AI Semantic Interoperability / RDF / OWL / SHACL Playbook
以下来源作为语义互操作和金融语义建模的官方锚点, 不构成法律、合规、模型验证或供应商选型意见。访问日期按 2026-06-29 记录。
AI Semantic Interoperability / RDF / OWL / SHACL Playbook
面向对象: AI Architect / AI PM / AI BA / Data Product Manager / Knowledge Governance Lead / EvalOps Lead / 金融零售 AI 平台负责人。 核心问题: 当 RAG、GraphRAG、AI agents、工具调用、事件流、指标语义层和审计追踪都在消费同一批业务概念时, 如何用 RDF / OWL / SHACL 设计可互操作、可验证、可追溯的 AI semantic contract。 学习目标: 能为 KYC、AML、信贷、客服、财富、监管响应等场景建立 graph facts、ontology semantics、constraint validation、schema boundary、eval contract 和 portfolio evidence。 前提假设: 读者已经具备成熟 PM / BA / CBAP 能力, 本手册不讲需求入门或语义网基础教程, 重点训练 AI 架构、知识治理、语义互操作和金融零售落地表达。
Source Anchors
以下来源作为语义互操作和金融语义建模的官方锚点, 不构成法律、合规、模型验证或供应商选型意见。访问日期按 2026-06-29 记录。
| Anchor | Official / primary source | 本手册使用方式 |
|---|---|---|
| W3C RDF | https://www.w3.org/RDF/ | 用 RDF 的 graph data model、IRI、triple、named graph 思路设计跨系统事实表达、证据链和 AI trace。 |
| W3C OWL | https://www.w3.org/OWL/ | 用 OWL 的 class、property、restriction、reasoning、一致性检查思路表达业务 ontology semantics。 |
| W3C SHACL | https://www.w3.org/TR/shacl/ | 用 SHACL 的 shapes graph、data graph、constraint validation、validation report 思路建立语义数据质量门禁。 |
| JSON Schema | https://json-schema.org/ | 用 JSON Schema 管理 JSON payload 的结构、类型、必填、枚举和 API contract, 并与 SHACL 划清边界。 |
| FIBO | https://spec.edmcouncil.org/fibo/ | 用 FIBO 作为金融服务业务概念、关系和 ontology 参考, 不机械照搬全量本体。 |
Standards-to-artifacts:
| Standard lens | 可产出的 artifact | 高级表达 |
|---|---|---|
| RDF | Semantic fact graph, evidence graph, named graph strategy, trace triple catalog | "我把 AI 使用的事实、证据、来源和版本表达成可交换 graph facts, 让跨系统含义可追踪。" |
| OWL | Ontology slice, class/property catalog, semantic ADR, reasoning impact memo | "我用 ontology semantics 定义概念边界和关系含义, 不是让模型靠相似文本猜业务词。" |
| SHACL | Constraint checklist, ingestion validation report, semantic release gate | "我把语义契约变成可执行 validation, 阻止错配字段、缺证据 claim 和过期政策进入 AI 上下文。" |
| JSON Schema | API payload schema, tool input contract, event envelope schema | "我用 JSON Schema 管接口形状, 用 SHACL 管图谱语义约束, 两者服务不同层。" |
| FIBO | Financial ontology reference, concept alignment map, term reuse decision | "我借鉴 FIBO 管金融概念一致性, 同时保留企业内部上下文和风险边界。" |
1. One-Sentence Positioning / 一句话定位
AI semantic interoperability 的目标不是把企业所有数据都改造成 RDF, 而是让 AI 系统在跨文档、跨图谱、跨 API、跨事件、跨指标和跨评估时使用同一套可治理语义契约。
business meaning
-> stable identifiers
-> RDF graph facts
-> OWL ontology semantics
-> SHACL constraint validation
-> RAG / GraphRAG / tool / event / eval contracts
-> auditable AI behavior
一句话:
Semantic interoperability for AI = 用可交换事实、明确语义和可执行约束, 降低 AI 幻觉、概念混淆、字段错配、指标歧义和审计不可追踪。
金融零售 AI 中最常见的失控点不是模型不会说话, 而是模型把语义边界弄错:
- "customer" 在 CRM、core banking、KYC、merchant acquiring、wealth 里不是同一个对象。
- "approval" 可能是信贷审批、合规批准、经理授权、系统状态或 AI recommendation accepted。
- "high risk" 在 AML、credit、customer vulnerability、cyber fraud、wealth suitability 中语义不同。
- "active account" 在运营指标、监管报送、产品资格和收费规则中口径不同。
- "evidence" 可能是原始交易、政策条款、case note、screening hit、customer document 或 evaluator rationale。
因此, AI Architect / PM / BA 需要交付的不是一个更长 prompt, 而是一个 semantic control plane:
| Control question | 低成熟度做法 | 高成熟度做法 |
|---|---|---|
| 这个概念到底指什么 | 在 prompt 里解释一次 | concept URI + glossary + ontology class + owner + eval cases |
| 这个事实从哪里来 | 让 RAG 引用文档链接 | RDF named graph + source version + evidence span + trace id |
| 这个字段能不能映射 | 用同名字段猜 | semantic mapping contract + SHACL validation + exception queue |
| 这个工具能否调用 | API schema 通过就放行 | JSON Schema + ontology-aware permission + business invariant check |
| 这个指标能否被解释 | dashboard 有数字就回答 | metric semantic contract + lineage + approved dimensions |
| 出错后如何审计 | 查日志和聊天记录 | answer claim -> graph fact -> source version -> constraint report -> release gate |
2. 为什么 AI 需要 Semantic Interoperability
传统系统集成主要关心 payload 能不能传、字段能不能解析、接口能不能调用。AI 系统还要解决更难的问题: 语义是否一致、证据是否支撑、上下文是否正确、约束是否可验证、输出是否可审计。
2.1 AI 放大语义缺陷的方式
| 语义缺陷 | 传统系统中的表现 | AI 系统中的风险 | 语义互操作控制 |
|---|---|---|---|
| 同名异义 | 报表字段解释不同 | LLM 把 AML risk 当成 credit risk 解释 | bounded context + concept URI + disjoint class |
| 异名同义 | 搜索召回不全 | RAG 找不到 UBO / beneficial owner / controlling person 的等价知识 | synonym scheme + ontology alignment |
| 字段错配 | ETL reconciliation 失败 | tool call 把 customer_id 传成 party_id, 返回错误客户事实 | JSON Schema + semantic mapping + SHACL node shape |
| 指标歧义 | dashboard 数字不一致 | AI summary 对经营原因作出错误解释 | metric contract + approved dimensions + lineage |
| 版本混用 | 手工对账发现 | AI 引用旧政策、旧产品条款或未生效规则 | named graph + effective date + validation gate |
| 证据断裂 | 审计追问困难 | 回答看似有引用, 但 claim 无法追到支撑证据 | claim graph + evidence span + citation validator |
| 上下文污染 | 数据湖标签混乱 | enterprise copilot 跨域使用敏感知识或错误规则 | context boundary + access policy + graph partition |
