
|
||
return [
|
||
{
|
||
"role": "system",
|
||
"content": (
|
||
"你是 X-Financial 小财管家的任务字段决策智能体。"
|
||
"你必须通过 function calling 返回下一步动作。"
|
||
"你的任务不是关键词匹配，而是结合用户意图、当前任务类型、canonical ontology 字段、"
|
||
"上游意图识别给出的缺失字段和字段目录，判断现在应先追问用户，还是可以展示核对结果。"
|
||
"所有 required_fields 和 missing_fields 只能使用 field_catalog 中的 canonical 字段。"
|
||
"如果字段是内部提示、示例、系统指令或可选项，不能当作用户已经提供。"
|
||
"费用申请场景中 amount 可由系统估算，不应作为用户必须手填字段。"
|
||
"费用申请生成核对表阶段，attachments 不阻塞生成，可在报销或归档阶段补充；"
|
||
"employee_no、department_name、employee_name 属于系统用户档案字段，必须从上下文读取，不能向用户追问。"
|
||
"差旅申请通常只有 transport_mode 这类会影响费用测算的字段才需要先追问。"
|
||
"如果缺失字段会影响后续测算、入库、附件归集或合规判断，应返回 ask_user；"
|
||
"如果信息足以生成可核对但未提交的结果，应返回 render_preview。"
|
||
"question 和 rationale 必须是面向用户的业务说明，不暴露内部推理链。"
|
||
),
|
||
},
|
||
{
|
||
"role": "user",
|
||
"content": json.dumps(context_payload, ensure_ascii=False),
|
||
},
|
||
]
|
||
|
||
@staticmethod
|
||
def _build_tool_schema() -> dict[str, Any]:
|
||
canonical_fields = list(BUSINESS_CANONICAL_FIELD_ORDER)
|
||
return {
|
||
"type": "function",
|
||
"function": {
|
||
"name": STEWARD_SLOT_DECISION_FUNCTION_NAME,
|
||
"description": "提交小财管家当前任务的字段缺口和下一步动作决策。",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"next_action": {
|
||
"type": "string",
|
||
"enum": ["ask_user", "render_preview"],
|
||
},
|
||
"required_fields": {
|
||
"type": "array",
|
||
"items": {"type": "string", "enum": canonical_fields},
|
||
},
|
||
"missing_fields": {
|
||
"type": "array",
|
||
"items": {"type": "string", "enum": canonical_fields},
|
||
},
|
||
"question": {"type": "string"},
|
||
"options": {
|
||
"type": "array",
|
||
"items": {
|
||
"type": "object",
|
||
"properties": {
|
||
"label": {"type": "string"},
|
||
"value": {"type": "string"},
|
||
"field_key": {"type": "string", "enum": canonical_fields},
|
||
"description": {"type": "string"},
|
||
},
|
||
"required": ["label", "value", "field_key"],
|
||
},
|
||
},
|
||
"rationale": {"type": "string"},
|
||
},
|
||
"required": ["next_action", "required_fields", "missing_fields", "question", "options", "rationale"],
|
||
},
|
||
},
|
||
}
|
||
|
||
def _build_response_from_model_payload(
|
||
self,
|
||
payload: dict[str, Any],
|
||
request: StewardSlotDecisionRequest,
|
||
traces: list[dict[str, Any]],
|
||
) -> StewardSlotDecisionResponse | None:
|
||
next_action = str(payload.get("next_action") or "").strip()
|
||
if next_action not in {"ask_user", "render_preview"}:
|
||
return None
|
||
required_fields = self._sanitize_fields(payload.get("required_fields"))
|
||
missing_fields = self._sanitize_fields(payload.get("missing_fields"))
|
||
required_fields = self._filter_blocking_fields(required_fields, request.task_type)
|
||
missing_fields = self._filter_blocking_fields(missing_fields, request.task_type)
|
||
missing_fields = [
|
||
key
|
||
for key in missing_fields
|
||
if key in required_fields or key in request.missing_fields
|
||
]
|
||
if next_action == "ask_user" and not missing_fields:
|
||
missing_fields = list(request.missing_fields)
|
||
if next_action == "ask_user" and not missing_fields:
|
||
next_action = "render_preview"
|
||
options = []
|
||
question = ""
|
||
rationale = "当前申请信息足以先生成核对结果；附件和员工编号不应作为用户补填项阻塞申请预览。"
|
||
else:
|
||
options = self._sanitize_options(payload.get("options"), missing_fields)
|
||
question = self._clean_text(payload.get("question"))
|
||
rationale = self._clean_text(payload.get("rationale"))
|
||
return StewardSlotDecisionResponse(
