Ontology in Production: 14 Chinese Enterprises Share Hard-Won Lessons

Why Ontology Deployment Remains Rare in China

The article opens with five core reasons why ontology (本体) has been discussed for over a decade in Chinese tech circles but few enterprises have successfully run it in production:

No unified standard: Palantir, Snowflake, Databricks, Apache Ossie, and Agent Memory each take different approaches — differences lie not in whether they use graphs, but in how knowledge is represented, executed, produced, and validated.

Tendency to become "big and comprehensive": Enterprise-wide ontologies built in one go become obsolete fastest; they bloat, break, and lose value when business changes.

Irreversible implementation path: Choosing between a semantic mapping layer (fast to launch, poor multi-hop performance) and a materialized object store (fast queries, but faces sync, consistency, cost, and "source of truth" challenges) is largely a one-way decision.

Definition conflicts: Business lines define metrics like "active user" differently; forced unification distorts understanding, while lack of unification leaves agents confused.

Depreciation after delivery: Without continuous iteration mechanisms, knowledge assets start depreciating from day one.

Section 1: Enterprises with Ontology Already Running in Production (7 Talks)

01 Ant Group: Financial Data Knowledge Ontology — From "Querying Data" to "Understanding Business"

Speaker: Cui Weimin, Senior Data Technology Expert, Ant Group.

Core problem: Business lines defined "active user" differently. Ant's solution: retain differences in the ontology, let agents fetch on demand.

Architecture:

LangGraph 6-node state machine: InputParser → Comprehender → EntityResolver → Reconciler → Synthesizer → Verifier — automates ontology synthesis and validation.

SQL Understander (sqlglot + regex): Reverse-extracts field semantics and cross-table relationships from historical task code, creating a "code-as-annotation" auto-generation pipeline.

Reconciler: 7-level cost-progressive fusion for cross-table same-name fields (auto-merge → human confirmation), core principle: flag conflicts but don't force merge.

Verifier: Pydantic + business rules + deep-pit detection; failures trigger L4 retry or downgrade warning.

Service layer — mutual-index architecture: Synchronously retrieves three knowledge types: KG_cs (business-view rules, e.g., Huabei's "active user" definition), KG_fr (entity-relation graph, e.g., "User A — holds — Huabei limit 50k"), RC (raw document evidence). Results fused, deduplicated, packaged as KnowledgePackage.

Results: Cross-opportunity identification reduced from 3–5 days to hour-level; anomaly attribution accuracy >85%; supports multiple marketing strategies in production.

"Ontology is not a dictionary; it's the textbook that lets the system 'understand business'."

02 Li Auto: Ontology-Driven Agent — From Query & Analysis to Sales Decisions

Speaker: Liang Wei, Senior Algorithm Engineer, Li Auto.

Background: Post-training achieved near 100% NL2SQL accuracy, but faced long knowledge-prep/training cycles, difficulty extending to new domains, and brittleness when business definitions changed.

Approach: Re-organized enterprise data with ontology so objects, relations, metrics, and business rules become semantic structures agents can understand and query.

Evolution path: NL2SQL → MCP → Skill → built-in Agent — four iterations with trade-offs. Sales decision support built on three layers: business fact layer, sales cognition layer, scenario execution layer.

03 Li Auto: Semantic Layer as Context — Production-Grade Agent Semantic Engineering

Speaker: Qian Han, Big Data Platform Lead, Li Auto.

Thesis: Production agent bottleneck is not model capability but context quality — whether agents get fragmented info scattered in prompts/tools/logs or semantically governed, traceable, structured context.

Case: Spark/Flink fault diagnosis. Knowledge split into four layers: Ontology, OSI, Playbook, General Execution Layer. Agents cross-reference logs, metrics, data warehouse to produce evidence-backed root-cause judgments; diagnostic gaps become versioned, verifiable, regressible semantic assets.

04 Tianyi Payment: Ontology Deployment Fork — Semantic Mapping vs. Object Storage

Speaker: Xu Dehua, Risk Management Director, Tianyi Payment.

Core dilemma: Is ontology a "translation layer" or a "data store"?

Mapping only: Metadata layer, translates to warehouse SQL at query time — no data movement, fast launch, no redundancy; but multi-hop traversal fails, definitions bound to underlying table structures, no place to land object state/execution results.

Materialized object store: Instances, attributes, relationships physically stored — fast queries, supports state & decision closure; but requires real-time sync, incremental consistency, storage cost, and "who is the truth" resolution.

Decision: Basically irreversible. Upstream issue: concept definitions and real business data shouldn't be mixed in one layer.

Solution: Hybrid — core objects materialized, long-tail kept as mapping; boundary fixed in design, not ad-hoc. Four selection criteria: query pattern, latency needs, need for state/write-back, change frequency.

Agent-facing data supply: Minimum necessary context, governed access interfaces, fallback conditions, risk quantification.