2.2 AI 语义互操作的五个目标
| Goal | 说明 | 金融零售例子 |
|---|---|---|
| Meaning consistency | 同一业务概念在不同系统中可对齐, 不同概念不会被误合并 | KYC customer, CRM customer, merchant beneficial owner 的映射边界 |
| Evidence grounding | AI claim 能追到结构化事实、文档条款、图谱路径或事件记录 | "需要 EDD" 必须指向政策条款、客户类型和风险触发条件 |
| Constraint enforcement | 语义约束能在 ingestion、retrieval、tool call、event processing、eval 前执行 | high-risk policy claim 必须有 source authority、effective period 和 citation |
| Cross-system exchange | RAG、GraphRAG、tool、event、metric、eval 能消费相同语义标识 | concept:beneficial-owner 跨 glossary、ontology、RAG metadata、eval case 复用 |
| Audit replay | 历史答案能按当时的 graph、ontology、constraints、source version 和 model trace 回放 | 监管问询时重建 2026-05-15 的 KYC assistant 回答证据链 |
2.3 语义互操作不是单一技术选型
| 误区 | 为什么危险 | 更好的做法 |
|---|---|---|
| 全部上 RDF 才算语义互操作 | 迁移成本高, 业务团队无法吸收 | 先定义 stable IDs、concept registry、mapping profile, 再选择 RDF / property graph / relational implementation |
| OWL ontology 越完整越好 | 全量企业本体会拖慢产品迭代 | 只建与 AI risk、retrieval、tool、eval、audit 相关的 ontology slice |
| SHACL 只用于数据质量 | 会错过 AI 语义约束和 release gate | 把 SHACL 用于 source、claim、evidence、tool input、event payload 的语义门禁 |
| JSON Schema 可以替代 SHACL | JSON Schema 管结构, 不表达 graph relation 和跨节点语义 | JSON Schema 管 payload shape, SHACL 管 graph meaning 和约束 |
| FIBO 可直接替代内部模型 | 行业本体不等于企业上下文和产品责任 | 以 FIBO 为 reference ontology, 做 context-specific alignment |
3. 核心框架: Semantic Contract Stack
Semantic Contract Stack 用来把标准、领域语言、系统接口和 AI 行为连接成可运行的治理链路。
| Layer | 核心对象 | 决策问题 | 核心产物 |
|---|---|---|---|
| 1. Domain context | bounded context, owner, risk tier | 这个语义契约属于哪个业务上下文 | Semantic Scope Memo |
| 2. Concept identity | concept URI, label, definition | 概念如何稳定命名, 如何处理同义词和禁用混淆词 | Concept Registry |
| 3. Graph facts | subject, predicate, object, named graph | 哪些事实、证据、版本和来源需要可交换 | RDF Mapping Profile |
| 4. Ontology semantics | class, property, domain/range, restriction | 哪些概念关系和不变量需要显式表达 | Ontology Slice |
| 5. Constraint validation | shapes graph, data graph, severity, report | 哪些语义错误必须在上线前被拦截 | SHACL Constraint Pack |
| 6. Payload contracts | JSON Schema, API/event/tool schema | AI 工具和事件如何保证结构正确 | JSON Schema Contract |
| 7. AI consumption | RAG, GraphRAG, tool, event, metric, eval | AI 能如何消费、组合和输出语义对象 | AI Semantic Contract |
| 8. Operations | drift, incident, versioning, stewardship | 语义变化如何监控、批准、回归和审计 | Semantic Governance Runbook |
一句话能力表达:
我能把金融零售 AI 的业务概念、事实图谱、本体语义、约束验证、接口 schema 和评估门禁设计成一套可运行的 semantic contract, 而不是把语义交给模型临场猜测。
4. RDF / OWL / SHACL 分工
RDF、OWL、SHACL 经常被混在一起讨论。架构表达中要先划清三者的职责:
RDF answers: what facts and relationships are asserted?
OWL answers: what do those concepts and relationships mean?
SHACL answers: does this graph satisfy the agreed constraints?
4.1 RDF: Graph Facts
RDF 适合表达可交换的事实和关系, 尤其当 AI 需要跨系统追踪 source、evidence、claim、version、permission 和 graph path 时。
| RDF capability | AI 架构用途 | 金融零售例子 |
|---|---|---|
| IRI / URI identity | 给概念、条款、客户、case、claim、evidence、metric 分配稳定标识 | concept:beneficial-owner, clause:kyc-v6-1-3-4, claim:kyc-doc-014 |
| Triple | 用 subject-predicate-object 表达事实关系 | claim:014 supportedBy clause:kyc-v6-1-3-4 |
| Named graph | 为事实集合绑定来源、版本、权限、业务日期和 ingestion snapshot | graph:kyc-policy-prod-2026-06-15 |
| RDF dataset | 同时管理多个 graph, 支持隔离和回放 | current policy graph, historical policy graph, evaluation graph |
| Standard interchange | 让 graph facts 可以跨工具、组织、推理器和验证器交换 | ontology registry, graph store, audit system, eval harness 共享事实 |
RDF 最小落地不一定是部署 triple store。可以先做 RDF-inspired mapping:
| 对象 | 示例 ID | 必备 metadata |
|---|---|---|
| Concept | concept:kyc.beneficial_owner | label, definition, context, owner, synonyms, source |
| Source | src:kyc-policy-v6-1 | authority, approval status, version, effective date, permission |
| Clause | clause:kyc-v6-1-3-4 | source, section, page, paragraph, validity |
| Claim | claim:kyc-doc-required-014 | statement, claim type, support status, risk tier |
| Evidence span | evidence:kyc-v6-1-p12-r3 | document, offset/page/table row, extraction confidence |
| Graph | graph:kyc-prod-2026-06-15 | graph type, source snapshot, ontology version, constraints version |
4.2 OWL: Ontology Semantics
OWL 适合表达概念、类、关系和可推理语义。它不是普通数据字典, 而是告诉系统 "这些东西在业务上意味着什么"。
| OWL-style decision | AI 架构用途 | 金融零售例子 |
|---|---|---|
| Class | 定义核心业务对象类型 | Customer, Party, Account, BeneficialOwner, PolicyClause, AMLCase |
| Object property | 定义对象之间的关系 | owns, controls, appliesTo, requires, overrides, supportedBy |
| Data property | 定义对象属性 | effectiveDate, authorityLevel, jurisdiction, confidenceScore |
