返回 Papers
AI 扩展计划 / Playbooks

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。

785AI_SEMANTIC_INTEROPERABILITY_RDF_OWL_SHACL_PLAYBOOK.md

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。

AnchorOfficial / primary source本手册使用方式
W3C RDFhttps://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 OWLhttps://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 SHACLhttps://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 Schemahttps://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。
FIBOhttps://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
RDFSemantic 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。"
OWLOntology 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。"
SHACLConstraint 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 SchemaAPI 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。"
FIBOFinancial 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 groundingAI 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 exchangeRAG、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 SHACLJSON 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 contextbounded context, owner, risk tier这个语义契约属于哪个业务上下文Semantic Scope Memo
2. Concept identityconcept 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 factssubject, 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 semanticsclass, property, domain/range, restriction哪些概念关系和不变量需要显式表达Ontology Slice
5. Constraint validationshapes 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 contractsJSON Schema, API/event/tool schemaAI 工具和事件如何保证结构正确JSON Schema Contract
7. AI consumptionRAG, GraphRAG, tool, event, metric, evalAI 能如何消费、组合和输出语义对象AI Semantic Contract
8. Operationsdrift, 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 capabilityAI 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 snapshotgraph: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
Conceptconcept:kyc.beneficial_ownerlabel, definition, context, owner, synonyms, source
Sourcesrc:kyc-policy-v6-1authority, approval status, version, effective date, permission
Clauseclause:kyc-v6-1-3-4source, section, page, paragraph, validity
Claimclaim:kyc-doc-required-014statement, claim type, support status, risk tier
Evidence spanevidence:kyc-v6-1-p12-r3document, offset/page/table row, extraction confidence
Graphgraph:kyc-prod-2026-06-15graph 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 decisionAI 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 / SOPSKOS-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 policyLightweight 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 graphDomain ontology + entity relation model + graph path semantics多跳关系和证据路径是分析核心
Credit decision supportOntology slice + business rule engine + HITL需保护信贷决策边界, 不让 LLM 直接做审批
Enterprise financial ontologyFIBO 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 capabilityAI 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 counthigh-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、infopermission missing 是 violation, stale glossary 是 warning
Validation report形成 release gate 和 audit evidenceingestion 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 gatesource, clause, metadata, evidence span不进入 production index 或 graph
Ontology release gateclass/property changes, mapping changes架构评审和回归评估
Retrieval context gateretrieved chunks / graph facts低权威、过期、无权限证据不得进入 prompt
Tool input gatetool semantic object before execution阻止跨客户、跨权限、越权动作
Event processing gatedomain event graph事件缺少 aggregate、actor、effective time 时进入 exception queue
Eval gateanswer claim, citation, graph pathcritical 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
字段是否必填、类型是否正确可做, 但不是最佳入口
API / tool input 的结构契约用于补充语义验证
事件 envelope 是否稳定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
跨节点关系是否一致
同一 concept 是否来自批准词表可枚举, 维护成本高
claim 是否由 evidence span 支撑
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