Cross-domain reuse & acceptance: Five-piece delivery per business domain: ontology fragment, rule library, case library, verified queries, evaluation set.

05 Shopee: Graph-Platform-Based Anti-Fraud Ontology Modeling & Intelligent Decisioning

Speaker: Zhang Songqing, Graph Platform Lead, Shopee Data Infra.

Problem: Traditional tabular models struggle with coordinated disguise across accounts, devices, transactions in e-commerce promotions and credit.

Platform: One-stop graph platform — graph DB + graph compute + graph learning synergy. Objects, relations, risk rules modeled as ontology → dynamic anti-fraud knowledge network.

Two real scenarios:

Monee Credit user relationship graph & credit risk control: Pre-/in-loan identification uplift; graph DB enables real-time eligibility approval for online loan applications.

Shopee order promotion knowledge graph & fraud detection: Graph compute + RGAT (Relational Graph Attention Network) identifies black-industry syndicates and anomalous transaction chains.

Engineering challenges: Large-scale heterogeneous relations, real-time updates, feature serving, training efficiency; solved with feature pre-fetch and batch-processing optimization.

06 Runhe Software: Domain Knowledge Engineering with Ontology — Financial Testing Case

Speaker: Yang Dongrui, AI Architect, Runhe Software.

Pain points: LLM direct generation suffers: methods not reusable, results not traceable, quality not evaluable.

Solution: Test Ontology + Test IR (Intermediate Representation). Business ontology precipitates testing methodology; Test IR connects requirements, outlines, cases. Agent handles semantic understanding & method selection; tools handle data calc, case assembly, validation — forming traceable chain: requirement parsing → IR → method application → case generation.

Four-route comparison: Direct LLM (fast but black-box), fixed workflow/prompt (single-scene but tight coupling), RAG knowledge retrieval (supplements knowledge but can't express method applicability constraints & output constraints), Ontology + Test IR (structured, suitable for long-term governance).

07 360: From "Querying Data" to "Understanding Business" — Progressive Semantic Layer Evolution & Quantitative Evaluation

Speaker: Guo Chaoyang, Lakehouse Data Agent Evolution Tech Lead, 360.

Gap: LLM writes SQL at 85% on academic benchmarks; in real warehouse (hundreds of tables, 10k+ fields, tribal definitions) accuracy collapses, and errors are opaque.

Strategy: "Half-finished cold start, grow while running"

Dual-path hybrid routing: Semantically complete → deterministic compilation engine (AST main path); only tables/docs → RAG fallback. Every result carries confidence score: ≥0.95 auto-enters reports; ≤0.65 system prompts "low confidence, suggest human review".

Thin API → Thick Skill → Ontology metadata: High-frequency usage patterns reverse-precipitate as semantic assets; Skills eventually degrade to pure orchestrators, knowledge fully provided by ontology layer.

Three-level evaluation + feedback flywheel: L1 programmatic result-set comparison → L2 LLM Judge semantic judgment → L3 human fallback. User thumbs-up (after anti-pollution filtering) auto-sampled into regression test set.

Current scope: Internal big-data cluster ops, S3/PoleFS storage ops metric analysis & attribution; Ops data-center ops, document search, membership middle-platform across departments.

Section 2: Ontology & Semantic Layer — Engineering Methodology & Selection Map (5 Talks)

08 Yuepoint Tech: Next Stop for Semantic Layer Is Business Ontology — 12 Years, ~100 Enterprises, Gains & Losses

Speaker: Ren Xinqi, Founder/CEO, Yuepoint Tech.

Key distinction: Domain ontology ≠ another name for warehouse semantic layer; it's an independent enterprise business world model — organizing business objects, relations, rules, executable actions.

Pitfalls observed: Terminology & metric definition inconsistency; "big and comprehensive" ontologies obsolete first; business/algorithm/data tri-party cognitive misalignment.

Methodology: Scenario-first lightweight ontology modeling; LLM-assisted knowledge extraction + expert review closed-loop iteration.

Ontology constraining LLMs: Concept verification & generation constraints across RAG & Agent full chain — unify definitions, reduce hallucination.

Case: Manufacturing quality traceability — cross CRM/MES/ERP/PLM trace from multi-person days to one person 5 minutes; yield improved 5%.

09 Datastrato: Open Semantics, Ontology & Metadata — Three Pillars of Agent Context

Speaker: Du Junping, Founder/CEO, Datastrato; Apache Gravitino initiator.

Three-layer progression:

Unified metadata: Apache Gravitino cross-cloud, cross-engine, multi-modal unified metadata management — consistent description & access interface for data assets.

Open semantic layer: Standardize metrics, dimensions, business definitions — eliminate "one metric, multiple definitions".

Ontology: Model business entities, relations, rules — enable agents to precisely map natural-language business intent to real data.