| Domain / range | 限制关系两端的语义类型 | requires domain PolicyClause, range DocumentRequirement |
| Subclass | 管理层级和继承 | HighRiskCustomer subclass CustomerSegment |
| Disjointness | 防止概念误合并 | AMLRedFlag disjointWith CreditApprovalReason |
| Restriction | 表达业务不变量或分类条件 | EnhancedDueDiligenceRequirement appliesTo HighRiskCustomer |
| Reasoning | 显式化隐含知识或发现不一致 | 如果 customer controls legal entity 且 ownership >= threshold, 推导 beneficial owner candidate |
OWL 的实际使用要按风险分层:
| 场景 | 推荐深度 | 原因 |
|---|---|---|
| 客服 FAQ / SOP | SKOS-style vocabulary + lightweight ontology | 核心是术语一致、版本和引用, 不需要复杂推理 |
| KYC / CDD policy | Lightweight OWL-style ontology + SHACL | 客户类型、文件要求、地区、例外、风险等级有明确关系 |
| AML investigation graph | Domain ontology + entity relation model + graph path semantics | 多跳关系和证据路径是分析核心 |
| Credit decision support | Ontology slice + business rule engine + HITL | 需保护信贷决策边界, 不让 LLM 直接做审批 |
| Enterprise financial ontology | FIBO alignment + internal bounded contexts | 需要跨域共用概念, 但必须保留上下文边界 |
4.3 SHACL: Constraint Validation
SHACL 适合把语义契约变成可执行验证。它检查的是 RDF data graph 是否满足 shapes graph 中定义的条件。
| SHACL capability | AI 架构用途 | 金融零售例子 |
|---|---|---|
| Node shape | 验证某类节点必须具备哪些属性和关系 | PolicyClaim 必须有 source, evidence, effective period |
| Property shape | 验证属性的数据类型、数量、枚举、范围 | authorityLevel 必须在 A0-A6 中 |
| Target | 指定哪些节点需要被验证 | 所有 ai:HighRiskClaim 都进入 validation |
| Cardinality | 验证 min/max count | high-risk answer claim 至少一个 evidence span |
| Class constraint | 验证对象类型 | supportedBy 的 object 必须是 EvidenceSpan |
| Value set | 验证枚举和允许值 | jurisdiction 必须来自 approved jurisdiction scheme |
| Severity | 区分 blocker、warning、info | permission missing 是 violation, stale glossary 是 warning |
| Validation report | 形成 release gate 和 audit evidence | ingestion 失败原因、影响节点、owner、修复状态 |
SHACL 在 AI lifecycle 中可以放在多个门:
| Gate | 验证对象 | 失败动作 |
|---|---|---|
| Source ingestion gate | source, clause, metadata, evidence span | 不进入 production index 或 graph |
| Ontology release gate | class/property changes, mapping changes | 架构评审和回归评估 |
| Retrieval context gate | retrieved chunks / graph facts | 低权威、过期、无权限证据不得进入 prompt |
| Tool input gate | tool semantic object before execution | 阻止跨客户、跨权限、越权动作 |
| Event processing gate | domain event graph | 事件缺少 aggregate、actor、effective time 时进入 exception queue |
| Eval gate | answer claim, citation, graph path | critical claim 缺 evidence 时 release fail |
5. JSON Schema 与 SHACL 的边界
JSON Schema 和 SHACL 都能做 validation, 但它们解决的问题不同。架构评审中要避免把接口结构检查误当成语义正确。
5.1 决策边界
| 问题 | JSON Schema 更适合 | SHACL 更适合 |
|---|---|---|
| Payload 是否是合法 JSON | 是 | 否 |
| 字段是否必填、类型是否正确 | 是 | 可做, 但不是最佳入口 |
| API / tool input 的结构契约 | 是 | 用于补充语义验证 |
| 事件 envelope 是否稳定 | 是 | 用于验证事件语义关系 |
| graph 中节点和关系是否满足业务语义 | 否 | 是 |
| 跨节点关系是否一致 | 弱 | 是 |
| 同一 concept 是否来自批准词表 | 可枚举, 维护成本高 | 是 |
| claim 是否由 evidence span 支撑 | 否 | 是 |
| policy version 是否与 business date 匹配 | 需要应用代码 | SHACL + SPARQL / rules 更合适 |
| validation report 是否可作为 RDF evidence | 否 | 是 |
5.2 推荐组合模式
JSON Schema:
validates message shape, types, required fields, enums, arrays, nested objects
Semantic mapper:
maps JSON payload fields to canonical concepts and graph objects
SHACL:
validates graph constraints, relationships, evidence, authority, version, permissions
5.3 示例: AI Tool Input Contract
{
"$id": "https://example.bank/schemas/tools/create-kyc-remediation-task.json",
"type": "object",
"required": ["case_id", "customer_party_id", "task_type", "reason_code", "evidence_claim_id"],
"properties": {
"case_id": { "type": "string", "pattern": "^KYC-[0-9]{8}$" },
"customer_party_id": { "type": "string" },
"task_type": { "type": "string", "enum": ["request_document", "escalate_review"] },
"reason_code": { "type": "string", "enum": ["missing_ubo_document", "expired_identity_document", "jurisdiction_exception"] },
"evidence_claim_id": { "type": "string" }
},
"additionalProperties": false
}
JSON Schema 能确认 payload 结构正确, 但不能单独回答:
customer_party_id是否属于case_id对应的 KYC case。- 当前用户是否有权为该客户创建 remediation task。
reason_code是否由当前有效政策支持。evidence_claim_id是否有 source authority 和 evidence span。- 该 task 是否越过人工审批边界。
这些应由 semantic mapper、SHACL validation、policy engine 和 audit trace 共同完成。
5.4 示例: SHACL-style Constraint Intent
ex:KYCRemediationTaskShape
a sh:NodeShape ;
sh:targetClass ai:KYCRemediationTask ;
sh:property [
sh:path ai:belongsToCase ;
sh:class ai:KYCCase ;
sh:minCount 1 ;
sh:maxCount 1 ;
sh:severity sh:Violation ;
] ;
sh:property [
sh:path ai:supportedBy ;
sh:class ai:PolicyClaim ;
sh:minCount 1 ;
sh:severity sh:Violation ;
] ;
sh:property [
sh:path ai:requestedBy ;
sh:class ai:AuthorizedActor ;
sh:minCount 1 ;
sh:severity sh:Violation ;
] .