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 statusFAQ 覆盖正式政策
Concept metadatachunk 绑定哪些 concept URI、bounded context、jurisdiction、product召回相似但错误上下文
Version validityeffective date, expiry date, business date filter引用过期或未生效政策
Claim extractionchunk 支撑哪些 claim, claim type, risk tieranswer 有 citation 但 claim 不被证据支持
Permissionrole, purpose, case binding, citation visibility检索到用户无权查看的 investigation note
Constraint gateingestion 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 identityentity resolution 规则、confidence、merge/split policy把两个同名客户或商户错误合并
Relation semanticsrelation 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
Evaluationpath precision、entity resolution accuracy、unsupported relation rateGraphRAG 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 modecreate_kyc_remediation_task 只能服务 KYC remediation
Input schemaJSON Schema 管字段形状case id, party id, reason code, evidence claim id
Semantic preconditionsSHACL / policy engine 管业务前置条件customer belongs to case, claim supported by active policy
Permissionactor role, purpose, entitlement, approval requirementbranch staff 不能创建 AML escalation
Output contract工具返回哪些 domain event 或 graph factsKYCRemediationTaskCreated 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 aggregatecase: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 snapshotgraph: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 claimclaim:kyc-doc-required-014
allowed_ai_useAI 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
Queryquery intent, bounded context, role, business date歧义时必须 clarify
Gold evidencesource 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 behavioranswer, 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 spanunsupported critical claim = fail
Semantic failure taxonomycontext 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 riskUBO、beneficial owner、controller、authorized signer、customer、party 在系统和政策中语义不同
RDF graph factscustomer party, legal entity, beneficial owner candidate, document requirement, policy clause, evidence span
OWL semanticsBeneficialOwner, ControllingPerson, LegalEntityCustomer, DocumentRequirement, EnhancedDueDiligenceRequirement
SHACL constraintshigh-risk KYC claim 必须有 approved policy source、effective period、jurisdiction、evidence span
JSON Schema roleKYC 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 slicescustomer type, jurisdiction, policy version, UBO threshold, document type, branch vs compliance role
Portfolio evidencesemantic 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 riskshared address、shared device、beneficial ownership、transaction counterparty、screening hit 的风险含义不同
RDF graph factscustomer-account-transaction-counterparty-device-address-case 关系, 每条边绑定来源和 valid time
OWL semanticsAMLCase, Alert, Transaction, Counterparty, RedFlag, TypologyIndicator, EvidencePath
SHACL constraintsGraphRAG path 必须区分 direct evidence、weak signal、investigation lead, 不得把弱关联当结论
JSON Schema rolealert 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 slicespath precision, entity resolution, missing red flag, unsupported relation, typology mismatch
Portfolio evidenceGraphRAG 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 riskapproval condition、adverse action、reason code、exception、policy override 具有高影响决策含义
RDF graph factsapplication, borrower, document, credit memo claim, policy clause, condition, human decision
OWL semanticsLoanApplication, CreditPolicyClause, ApprovalCondition, AdverseActionReason, ExceptionReview
SHACL constraintsAI-generated credit memo claim 必须绑定 source evidence, 不得产生 automated approval / decline claim
JSON Schema rolememo 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 slicespolicy version, product, jurisdiction, protected attribute exclusion, reason code consistency
Portfolio evidencedecision 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 riskcomplaint、inquiry、dispute、escalation、service recovery、fee waiver 的业务含义和监管后果不同
RDF graph factsservice case, interaction, product term, complaint category, escalation reason, response claim
OWL semanticsServiceCase, Complaint, FeeDispute, ProductTerm, EscalationRequirement, CustomerVulnerability
SHACL constraintscustomer-facing response claim 必须区分 explanation、commitment、escalation; fee waiver commitment 需要 human approval
JSON Schema roleresponse 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 slicesvulnerable customer, product line, channel, complaint category, false promise, escalation correctness
Portfolio evidencecustomer-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 riskadoption、usage、deflection、containment、resolution、ROI、risk reduction 的指标口径容易混用
RDF graph factsmetric, measure, dimension, owner, formula, data contract, dashboard, eval report, release decision
OWL semanticsMetric, BusinessOutcome, EvalMetric, WorkflowMetric, GuardrailMetric, AIUseCase
SHACL constraintsofficial AI metric 必须有 owner、formula、grain、allowed dimensions、lineage、quality SLO
JSON Schema rolemetric API request/response, portfolio dashboard event
AI controlsNL analytics 只能解释 approved metrics, 歧义指标必须追问
Eval slicesmetric ambiguity, restricted slice, causal overclaim, stale metric definition
Portfolio evidencemetric 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

ActivityProduct / BADomain OwnerArchitectureKnowledge / Data StewardRisk / ComplianceEvalOps
Define semantic scopeARRCCC
Approve concept definitionRACRCC
Design ontology sliceCRARCC
Align FIBO / external ontologyCCARCC
Define SHACL constraintsRCARCR
Approve source authorityCACRRC
Validate tool/event semanticsRCARRR
Maintain eval semantic taxonomyARCCRR
Review semantic driftRARRCR
Decide release gateARRCRR

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 changebeneficial owner threshold 或适用范围变化domain owner approval, ontology version, eval regression
Relationship changecontrols relation 新增 indirect control pathontology ADR, SHACL update, GraphRAG path regression
Source authority changeFAQ 降级, policy 新版本生效source registry update, RAG index refresh, version tests
Tool semantic changecreate task tool 新增 reason codeJSON Schema version, SHACL shape update, permission review
Metric semantic changeadoption metric 从 login 改为 completed workflowmetric contract major version, consumer impact analysis
FIBO alignment changeinternal Party 映射到 FIBO party conceptalignment 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 indexsource deprecation, synonym review, retrieval regression
Tool validation failure 增加API schema 变化未同步 semantic mapperschema-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

ElementExample
Slice namekyc-cdd-requirement-slice
ScopeKYC policy explanation, document requirement reasoning, remediation task support
Out of scopeFinal KYC approval, customer risk rating override, automated account opening
Reference ontologyFIBO for Party / Organization reference; internal KYC ontology for policy-specific obligations
Core classesParty, NaturalPerson, LegalEntityCustomer, BeneficialOwner, ControllingPerson, KYCRequirement, DocumentRequirement, PolicyClause, EvidenceSpan
Core object propertiesowns, controls, beneficialOwnerOf, appliesTo, requires, supportedBy, overrides
Core data propertiesownershipPercentage, effectiveDate, expiryDate, jurisdiction, authorityLevel, confidenceScore
DisjointnessBeneficialOwner disjointWith AuthorizedSignerOnly; AMLRedFlag disjointWith CreditApprovalReason
Reasoning useIdentify candidate document requirement from customer type, jurisdiction, and beneficial owner status
SHACL gateValidate every DocumentRequirement claim has evidence, source authority, effective period, and allowed jurisdiction
Eval impactGolden set sliced by customer type, ownership threshold, jurisdiction, policy version, hard negative terms
OwnerKYC Policy Owner + AI Knowledge Architect