Two context-building strategies compared: Top-down (task-driven, fast but creates context islands) vs. Bottom-up (metadata-governed, solid foundation but can't infer business meaning). Sustainable hybrid: business-scenario-driven modeling + unified metadata connecting business concepts, data assets, governance rules + continuous iteration via business validation & agent feedback.

10 Cloud器 (CloudQi): No Unified Ontology Standard — Five Routes, Which to Follow?

Speaker: Guan Tao, Co-founder & CTO, Cloud器.

Five mainstream routes: Palantir, Snowflake, Databricks, Apache Ossie, Agent Memory. Real difference not "graph or not" but how knowledge is represented, executed, produced, validated.

From DIKW: Deconstruct knowledge & ontology evolution; compare five routes; give enterprise path from data engineering to knowledge engineering.

11 Dongchedi (懂车帝): AI-Native Data Semantic Platform — Ontology as Strong-Constraint Knowledge Layer

Speaker: Miao Zhiyong, Data Warehouse Lead, Dongchedi.

Problem: LLM direct answers on business data → hallucination, instability, low confidence. Text2SQL relying solely on model training or text recall can't guarantee accuracy/consistency.

Semantic layer value: Strong-constraint knowledge between LLM and data assets. Structurally manages metrics, dimensions, entities, business processes, physical implementations, SQL rules — lets AI understand business, generate SQL, explain results within explicit boundaries. vs. single text KB, RAG vector recall, full metadata recall: semantic layer has clearer query constraints, explainability, lower inference cost.

Platformization solves: Stale definitions, physical table updates, local version inconsistency, scattered knowledge maintenance — platform as single source of truth, continuously manages semantic definitions, versions, mappings, consumption governance.

12 Yiche (易车): From DataAgent to Action Closure — Ontology as One of Three Foundational Layers

Speaker: Wang Linhong, Data Platform Lead, Yiche.

Evolution: ChatBI → DataAgent → DataWork. Three-layer foundation covering trusted data, semantics, ontology & knowledge, MCP & Harness. Query evolution + proactive follow-through connect analysis, action, result verification — forming enterprise data intelligence progression from trusted querying → assisted analysis → controlled execution.

Section 3: Frontier Exploration & Cross-Industry Extension (2 Talks)

13 Renmin University: Self-Evolving Data Agents — Data–Ontology–Agent Co-Evolution

Speaker: Zhang Shaolei, Assistant Professor, RUC School of Information.

Problem: Existing data agents face alignment gaps among business intent, data semantics, tool execution on heterogeneous data (tables, databases, docs). Fixed workflows, RAG, static KGs lack continuous adaptation to open tasks & dynamic data.

Framework: Ontology as executable semantic middle layer — unifies domain concepts, tool capabilities, data schemas. Based on agent interaction trajectories: failure attribution → knowledge correction → pairwise evaluation gating.

Key insight: Typical agent errors stem not from model hallucination but from data object identification, field semantic mapping, tool selection errors.

Three-layer ontology: Content Layer (business concepts/metric definitions/entity relations), Tool Layer (tool capabilities & constraints), Schema Layer (logical concept → physical table/field mapping).

Evolution loop: Interaction trajectory failure attribution → candidate modification generation → hold-out validation pairwise evaluation gating → versioned release / reject rollback. Forms continuous loop: "data exposes problem → agent produces trajectory → ontology precipitates knowledge → agent re-validates".

Status: Running 6 months; 2000+ registered users, ~400 deep users; 300+ universities/institutes requested trial.

14 Global Luxury Retailer × Shinx (数新智能): Apache Ossie Ontology Spec & Semantic Enhancement

Speakers: Tan Zong (APAC Big Data Lead, global top luxury retailer) × Yuan Panfeng (CTO, Shinx).

Context: LLMs evolving from single-point Copilot to Agent paradigm; enterprise data scattered across multi-cloud multi-brand, strict compliance, fragmented AI apps — agents "no data available".

Two-part talk: Business perspective (pain points: data silos, compliance walls, fragmented AI limits, Agent-Ready data foundation demands) + Technical perspective (multi-cloud AI-native platform DataAgent architecture, enterprise semantic layer, sandbox security, closed-loop iteration).

Results: End-to-end data delivery efficiency +70%; Agent task diagnosis accuracy +50%; via semantic knowledge base, code generation accuracy from 30% to 90%.

Engineering highlights: Apache Ossie ontology spec, CodeWiki semantic enhancement to补齐 business knowledge gaps, multi-version compatibility/fault-tolerance.

Closing Synthesis

The article concludes that ontology deployment in China is scarce because it demands a chain of upfront judgments: separate concept vs. instance layers? mapping vs. object store? force-unify definitions or preserve differences? one-off delivery or evolving infrastructure? how to prove superiority over prior approach? No standard answers — only trade-offs made under each enterprise's constraints. The value lies in those trade-offs and their costs. The 14 talks at DACon 2026 Beijing (Oct 23–24, Hilton Beijing) represent the densest public sharing to date of Chinese enterprises' ontology production experiences.