作品集表达:
JSON Schema 保证 AI 工具调用的消息形状正确; SHACL 保证消息背后的业务对象、证据、权限和语义关系正确。高风险金融零售 AI 需要两者组合, 不能只靠 API schema。
6. RAG / GraphRAG / Tool / Event / Eval 的语义契约
AI 语义互操作的价值在于让多个 AI 消费面使用同一套 concept identity、source authority、graph facts、constraints 和 audit trace。
6.1 RAG Semantic Contract
| Contract area | 必须声明 | 失败模式 |
|---|---|---|
| Source authority | 哪些 source 可进入 index, authority order, approval status | FAQ 覆盖正式政策 |
| Concept metadata | chunk 绑定哪些 concept URI、bounded context、jurisdiction、product | 召回相似但错误上下文 |
| Version validity | effective date, expiry date, business date filter | 引用过期或未生效政策 |
| Claim extraction | chunk 支撑哪些 claim, claim type, risk tier | answer 有 citation 但 claim 不被证据支持 |
| Permission | role, purpose, case binding, citation visibility | 检索到用户无权查看的 investigation note |
| Constraint gate | ingestion 和 retrieval 前通过哪些 SHACL shapes | 缺 owner、缺 evidence span 的知识进入 prompt |
RAG 的关键不是 "top-k 更准", 而是 "top-k 中的每条证据是否语义可用"。
6.2 GraphRAG Semantic Contract
| Contract area | 必须声明 | 失败模式 |
|---|---|---|
| Entity identity | entity resolution 规则、confidence、merge/split policy | 把两个同名客户或商户错误合并 |
| Relation semantics | relation type、domain/range、direction、valid time | 把 "shared address" 当成 "controls" |
| Path admissibility | 哪些 graph paths 可用于解释, 哪些只作 investigation lead | 模型把弱关联说成事实结论 |
| Evidence binding | 每条边是否有 source、event、document 或 system record 支撑 | graph path 看似合理但无证据 |
| Human control | 哪些 path 只能提示人工复核 | 自动判定可疑活动或客户风险 |
| Evaluation | path precision、entity resolution accuracy、unsupported relation rate | GraphRAG 讲出漂亮但不可审计的关系故事 |
6.3 Tool Semantic Contract
| Contract area | 必须声明 | 金融零售例子 |
|---|---|---|
| Tool intent | 工具服务哪个 bounded context 和 task mode | create_kyc_remediation_task 只能服务 KYC remediation |
| Input schema | JSON Schema 管字段形状 | case id, party id, reason code, evidence claim id |
| Semantic preconditions | SHACL / policy engine 管业务前置条件 | customer belongs to case, claim supported by active policy |
| Permission | actor role, purpose, entitlement, approval requirement | branch staff 不能创建 AML escalation |
| Output contract | 工具返回哪些 domain event 或 graph facts | KYCRemediationTaskCreated event + audit trace |
| Reversal / audit | 是否可撤销, 如何回放 | task cancellation reason, actor, timestamp, source claim |
6.4 Event Semantic Contract
AI agents 不应只消费 "消息", 而应消费带语义的 domain events。
| Event field | 语义要求 | 示例 |
|---|---|---|
| event_id | 全局唯一, 可追踪 | evt:kyc-document-missing-20260629-001 |
| event_type | 来自批准事件词表 | KYC.DocumentRequirementMissing |
| aggregate_id | 事件属于哪个 aggregate | case:KYC-20260629 |
| actor | 事件发起者或系统 | user:kyc-analyst-17 |
| business_time | 业务发生时间 | 2026-06-29T09:21:00-05:00 |
| source_graph | 事件依据的 graph snapshot | graph:kyc-prod-2026-06-15 |
| policy_version | 使用的政策版本 | src:kyc-policy-v6-1 |
| evidence_claim | 支撑事件的 claim | claim:kyc-doc-required-014 |
| allowed_ai_use | AI 可如何消费该事件 | summarize, alert, eval, audit |
6.5 Eval Semantic Contract
Eval 不是只给答案打分, 而是验证 AI 是否遵守语义契约。
| Eval object | 语义契约 | Gate behavior |
|---|---|---|
| Query | query intent, bounded context, role, business date | 歧义时必须 clarify |
| Gold evidence | source authority, clause, graph path, valid time | 找不到 gold evidence 不得硬答 |
| Hard negatives | 相似但错误来源、旧版本、错误上下文 | 被检索进入 final context 视为 failure |
| Expected behavior | answer, refuse, escalate, draft only, ask clarification | 高风险越权输出为 release blocker |
| Claim support | 每个 critical claim 绑定 evidence span | unsupported critical claim = fail |
| Semantic failure taxonomy | context confusion, authority inversion, version error, field mismatch | 按 failure type 回流治理 owner |
7. 金融零售案例
7.1 KYC / CDD Policy Assistant
| 设计面 | 决策 |
|---|---|
| Semantic risk | UBO、beneficial owner、controller、authorized signer、customer、party 在系统和政策中语义不同 |
| RDF graph facts | customer party, legal entity, beneficial owner candidate, document requirement, policy clause, evidence span |
| OWL semantics | BeneficialOwner, ControllingPerson, LegalEntityCustomer, DocumentRequirement, EnhancedDueDiligenceRequirement |
| SHACL constraints | high-risk KYC claim 必须有 approved policy source、effective period、jurisdiction、evidence span |
| JSON Schema role | KYC remediation tool input、document request event、case update payload |
| AI controls | 不自动通过 KYC, 只解释文件要求、缺失材料、政策依据和人工升级 |
| Eval slices | customer type, jurisdiction, policy version, UBO threshold, document type, branch vs compliance role |
| Portfolio evidence | semantic contract card, ontology slice, SHACL validation report, version regression pack |
高级表达:
KYC assistant 的核心不是把政策 PDF 放进向量库, 而是把客户身份、控制关系、文件要求、政策条款和证据引用建立成可验证语义契约, 防止 AI 把相似术语混成同一个合规义务。
7.2 AML Investigation Graph / GraphRAG
| 设计面 | 决策 |
|---|---|
| Semantic risk | shared address、shared device、beneficial ownership、transaction counterparty、screening hit 的风险含义不同 |
| RDF graph facts | customer-account-transaction-counterparty-device-address-case 关系, 每条边绑定来源和 valid time |
| OWL semantics | AMLCase, Alert, Transaction, Counterparty, RedFlag, TypologyIndicator, EvidencePath |
| SHACL constraints | GraphRAG path 必须区分 direct evidence、weak signal、investigation lead, 不得把弱关联当结论 |
| JSON Schema role | alert event、case note draft、analyst feedback、tool action payload |
| AI controls | 可总结模式和草拟 narrative, 不自动提交 SAR、不关闭 case |
| Eval slices | path precision, entity resolution, missing red flag, unsupported relation, typology mismatch |
| Portfolio evidence | GraphRAG semantic path contract, entity resolution decision memo, SHACL path validation checklist |
高级表达:
AML GraphRAG 的价值不是展示更复杂的图, 而是让每条风险路径有可解释关系类型、证据来源、时间有效性和人工复核边界。
7.3 Credit Origination Assistant
| 设计面 | 决策 |
|---|---|
| Semantic risk | approval condition、adverse action、reason code、exception、policy override 具有高影响决策含义 |
| RDF graph facts | application, borrower, document, credit memo claim, policy clause, condition, human decision |