9.3 SHACL Constraint Checklist

Constraint groupCheckSeverityRelease action
IdentityEvery concept, source, clause, claim, evidence span has stable URIViolationBlock ingestion
AuthorityHigh-risk claim source authority is A0 or A1ViolationBlock retrieval and release
VersionSource has effective date, expiry or current status, and versionViolationBlock production index
JurisdictionPolicy claim includes approved jurisdiction valueViolationRoute to steward queue
PermissionEvidence span has permission group and allowed AI useViolationBlock prompt inclusion
CitationCritical answer claim has at least one evidence spanViolationEval gate fail
Domain/rangeRelation subject/object match ontology domain and rangeViolationGraph publish fail
DisjointnessNode is not assigned mutually exclusive classesViolationOntology/data reconciliation
Path semanticsGraphRAG path labels are allowed for the answer typeViolationBlock path explanation
Weak signalInvestigation lead is not phrased as confirmed factWarning / Violation by contextRequire wording repair or HITL
Source freshnessSource modified date within freshness SLOWarningStale warning or route downgrade
Audit replaygraph, ontology, shapes, source snapshot ids presentViolationBlock release

9.4 Semantic Drift Log

DateDomainDrift signalEvidenceRoot causeDecisionRegression addedOwner
2026-06-29KYCBeneficial owner and authorized signer confused in 7 of 120 eval casesKYC-TERM-004, KYC-TERM-009Training FAQ used "account controller" looselyAdded disjoint classes and hard negatives; FAQ downgraded to A4KYC-BO-019 through KYC-BO-024KYC Steward
2026-06-29AMLGraphRAG path described shared device as control relationshipAML-PATH-017Relation label mapping collapsed sharesDeviceWith into associatedWithRestored relation-specific labels and path wording policyAML-PATH-041AML Graph Owner
2026-06-29MetricsAI adoption metric interpreted login as workflow completionVALUE-METRIC-006Metric glossary not synced with semantic layerMetric contract major version and NL analytics clarification ruleVALUE-METRIC-021AI 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
1Use 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
2Concept inventory提取 30-50 个关键业务术语, 标注同义词、禁用混淆词、ownerConcept Registry v1
3Source authority建立 source authority 层级和 source registrySource Authority Matrix
4FIBO alignment对 Party、Organization、Account、Agreement 等概念做 reference alignmentFIBO Alignment Note
5URI strategy设计 concept、source、clause、claim、evidence、graph 的 ID 规则URI / ID Strategy
6RDF mapping把 20 个政策 claim 或 graph facts 转成 triple-style mappingRDF Mapping Profile
7Named graph model设计 graph version、source snapshot、jurisdiction、permission 的 graph strategyNamed Graph Strategy
8Ontology scope选 8-12 个核心 class 和 10-15 个 propertyOntology Slice v1
9Domain/range为核心关系定义 domain、range、direction、valid timeProperty Catalog
10Disjointness定义容易混淆的互斥概念Disjointness Decision Memo
11Restriction定义关键业务不变量和推理边界Restriction Catalog
12JSON Schema boundary为一个 tool 或 event 设计 JSON SchemaTool / Event JSON Schema
13Semantic mapper映射 JSON 字段到 canonical concept 和 graph objectSchema-to-Ontology Mapping
14SHACL source shapes定义 source、clause、claim、evidence 的 validation intentSource SHACL Checklist
15SHACL graph shapes定义 relation、path、domain/range、permission validationGraph SHACL Checklist
16RAG contract设计 RAG metadata、authority、version、permission、citation contractRAG Semantic Contract
17GraphRAG contract设计 path semantics、edge evidence、weak signal wordingGraphRAG Semantic Contract
18Tool contract设计 semantic preconditions、blocked actions、audit outputTool Semantic Contract
19Event contract设计 domain event fields、source graph、allowed AI useEvent Semantic Contract
20Eval contract设计 hard negatives、critical failures、claim support rubricEval Semantic Contract
21Validation report生成一份模拟 SHACL validation report 和 release interpretationConstraint Validation Report
22Semantic drift设计 drift signals、triage、owner、regression loopSemantic Drift Log
23Audit trace设计 answer-to-evidence replay schemaAudit Replay Trace
24Release gate定义 blockers、warnings、limited go 条件Semantic Release Gate Memo
25Case study 1完成 KYC 或 AML 端到端案例Financial Retail Case Study
26Case study 2完成 credit、customer servicing 或 metrics 案例Second Case Study
27Architecture review用 review questions 自评并修订Architecture Review Report
28Portfolio pack整理 scope、registry、mapping、ontology、constraints、eval、releasePortfolio Evidence Pack
29Interview story准备 30 秒、2 分钟、Enterprise Architect 三种表达Interview Narrative
30Executive 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 modelingclaim、evidence、source、version、permission 能用 RDF-style mapping 追踪
Ontology semanticsclass/property、domain/range、disjointness、restriction 能支持 AI 风险控制
Constraint validationSHACL 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 consumptionRAG、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。

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.