feat(steward): 拦截业务无关输入返回 off_topic 计划

- schemas/steward.py:StewardPlanResponse 新增 suggested_prompts 字段
- steward_planner.py:新增 STEWARD_BUSINESS_SIGNAL_KEYWORDS 与
  _is_business_irrelevant_input 守卫,在 build_plan 入口前置;
  新增 _build_off_topic_plan 构造 plan_status=off_topic 的引导计划
- steward_intent_agent.py:system prompt 追加业务无关约束
- test_steward_planner.py:覆盖 123/你好/纯标点走 off_topic,
  并验证正常业务输入不受守卫影响
This commit is contained in:
caoxiaozhu
2026-06-18 14:15:20 +08:00
parent b8915a29c0
commit cce19e4c40
4 changed files with 143 additions and 0 deletions

View File

@@ -115,6 +115,10 @@ class StewardIntentAgent:
"如果输入里出现 occurred_date、transport_type、reason_value 等别名,必须映射为 canonical 字段。"
"相对日期必须以 base_date 为准转换为明确日期。"
"thinking_events 只能是面向用户的过程摘要,不能暴露内部推理链。"
"如果用户输入与出差、费用、报销、申请等财务事项完全无关"
"(例如纯数字、问候、闲聊、无意义字符、单字符重复),"
"必须让 tasks 返回空数组,并在 thinking_events 中明确说明“未识别到财务事项”,"
"不要强行把无关输入识别为 expense_application 或 reimbursement 任务。"
),
},
{

View File

@@ -49,6 +49,27 @@ CITY_NAMES = (
"无锡",
)
# 业务信号关键词:用于判定输入是否与小财管家支持的财务事项相关。
# 只要清洗后的消息命中其中任意一个关键词,就视为业务相关;否则进入 off_topic 拦截。
STEWARD_BUSINESS_SIGNAL_KEYWORDS: tuple[str, ...] = (
# 动作词
"申请", "报销", "草稿", "提交", "审批", "保存", "发起", "创建", "核对", "归集",
# 差旅场景
"出差", "差旅", "费用", "交通", "住宿", "招待", "酒店", "机票", "航班", "高铁",
"动车", "火车", "出租车", "的士", "网约车", "打车", "地铁", "公交", "用餐", "餐饮", "宴请",
# 票据/凭证
"票据", "发票", "凭证", "行程单", "付款截图", "付款", "小票", "收据",
# 业务对象
"客户", "项目", "拜访", "会议", "培训", "部署", "实施", "支撑", "支持", "协助",
"调研", "驻场", "上线", "验收", "审核",
# 时间信号
"昨天", "前天", "明天", "后天", "下周", "下月", "近期", "月底", "今天", "上周", "上月",
# 金额/数量("天"用于"出差3天"等表达)
"金额", "", "", "", "", "",
# 复用城市名信号
*CITY_NAMES,
)
APPLICATION_SPLIT_PATTERN = re.compile(r"(?:^|[,。;;])[^,。;;]*?(?:申请|出差申请|差旅申请)[^,。;;]*")
REIMBURSEMENT_PATTERN = re.compile(r"(?:我要报销|还需要报销|需要报销|报销)([^,。;;?!\n]+)")
MONTH_DAY_PATTERN = re.compile(r"(?P<month>\d{1,2})\s*月\s*(?P<day>\d{1,2})\s*(?:日|号)?")
@@ -107,6 +128,9 @@ class StewardPlannerService:
raise ValueError("小财管家需要一段任务描述。")
base_date = self._resolve_base_date(request.client_now_iso, request.context_json)
# 业务无关输入拦截(纯数字、问候、闲聊、乱码等):在进入 LLM/规则兜底之前直接返回 off_topic 计划。
if self._is_business_irrelevant_input(message, request):
return self._build_off_topic_plan(request)
model_call_traces: list[dict[str, Any]] = []
fallback_reason = ""
if self.intent_agent is not None and self._should_use_model_intent_recognition(message, base_date, request):
@@ -159,6 +183,51 @@ class StewardPlannerService:
return False
return self._has_multiple_financial_demands(message)
@staticmethod
def _is_business_irrelevant_input(message: str, request: StewardPlanRequest) -> bool:
"""判断输入是否与小财管家支持的财务事项完全无关。
判定规则:消息去除所有空白后不含任何业务信号关键词,且没有上传附件,
即视为业务无关输入(如纯数字、问候、闲聊、乱码)。
"""
if request.attachments:
return False
compact = re.sub(r"\s+", "", message)
if not compact:
return False
return not any(keyword in compact for keyword in STEWARD_BUSINESS_SIGNAL_KEYWORDS)
def _build_off_topic_plan(self, request: StewardPlanRequest) -> StewardPlanResponse:
"""业务无关输入的兜底计划:明确告知用户未识别到财务事项,并给出话术示例。"""
return StewardPlanResponse(
plan_id=f"steward_plan_{uuid.uuid4().hex[:12]}",
plan_status="off_topic",
planning_source="rule_fallback",
next_action="none",
summary="这看起来跟财务任务没什么关系,小财管家没识别到费用申请或费用报销的意图。",
thinking_events=[
StewardThinkingEvent(
event_id="intent_agent_off_topic",
stage="off_topic",
title="未识别到财务事项",
content=(
"我检查了这句话,没有发现费用申请、报销、出差、交通、招待等财务线索。"
"如果你确实是要处理财务任务,可以参考下面的示例换一种说法。"
),
)
],
tasks=[],
attachment_groups=[],
confirmation_groups=[],
candidate_flows=[],
suggested_prompts=[
"我想要申请明天去北京出差3天支撑客户现场实施",
"我要报销昨天的交通费",
"报销上周出差上海的费用",
],
model_call_traces=[],
)
def _build_rule_fallback_plan(
self,
request: StewardPlanRequest,