| OWL semantics | LoanApplication, CreditPolicyClause, ApprovalCondition, AdverseActionReason, ExceptionReview |
| SHACL constraints | AI-generated credit memo claim 必须绑定 source evidence, 不得产生 automated approval / decline claim |
| JSON Schema role | memo draft payload、condition checklist、review workflow event |
| AI controls | 草拟 memo、检查材料、引用政策; 最终 approve / decline 留给受控流程 |
| Eval slices | policy version, product, jurisdiction, protected attribute exclusion, reason code consistency |
| Portfolio evidence | decision boundary shape, memo claim support report, prohibited output regression set |
高级表达:
信贷 AI 的语义互操作要服务于 decision support, 不是 decision automation。它要把政策、申请事实、reason code 和人工决策边界连接起来, 并且让越界行为可测试。
7.4 Customer Servicing / Complaint AI
| 设计面 | 决策 |
|---|---|
| Semantic risk | complaint、inquiry、dispute、escalation、service recovery、fee waiver 的业务含义和监管后果不同 |
| RDF graph facts | service case, interaction, product term, complaint category, escalation reason, response claim |
| OWL semantics | ServiceCase, Complaint, FeeDispute, ProductTerm, EscalationRequirement, CustomerVulnerability |
| SHACL constraints | customer-facing response claim 必须区分 explanation、commitment、escalation; fee waiver commitment 需要 human approval |
| JSON Schema role | response draft, escalation task, complaint event payload |
| AI controls | 可解释政策和草拟回复; 不承诺费用减免、授信结果、监管解释或赔偿 |
| Eval slices | vulnerable customer, product line, channel, complaint category, false promise, escalation correctness |
| Portfolio evidence | customer-facing claim taxonomy, semantic response validator, complaint metric linkage |
高级表达:
客服 AI 的风险不是回答不够流畅, 而是把解释、承诺、升级和监管表述混在一起。语义契约要让这些 speech acts 可区分、可限制、可审计。
7.5 AI Value Office / Metrics Semantic Layer
| 设计面 | 决策 |
|---|---|
| Semantic risk | adoption、usage、deflection、containment、resolution、ROI、risk reduction 的指标口径容易混用 |
| RDF graph facts | metric, measure, dimension, owner, formula, data contract, dashboard, eval report, release decision |
| OWL semantics | Metric, BusinessOutcome, EvalMetric, WorkflowMetric, GuardrailMetric, AIUseCase |
| SHACL constraints | official AI metric 必须有 owner、formula、grain、allowed dimensions、lineage、quality SLO |
| JSON Schema role | metric API request/response, portfolio dashboard event |
| AI controls | NL analytics 只能解释 approved metrics, 歧义指标必须追问 |
| Eval slices | metric ambiguity, restricted slice, causal overclaim, stale metric definition |
| Portfolio evidence | metric semantic contract, lineage graph, NL analytics guardrail report |
高级表达:
AI Value Office 的指标层必须是语义契约, 不是 dashboard 拼接。否则 AI 会把 adoption、usage、business value 和 risk reduction 混成一套看似合理但不可审计的故事。
8. Semantic Governance Operating Model
语义互操作需要治理角色, 不能只靠平台团队维护 ontology 文件。
8.1 RACI
| Activity | Product / BA | Domain Owner | Architecture | Knowledge / Data Steward | Risk / Compliance | EvalOps |
|---|---|---|---|---|---|---|
| Define semantic scope | A | R | R | C | C | C |
| Approve concept definition | R | A | C | R | C | C |
| Design ontology slice | C | R | A | R | C | C |
| Align FIBO / external ontology | C | C | A | R | C | C |
| Define SHACL constraints | R | C | A | R | C | R |
| Approve source authority | C | A | C | R | R | C |
| Validate tool/event semantics | R | C | A | R | R | R |
| Maintain eval semantic taxonomy | A | R | C | C | R | R |
| Review semantic drift | R | A | R | R | C | R |
| Decide release gate | A | R | R | C | R | R |
8.2 Semantic Change Control
| Change type | 示例 | Required control |
|---|---|---|
| Label change | "UBO" display label 改为 "Beneficial Owner" | glossary update, no graph migration if URI stable |
| Definition change | beneficial owner threshold 或适用范围变化 | domain owner approval, ontology version, eval regression |
| Relationship change | controls relation 新增 indirect control path | ontology ADR, SHACL update, GraphRAG path regression |
| Source authority change | FAQ 降级, policy 新版本生效 | source registry update, RAG index refresh, version tests |
| Tool semantic change | create task tool 新增 reason code | JSON Schema version, SHACL shape update, permission review |
| Metric semantic change | adoption metric 从 login 改为 completed workflow | metric contract major version, consumer impact analysis |
| FIBO alignment change | internal Party 映射到 FIBO party concept | alignment decision memo, downstream mapping test |
8.3 Semantic Drift Signals
| Drift signal | 可能根因 | 治理动作 |
|---|---|---|
| Eval 中 context confusion 上升 | 新业务术语未入 glossary 或 ontology | 更新 concept registry, 添加 hard negatives |
| Retrieval 命中旧术语 | 同义词治理缺口或旧文档仍在 production index | source deprecation, synonym review, retrieval regression |
| Tool validation failure 增加 | API schema 变化未同步 semantic mapper | schema-to-ontology impact review |
| GraphRAG unsupported relation 增加 | entity resolution 或关系抽取质量下降 | relation evidence audit, extraction model rollback |
| 指标解释歧义增加 | metric glossary 和 semantic layer 未同步 | metric contract refresh, NL analytics guardrail update |
| 审计回放失败 | graph version、constraint version 或 source snapshot 缺失 | trace schema hardening, release gate block until fixed |
9. Templates / Portfolio Artifacts
下面模板使用具体示例值, 目的是形成作品集证据和架构评审材料, 不是空白表单。
9.1 Semantic Contract Card
# Semantic Contract Card: KYC Beneficial Owner Requirement
## Identity
- Contract ID: semcon:kyc-beneficial-owner-requirement-v1
- Bounded context: KYC / Customer Due Diligence
- Risk tier: High
- Accountable owner: Head of KYC Policy
- Responsible steward: KYC Knowledge Governance Lead
- Architecture owner: AI Knowledge Platform Architect
## Business Meaning
- Approved concept URI: concept:kyc.beneficial_owner
- Approved definition: A natural person who owns, controls, or benefits from a legal entity customer under the active KYC policy threshold and jurisdiction rules.