|
||
decision_source="llm_function_call",
|
||
next_action=next_action, # type: ignore[arg-type]
|
||
required_fields=required_fields,
|
||
missing_fields=missing_fields,
|
||
question=question,
|
||
options=options,
|
||
rationale=rationale,
|
||
model_call_traces=traces,
|
||
)
|
||
|
||
@staticmethod
|
||
def _filter_blocking_fields(fields: list[str], task_type: str) -> list[str]:
|
||
if task_type != "expense_application":
|
||
return fields
|
||
return [field for field in fields if field not in APPLICATION_NON_BLOCKING_FIELDS]
|
||
|
||
@staticmethod
|
||
def _sanitize_fields(raw_fields: Any) -> list[str]:
|
||
fields: list[str] = []
|
||
if not isinstance(raw_fields, list):
|
||
return fields
|
||
for item in raw_fields:
|
||
key = str(item or "").strip()
|
||
if key in BUSINESS_CANONICAL_FIELDS and key not in fields:
|
||
fields.append(key)
|
||
return fields
|
||
|
||
def _sanitize_options(self, raw_options: Any, missing_fields: list[str]) -> list[StewardSlotOption]:
|
||
options: list[StewardSlotOption] = []
|
||
if isinstance(raw_options, list):
|
||
for item in raw_options:
|
||
if not isinstance(item, dict):
|
||
continue
|
||
field_key = str(item.get("field_key") or "").strip()
|
||
label = self._clean_text(item.get("label"))
|
||
value = self._clean_text(item.get("value")) or label
|
||
if not field_key or field_key not in BUSINESS_CANONICAL_FIELDS or not label or not value:
|
||
continue
|
||
options.append(
|
||
StewardSlotOption(
|
||
field_key=field_key,
|
||
label=label,
|
||
value=value,
|
||
description=self._clean_text(item.get("description")),
|
||
)
|
||
)
|
||
if not options and missing_fields and missing_fields[0] == "transport_mode":
|
||
options = [
|
||
StewardSlotOption(field_key="transport_mode", label="火车", value="火车", description="选择火车或高铁出行。"),
|
||
StewardSlotOption(field_key="transport_mode", label="飞机", value="飞机", description="选择飞机出行。"),
|
||
StewardSlotOption(field_key="transport_mode", label="轮船", value="轮船", description="选择轮船出行。"),
|
||
]
|
||
return options[:6]
|
||
|
||
def _build_rule_fallback(
|
||
self,
|
||
request: StewardSlotDecisionRequest,
|
||
traces: list[dict[str, Any]],
|
||
) -> StewardSlotDecisionResponse:
|
||
missing_fields = list(request.missing_fields)
|
||
if missing_fields:
|
||
field = missing_fields[0]
|
||
return StewardSlotDecisionResponse(
|
||
decision_source="rule_fallback",
|
||
next_action="ask_user",
|
||
required_fields=list(dict.fromkeys([*request.ontology_fields.keys(), *missing_fields])),
|
||
missing_fields=missing_fields,
|
||
question=self._build_fallback_question(field),
|
||
options=self._sanitize_options([], [field]),
|
||
rationale="模型字段决策暂不可用，我先按上游意图识别给出的本体缺口向你确认。",
|
||
model_call_traces=traces,
|
||
)
|
||
return StewardSlotDecisionResponse(
|
||
decision_source="rule_fallback",
|
||
next_action="render_preview",
|
||
required_fields=list(request.ontology_fields.keys()),
|
||
missing_fields=[],
|
||
question="",
|
||
options=[],
|
||
rationale="当前任务没有上游标记的关键字段缺口，可以先生成核对结果供你确认。",
|
||
model_call_traces=traces,
|
||
)
|
||
|
||
@staticmethod
|
||
def _build_fallback_question(field: str) -> str:
|
||
label = FIELD_CATALOG.get(field, {}).get("label") or field
|
||
if field == "transport_mode":
|
||
return "请问你这次打算怎么出行？可以选择火车、飞机或轮船。"
|
||
return f"当前还缺少{label}，请先补充后我再继续处理。"
|
||
|
||
@staticmethod
|
||
def _clean_text(value: Any) -> str:
|
||
return str(value or "").strip()
|