Files
X-Financial/document/development/agent plan/02_semantic_ontology.md

5.3 KiB
Raw Blame History

语义本体协议设计

1. 定位

语义本体协议是用户问题、定时任务、规则中心、MCP、数据库查询和 Agent 之间的统一中间层。

它解决的问题是:

  • 用户到底在问哪个业务域?
  • 这属于什么场景?
  • 用户想做什么?
  • 问题中涉及哪些对象?
  • 有没有时间、金额、状态、部门等过滤条件?
  • 是否涉及风险?
  • 下一步应该查知识库、查数据库、跑规则、调 MCP还是追问

2. 第一版核心字段

第一版建议只强制落 8 个字段。

{
  "domain": "",
  "scenario": "",
  "intent": "",
  "entities": [],
  "time_range": {},
  "constraints": {},
  "risk_signals": [],
  "next_step": ""
}

2.1 domain

一级业务域。

建议枚举:

reimbursement
accounts_receivable
accounts_payable
general_finance
system_operation

含义:

  • reimbursement:报销、差旅、发票、补件。
  • accounts_receivable:应收账款、客户开票、收款、账龄。
  • accounts_payable:应付账款、供应商发票、付款、对账。
  • general_finance:通用财务知识、制度、统计。
  • system_operation系统巡检、任务运行、规则维护、MCP 健康检查。

2.2 scenario

细分场景。

报销:

travel_reimbursement
daily_expense
invoice_validation
attachment_review
policy_overrun
reimbursement_audit

应收:

customer_invoice
collection_followup
receivable_aging
payment_matching
bad_debt_risk
contract_receivable

应付:

vendor_invoice
payment_request
payable_aging
vendor_reconciliation
invoice_matching
cash_outflow_forecast

系统运营:

daily_risk_scan
daily_finance_statistics
knowledge_accumulation
mcp_health_check
rule_quality_review

2.3 intent

用户或任务的意图。

建议枚举:

query
explain
create
validate
summarize
reconcile
monitor
predict
remind
generate
optimize

2.4 entities

识别出的业务对象。

统一结构:

{
  "type": "invoice",
  "value": "INV-202605001",
  "normalized_value": "INV-202605001",
  "role": "target",
  "confidence": 0.92
}

常见实体:

employee
department
customer
vendor
invoice
contract
reimbursement_request
payment_order
receipt
bank_transaction
cost_center
project
policy
approval_node
rule
task

2.5 time_range

统一描述时间。

{
  "raw": "上个月",
  "start": "2026-04-01",
  "end": "2026-04-30",
  "granularity": "month"
}

Hermes 定时任务也使用同一字段。

例如每日风险巡检:

{
  "raw": "昨日",
  "start": "2026-05-09",
  "end": "2026-05-09",
  "granularity": "day"
}

2.6 constraints

查询、判断或执行条件。

{
  "status": "overdue",
  "aging_days": ">30",
  "amount": {
    "operator": ">",
    "value": 50000,
    "currency": "CNY"
  },
  "department": "销售部",
  "risk_level": ["medium", "high"]
}

2.7 risk_signals

风险信号。

建议枚举:

duplicate_invoice
missing_attachment
policy_overrun
over_budget
overdue_receivable
bad_debt_risk
vendor_payment_risk
payment_mismatch
contract_mismatch
cashflow_pressure
mcp_unavailable
rule_quality_issue

2.8 next_step

下一步动作。

建议枚举:

answer
ask_clarification
query_database
run_rule
call_mcp
search_knowledge
create_draft
create_task
generate_report
notify_user
escalate_to_human

3. 扩展字段

后续可以增加:

{
  "schema_version": "1.1",
  "confidence": 0.86,
  "ambiguity": [],
  "missing_slots": [],
  "required_capabilities": [],
  "normalized_query": "",
  "permission_scope": {},
  "audit_tags": []
}

扩展原则:

  • 先不要把所有字段都做成数据库列。
  • 语义结果建议存 JSONB。
  • 使用 schema_version 管理版本。
  • Orchestrator 只依赖稳定字段。
  • 新字段以可选方式加入,不影响老任务。

4. 示例

4.1 用户查询应收账龄

用户问:

上个月哪些客户应收逾期超过 30 天?

解析:

{
  "domain": "accounts_receivable",
  "scenario": "receivable_aging",
  "intent": "query",
  "entities": [
    {
      "type": "customer",
      "value": "客户",
      "role": "group_by"
    }
  ],
  "time_range": {
    "raw": "上个月",
    "start": "2026-04-01",
    "end": "2026-04-30",
    "granularity": "month"
  },
  "constraints": {
    "aging_days": ">30",
    "status": "overdue"
  },
  "risk_signals": ["overdue_receivable"],
  "next_step": "query_database"
}

4.2 用户解释发票拦截

用户问:

这张发票为什么报销被拦截?

解析:

{
  "domain": "reimbursement",
  "scenario": "invoice_validation",
  "intent": "explain",
  "entities": [
    {
      "type": "invoice",
      "value": "这张发票",
      "role": "target"
    }
  ],
  "time_range": {},
  "constraints": {},
  "risk_signals": ["unknown"],
  "next_step": "run_rule"
}

4.3 Hermes 每日风险巡检

任务配置:

{
  "domain": "reimbursement",
  "scenario": "daily_risk_scan",
  "intent": "monitor",
  "entities": [],
  "time_range": {
    "raw": "昨日"
  },
  "constraints": {
    "risk_level": ["medium", "high"]
  },
  "risk_signals": [
    "duplicate_invoice",
    "missing_attachment",
    "policy_overrun"
  ],
  "next_step": "run_rule"
}