- Not the same as: authorized signer, relationship manager, primary contact, credit guarantor.
- Source authority: src:kyc-policy-v6-1, src:regulatory-kyc-guidance-us-2026.
## RDF Facts
- Named graph: graph:kyc-prod-2026-06-15
- Core triples:
- clause:kyc-v6-1-3-4 ai:defines concept:kyc.beneficial_owner
- claim:kyc-bo-doc-014 ai:supportedBy evidence:kyc-v6-1-p12-r3
- claim:kyc-bo-doc-014 ai:appliesTo concept:customer.legal_entity
## Ontology Semantics
- Class: ai:BeneficialOwner
- Superclass: ai:ControllingPerson
- Related classes: ai:LegalEntityCustomer, ai:DocumentRequirement, ai:EnhancedDueDiligenceRequirement
- Disjoint classes: ai:AuthorizedSignerOnly, ai:RelationshipManager
## SHACL Constraints
- High-risk KYC claims require source authority A0 or A1.
- Claims require at least one evidence span with source version and effective period.
- Claims must include jurisdiction and customer segment.
- Any claim using retired source status fails release gate.
## AI Consumption
- RAG: allowed for policy explanation with claim-level citation.
- GraphRAG: allowed for ownership/control path explanation when evidence path is present.
- Tool: allowed to create remediation task only after SHACL validation and user entitlement check.
- Eval: included in KYC-BO-001 through KYC-BO-018 regression cases.
## Audit
- Required trace: user role, query, business date, graph version, ontology version, shapes version, source version, retrieved evidence, answer claims.
- Release gate: unsupported high-risk claim count must be 0.
9.2 Ontology Slice
| Element | Example |
|---|---|
| Slice name | kyc-cdd-requirement-slice |
| Scope | KYC policy explanation, document requirement reasoning, remediation task support |
| Out of scope | Final KYC approval, customer risk rating override, automated account opening |
| Reference ontology | FIBO for Party / Organization reference; internal KYC ontology for policy-specific obligations |
| Core classes | Party, NaturalPerson, LegalEntityCustomer, BeneficialOwner, ControllingPerson, KYCRequirement, DocumentRequirement, PolicyClause, EvidenceSpan |
| Core object properties | owns, controls, beneficialOwnerOf, appliesTo, requires, supportedBy, overrides |
| Core data properties | ownershipPercentage, effectiveDate, expiryDate, jurisdiction, authorityLevel, confidenceScore |
| Disjointness | BeneficialOwner disjointWith AuthorizedSignerOnly; AMLRedFlag disjointWith CreditApprovalReason |
| Reasoning use | Identify candidate document requirement from customer type, jurisdiction, and beneficial owner status |
| SHACL gate | Validate every DocumentRequirement claim has evidence, source authority, effective period, and allowed jurisdiction |
| Eval impact | Golden set sliced by customer type, ownership threshold, jurisdiction, policy version, hard negative terms |
| Owner | KYC Policy Owner + AI Knowledge Architect |
9.3 SHACL Constraint Checklist
| Constraint group | Check | Severity | Release action |
|---|---|---|---|
| Identity | Every concept, source, clause, claim, evidence span has stable URI | Violation | Block ingestion |
| Authority | High-risk claim source authority is A0 or A1 | Violation | Block retrieval and release |
| Version | Source has effective date, expiry or current status, and version | Violation | Block production index |
| Jurisdiction | Policy claim includes approved jurisdiction value | Violation | Route to steward queue |
| Permission | Evidence span has permission group and allowed AI use | Violation | Block prompt inclusion |
| Citation | Critical answer claim has at least one evidence span | Violation | Eval gate fail |
| Domain/range | Relation subject/object match ontology domain and range | Violation | Graph publish fail |
| Disjointness | Node is not assigned mutually exclusive classes | Violation | Ontology/data reconciliation |
| Path semantics | GraphRAG path labels are allowed for the answer type | Violation | Block path explanation |
| Weak signal | Investigation lead is not phrased as confirmed fact | Warning / Violation by context | Require wording repair or HITL |
| Source freshness | Source modified date within freshness SLO | Warning | Stale warning or route downgrade |
| Audit replay | graph, ontology, shapes, source snapshot ids present | Violation | Block release |
9.4 Semantic Drift Log
| Date | Domain | Drift signal | Evidence | Root cause | Decision | Regression added | Owner |
|---|---|---|---|---|---|---|---|
| 2026-06-29 | KYC | Beneficial owner and authorized signer confused in 7 of 120 eval cases | KYC-TERM-004, KYC-TERM-009 | Training FAQ used "account controller" loosely | Added disjoint classes and hard negatives; FAQ downgraded to A4 | KYC-BO-019 through KYC-BO-024 | KYC Steward |
| 2026-06-29 | AML | GraphRAG path described shared device as control relationship | AML-PATH-017 | Relation label mapping collapsed sharesDeviceWith into associatedWith | Restored relation-specific labels and path wording policy | AML-PATH-041 | AML Graph Owner |
| 2026-06-29 | Metrics | AI adoption metric interpreted login as workflow completion | VALUE-METRIC-006 | Metric glossary not synced with semantic layer | Metric contract major version and NL analytics clarification rule | VALUE-METRIC-021 | AI Value Office |
9.5 Portfolio Evidence Pack
| Artifact | 证明的能力 | 面试中怎么讲 |
|---|---|---|
| Semantic Scope Memo | 能划定 AI 语义边界 | 说明为什么只建 KYC / AML / credit 的 ontology slice, 不做全企业大一统 |
| Concept Registry | 能治理业务概念和同义词 | 展示 concept URI、definition、owner、blocked synonyms、eval cases |
| RDF Mapping Profile | 能把事实、证据、来源、版本表达成可交换 graph | 展示 claim-to-evidence lineage 和 named graph strategy |
| Ontology Slice | 能定义概念、关系、不变量和推理边界 | 讲清 class/property、domain/range、disjointness、restriction |
| SHACL Constraint Pack | 能把语义契约变成门禁 | 展示 validation report 如何阻止旧政策、无证据 claim、字段错配进入 AI |
| JSON Schema Boundary Memo | 能区分 payload schema 和 semantic validation | 说明 tool call 为什么需要 JSON Schema + SHACL + policy engine |
| RAG / GraphRAG Semantic Contract | 能把检索和图谱消费治理化 | 展示 source authority、path semantics、hard negatives 和 eval gate |
| Tool / Event Contract | 能控制 agentic action 风险 | 展示工具输入、业务前置条件、事件语义和 audit trail |
| Semantic Drift Log | 能运营语义质量 | 展示漂移信号、根因、治理动作和回归样本 |
| Release Gate Memo | 能做生产级上线判断 | 用 constraint report、eval report、incident runbook 支撑 limited go / no-go |
10. Architecture Review Questions
10.1 Semantic Scope
- 这个 AI capability 属于哪个 bounded context, 是否存在一词多义风险?
