12 Chinese Tech Giants Share Production Semantic Layer Architectures for AI Data Agents

Conference Overview: DACon 2026 Beijing

On October 23-24, 2026, DACon 2026 Beijing hosted 12 sessions from companies that have implemented semantic layers at scale. The central theme: large models achieve 85% SQL accuracy on academic benchmarks but plummet to ~50% in real warehouses with hundreds of tables, thousands of columns, and tribal knowledge. The solution is a semantic layer — a "strong constraint knowledge" layer between LLMs and data assets.

Four Universal Hurdles

Semantic maintenance: Definitions expire, physical tables change, local versions diverge, knowledge scattered.

Execution reliability: LLMs generate SQL that fails silently; errors are undetectable without deterministic compilation.

Permission inheritance: Metadata visibility ≠ data access; most systems only handle the former.

Effect quantification: Academic benchmarks compare SQL text, not query results; they ignore business definitions, multi-turn dialogue, and hallucination detection.

Company Case Studies

1. 懂车帝 — AI-Native Semantic Platform

Speaker: 苗治勇 (Data Warehouse Lead). Defines semantic layer as "strong constraint knowledge" structuring metrics, dimensions, entities, business processes, physical mappings, and SQL rules. Compared to alternatives: vs. pure text knowledge base — clearer query constraints, better explainability, lower inference cost; vs. RAG vector retrieval — defines rules not similarity; vs. full metadata dump — prevents AI guessing across thousands of tables. Platform solves definition expiration, table changes, version inconsistency, scattered maintenance via unified fact source.

2. 知乎 — Three-Stage Evolution: MetricFlow → Knowledge Base → Lightweight Knowledge Graph

Speaker: 汤晋瑄 (Data Platform Engineer). Stage 1: dbt MetricFlow + BM25 retrieval. SQL generated by MetricFlow engine — standardized, type-safe, accurate. Limitations: complex, high cost for special metrics, inflexible, cannot leverage LLM fully, high config maintenance. Stage 2: LLM auto-generates metric definitions stored as Markdown; closed loop: LLM generate → knowledge base ingest → enhanced retrieval → LLM use. Scale: 1,400+ metrics, 300+ wide/certified tables, 25,000+ SQL templates, business docs. Stage 3: Knowledge graph for multi-hop relations. AI-native, lightweight: AI-led construction, maintenance, querying; no heavy graph compute. Core tech: custom chunking (per metric, layered tables, per SQL statement, fallback), hybrid retrieval (full-text + vector + WRRF fusion + name matching), hierarchical matching (table → field), business priority (core wide tables > certified tables). Result: team penetration >65%, simple queries 100% satisfied, token consumption drastically reduced.

3. 蚂蚁集团 — Apache Ossie Semantic Layer: 100+ Models, 5,000+ Metrics

Speaker: 王小军 (Semantic Platform Lead). Problem: Data Agents write SQL but don't understand business. Solution: Open-source OSI protocol with "base standard + business extensions". Asset platform hosts single semantic source of truth (SSOT), unifying models, lineage, consumption — "define once, consume everywhere". Converts tables, BI models, queries into traceable human+machine co-built models. Forward dev pipeline + lineage monitoring for freshness. Subgraph retrieval, confidence scoring, ranking for agent consumption. Current: 100+ semantic models, 5,000+ metrics, multi-scenario consumption (human, Agent, platform). Challenges: OSI native standard lacks enterprise features (multi-grain metrics, equivalent metrics, cross-model refs) — solved by controlled extensions; model drift — solved by full-chain monitoring (feedback, patrol, change analysis) + forward dev pipeline embedding semantic iteration into data dev lifecycle; consumption accuracy at scale — solved by "hybrid retrieval + progressive disclosure" (keyword+vector+tags+graph, confidence scoring, re-ranking, return task-specific subgraphs).

4. 理想汽车 — Semantic Layer as Context for Production Agents

Speaker: 钱瀚 (Big Data Platform Lead). Thesis: Production Agent bottleneck is context quality, not model capability. Uses Spark/Flink fault diagnosis example. Splits scattered Prompt/tool knowledge into Ontology, OSI, Playbook, generic execution layer. Agent cross-references logs, metrics, warehouse for evidence-based root cause. Diagnoses gaps become versioned, verifiable, regressible semantic assets. Discusses lightweight modeling, query constraints, exploration vs. determinism trade-offs. Challenges: manual ontology/playbook effort limits cross-business replication; need standardized auto-generation. Future: permission control, execution tracing, evidence status, result validation, multi-environment adaptation, governable/observable/extensible Agent control plane.