- 哪些概念必须使用 stable URI, 哪些只需要普通 metadata?
- 哪些外部 ontology, 如 FIBO, 适合作为 reference, 哪些必须由内部上下文定义?
- 哪些概念不能跨上下文复用, 必须建立 anti-corruption mapping?
10.2 RDF / Graph Facts
- 哪些事实需要表达成 graph facts, 而不是只存在于文档 chunk 或 JSON payload?
- 每个 claim 是否能追到 source、version、evidence span、permission 和 business date?
- named graph 如何表达 source snapshot、environment、jurisdiction、ontology version 和 constraint version?
- 历史 graph 是否保留到足以支持 audit replay?
10.3 OWL / Ontology Semantics
- 哪些 class 和 property 是 AI 风险关键, 需要显式建模?
- 是否定义了 domain/range、disjointness、subclass 和 restriction?
- 推理结果是 production answer 的依据, 还是 investigation lead?
- ontology 变更会影响哪些 RAG metadata、GraphRAG path、tool contract 和 eval cases?
10.4 SHACL / Validation
- 哪些 constraints 是 release blockers, 哪些只是 warning?
- validation 发生在 ingestion、retrieval、tool execution、event processing 还是 eval 阶段?
- SHACL validation report 是否进入 release evidence 和 audit trace?
- constraint failure 的 owner、SLA、exception process 是否明确?
10.5 JSON Schema Boundary
- JSON Schema 验证的是 payload shape 还是业务语义?
- API payload 字段如何映射到 canonical concept URI?
- 字段名相同但语义不同的情况如何处理?
- Tool call 是否在 JSON Schema 通过后还经过 semantic precondition 和 permission check?
10.6 AI Consumption
- RAG retrieved context 是否通过 authority、version、permission 和 semantic constraints?
- GraphRAG path 是否区分 confirmed fact、weak signal、investigation lead 和 recommendation?
- Event-driven agents 是否只消费 approved domain events?
- Eval 是否覆盖 context confusion、field mismatch、metric ambiguity、unsupported claim 和 audit replay?
11. 30-Day Lab: 做一个 Semantic Interoperability Portfolio Case
目标: 30 天内完成一个面向金融零售 AI 的 semantic interoperability 作品集包。推荐选择 KYC Policy Assistant、AML Investigation GraphRAG、Credit Memo Assistant 或 AI Value Office Metrics Assistant。
| Day | 主题 | 实操任务 | 产出 |
|---|---|---|---|
| 1 | Use case framing | 选择一个高语义风险 AI capability, 定义 bounded context、用户、风险等级、禁止动作 | Semantic Scope Memo |
| 2 | Concept inventory | 提取 30-50 个关键业务术语, 标注同义词、禁用混淆词、owner | Concept Registry v1 |
| 3 | Source authority | 建立 source authority 层级和 source registry | Source Authority Matrix |
| 4 | FIBO alignment | 对 Party、Organization、Account、Agreement 等概念做 reference alignment | FIBO Alignment Note |
| 5 | URI strategy | 设计 concept、source、clause、claim、evidence、graph 的 ID 规则 | URI / ID Strategy |
| 6 | RDF mapping | 把 20 个政策 claim 或 graph facts 转成 triple-style mapping | RDF Mapping Profile |
| 7 | Named graph model | 设计 graph version、source snapshot、jurisdiction、permission 的 graph strategy | Named Graph Strategy |
| 8 | Ontology scope | 选 8-12 个核心 class 和 10-15 个 property | Ontology Slice v1 |
| 9 | Domain/range | 为核心关系定义 domain、range、direction、valid time | Property Catalog |
| 10 | Disjointness | 定义容易混淆的互斥概念 | Disjointness Decision Memo |
| 11 | Restriction | 定义关键业务不变量和推理边界 | Restriction Catalog |
| 12 | JSON Schema boundary | 为一个 tool 或 event 设计 JSON Schema | Tool / Event JSON Schema |
| 13 | Semantic mapper | 映射 JSON 字段到 canonical concept 和 graph object | Schema-to-Ontology Mapping |
| 14 | SHACL source shapes | 定义 source、clause、claim、evidence 的 validation intent | Source SHACL Checklist |
| 15 | SHACL graph shapes | 定义 relation、path、domain/range、permission validation | Graph SHACL Checklist |
| 16 | RAG contract | 设计 RAG metadata、authority、version、permission、citation contract | RAG Semantic Contract |
| 17 | GraphRAG contract | 设计 path semantics、edge evidence、weak signal wording | GraphRAG Semantic Contract |
| 18 | Tool contract | 设计 semantic preconditions、blocked actions、audit output | Tool Semantic Contract |
| 19 | Event contract | 设计 domain event fields、source graph、allowed AI use | Event Semantic Contract |
| 20 | Eval contract | 设计 hard negatives、critical failures、claim support rubric | Eval Semantic Contract |
| 21 | Validation report | 生成一份模拟 SHACL validation report 和 release interpretation | Constraint Validation Report |
| 22 | Semantic drift | 设计 drift signals、triage、owner、regression loop | Semantic Drift Log |
| 23 | Audit trace | 设计 answer-to-evidence replay schema | Audit Replay Trace |
| 24 | Release gate | 定义 blockers、warnings、limited go 条件 | Semantic Release Gate Memo |
| 25 | Case study 1 | 完成 KYC 或 AML 端到端案例 | Financial Retail Case Study |
| 26 | Case study 2 | 完成 credit、customer servicing 或 metrics 案例 | Second Case Study |
| 27 | Architecture review | 用 review questions 自评并修订 | Architecture Review Report |
| 28 | Portfolio pack | 整理 scope、registry、mapping、ontology、constraints、eval、release | Portfolio Evidence Pack |
| 29 | Interview story | 准备 30 秒、2 分钟、Enterprise Architect 三种表达 | Interview Narrative |
| 30 | Executive memo | 写给 VP / CTO / Risk 的一页上线建议 | Executive Semantic Risk Memo |
验收标准:
| Capability | 可验证证据 |
|---|---|
| Semantic scope | 能解释为什么选这个 bounded context 和 ontology slice |
| Concept governance | 核心概念有 stable URI、definition、owner、synonyms、blocked confusion terms |
| Graph fact modeling | claim、evidence、source、version、permission 能用 RDF-style mapping 追踪 |
| Ontology semantics | class/property、domain/range、disjointness、restriction 能支持 AI 风险控制 |
| Constraint validation | SHACL checklist 能阻止错配、缺证据、过期、越权和不可回放 |
| Schema boundary | 能清晰区分 JSON Schema、semantic mapper、SHACL、policy engine 的职责 |
| AI consumption | RAG、GraphRAG、tool、event、eval 都有语义契约 |
| Operations | 有 semantic drift log、release gate、incident and audit replay design |
12. 面试表达
12.1 30 秒版本
AI semantic interoperability 是把业务概念、事实、证据、接口和评估统一成可治理语义契约。RDF 负责把事实和关系表达成可交换 graph facts, OWL 负责定义概念和关系的业务语义, SHACL 负责验证这些 graph 是否满足约束。JSON Schema 仍然重要, 但它主要管 API 或 tool payload 的结构; 高风险金融零售 AI 还需要 SHACL 和 ontology 来管证据、权限、版本、上下文和审计。
12.2 2 分钟版本
我会把语义互操作拆成四层。
第一层是 concept identity。金融零售里 "customer"、"risk"、"approval"、"case" 这类词在 AML、信贷、客服、财富里含义不同, 所以要先有 bounded context、concept URI、definition、owner、synonym 和 blocked confusion terms。
第二层是 graph facts。用 RDF-style mapping 把 claim、evidence、source、policy clause、graph path、version、permission 表达出来, 尤其用 named graph 管 source snapshot、effective date、jurisdiction 和 audit replay。
第三层是 ontology semantics。用 OWL-style class、property、domain/range、disjointness 和 restriction 定义业务含义。例如 BeneficialOwner 不是 AuthorizedSignerOnly, AML red flag 不是 credit approval reason。这样 AI 不靠相似文本猜语义。
第四层是 constraint validation。用 SHACL-style shapes 在 ingestion、retrieval、tool execution、event processing 和 eval gate 前做检查。比如 high-risk KYC claim 必须有 A0/A1 source、effective period、jurisdiction、evidence span 和 permission group。缺任何一个都不能进入生产回答。
JSON Schema 我会用于 tool input 和 event payload 的形状验证, 但不会让它替代语义验证。生产级 AI agent 必须先通过 JSON Schema, 再通过 semantic mapper、SHACL、permission 和 policy gate。
12.3 Enterprise Architect 版本
在企业架构视角, semantic interoperability 是 AI control plane 的一部分。它连接业务架构的 bounded context、数据架构的 canonical concepts、应用架构的 API/event/tool contracts、知识架构的 RAG/GraphRAG、治理架构的 eval/release/audit。
我会用 RDF 表达可交换事实和 lineage, 用 OWL 定义核心领域语义和推理边界, 用 SHACL 把语义质量转成可执行门禁, 用 JSON Schema 管接口 payload, 用 FIBO 做金融行业概念参考。落地不是一次性建设全企业 ontology, 而是选择高风险 AI use case 的 ontology slice, 先让 KYC、AML、credit、customer servicing 或 AI metrics 的关键语义可治理、可验证、可审计。
判断标准是: 如果一个 AI capability 无法说明它使用哪些 concept URI、哪些 graph facts、哪些 ontology semantics、哪些 SHACL constraints、哪些 source versions、哪些 eval hard negatives 和哪些 audit trace, 它就不应该进入高风险生产流程。
12.4 高阶追问速答
| 问题 | 回答要点 |
|---|---|
| RDF 和 property graph 如何取舍? | RDF 强在标准化语义交换、IRI、named graph、ontology/SHACL 生态; property graph 强在工程落地和路径查询。高风险互操作和审计用 RDF-style contracts, 运营路径查询可映射到 property graph。 |
| OWL 会不会过度设计? | 会, 如果从全企业大本体开始。我的做法是 ontology slice: 只建与 AI 风险、检索、工具、评估、审计相关的 class/property, 并用 ADR 说明不用 full OWL reasoner 的原因。 |
| SHACL 与数据质量规则有什么区别? | 普通数据质量常看 null、type、range; SHACL 能表达 graph node、relation、class、cardinality、domain/range、evidence、authority、permission 等语义约束, 适合做 AI release gate。 |
| JSON Schema 为什么不够? | 它能验证 payload 形状, 不能证明客户属于 case、claim 有政策证据、source 当前有效、用户有权限、工具动作不越界。这些需要 semantic validation 和 policy gate。 |
| FIBO 怎么用? | FIBO 是金融概念 reference ontology。用它对齐 Party、Account、Agreement 等行业概念, 但内部 KYC、AML、credit 语义仍按 bounded context 和企业政策建模。 |
| 如何减少 hallucination? | 不只靠 prompt, 要把 answer claim 绑定 evidence graph, 检索前后用 source authority、version、permission、SHACL constraints 过滤, eval 中把 unsupported claim 作为 release blocker。 |
| 如何处理 semantic drift? | 监控 eval failure、retrieval mismatch、tool validation failure、metric ambiguity、graph relation error; 每类 drift 绑定 owner、root cause、ontology/schema/constraint update 和 regression case。 |
| 如何证明可审计? | trace 必须包括 user role、business date、graph version、ontology version、shapes version、source version、retrieved evidence、answer claims、tool events 和 validation reports。 |
13. 一页总结
AI semantic interoperability 的核心, 是把 "AI 理解业务" 从 prompt 期望升级为可治理工程契约。
| 从 | 到 |
|---|---|
| 字段名相似 | concept URI + semantic mapping |
| 文档 chunk 相似 | source authority + evidence span + claim graph |
| API schema 通过 | JSON Schema + semantic precondition + permission gate |
| 图谱路径漂亮 | relation semantics + evidence path + human review boundary |
| 本体很大 | ontology slice + risk-based constraints |
| 数据质量检查 | SHACL semantic validation + release gate |
| 平均回答准确率 | failure taxonomy + hard negatives + claim support |
| 事后查日志 | answer-to-evidence audit replay |
作品集表达可以浓缩成一句话:
I design AI semantic contracts with RDF-style facts, OWL-style meaning and SHACL-style validation, so financial retail AI systems can exchange knowledge without confusing concepts, mismatching fields, inventing metrics or losing audit traceability.