5. Datastrato — Three Pillars of Agent Context: Unified Metadata, Open Semantic Layer, Ontology

Speaker: 堵俊平 (Founder/CEO, Apache Gravitino PMC). Layered design: Unified Metadata (Gravitino: cross-cloud, cross-engine, multimodal); Open Semantic Layer (standardized metrics, dimensions, business definitions — eliminate multiple calibers); Ontology (business entities, relationships, rules — map natural language intent to real data). Layers build progressively. Two build paths: top-down (app-driven) fast value but creates context silos; bottom-up (metadata-driven) broad coverage but cannot derive business meaning from technical structure alone. Sustainable hybrid: business scenarios drive modeling, unified metadata connects concepts/assets/governance, continuous iteration via business validation and agent feedback. Challenges: fragmented semantic modeling across apps; business meaning not auto-derivable from tech structure; lack of standards for multimodal/cross-cloud/cross-engine metadata; context freshness — solved by unified metadata base + open semantic layer + ontology rules/boundaries + scenario-driven modeling + feedback iteration.

6. 高德 (Talk A) — Semantic Governance, Attribution Diagnosis, Management Decision

Speaker: 钟雨洁 (AI Application Data Analyst). Local life business: 20+ industries, vastly different models/calibers/focus. Manual analysis slow, experience not reusable. Three-layer Agent architecture: Semantic Governance unifies business calibers/knowledge; Attribution Diagnosis locates metric changes and business root causes; Management Decision converts conclusions to actions. Knowledge architecture: Edit-Compile-Publish-Consume. Edit layer manages industry knowledge, metric rules, operational experience. Compile layer transforms to business entities, strong-typed relations, atomic knowledge statements. Publish layer generates immutable versions with content hashes. Consume layer serves unified interface to analysis/decision Agents. Query pipeline: capability catalog → identity resolution → relation retrieval → task routing → knowledge assembly; ambiguity, missing, expired, truncated explicitly returned. Results: ~66,000 traceable knowledge statements, 198 consumption paths (186 direct-service). Key division: engineering system guards numbers, identities, scopes, versions; model handles business interpretation, synthesis, natural language. Numbers and attribution computed by deterministic engine; model explains within fixed facts/evidence boundaries. Reports closed via structured contracts, execution receipts, content fingerprints, file hashes; Guard intercepts missing signals, unsourced numbers, causal leaps. Efficiency up 64x, business acceptance 100%.

7. 高德 (Talk B) — Real Architecture Migration: From "Answering Right" to "Seeing Through"

Speaker: 林宇航 (Data Product Expert). Pain: dashboards never finished, long-tail uncovered; traditional query only answers "how much", not "why/what now". Did not stack larger models; transformed multi-year BI base into routable, domain-knowledge-loadable Agent base. Real migration in two months: MVP 39 tables → 89.4% accuracy; full 331 tables → end-to-end accuracy crashed to 50%; switched from RAG+Workflow to Skill architecture → recovered to 95%. Scale challenge: single-domain MVP works, but full-scale exposes RAG+Workflow flaws — one-shot knowledge retrieval, multi-step info loss, no cross-table complex calc. Skill (progressive disclosure) with on-demand retrieval + AI self-reflection correction restored accuracy. Core judgment: semantic modeling is key, not model itself. Gartner L2→L3→L4 mapping: currently L2→L3 transition, L4 autonomous decision blank, 12-18 month window. Dual-engine: 小豪 ChatBI for L2普惠; Cici Skill+Agent for deep L3-L4. AI-friendly knowledge base: three sources (auto metadata, human-fed business knowledge, routing rules). Evaluation flywheel: AI-friendly standard score, NL2SQL EM+EX, end-to-end vs OpenAI Trace Grading.

8. Datastrato (Talk B) — Enterprise-Grade Trusted Query Engineering Loop

Speaker: 李明皇 (Senior Engineer, Apache Gravitino PMC). Four hurdles (same as above). Engineering loop: modeling → compilation → maintenance → evaluation → governance. Unified metric semantics compiled to executable SQL. Continuous semantic drift detection, evaluation-set-driven regression. Business domain + tag policies for fine-grained permissions. Uses Apache Gravitino with thousands of metrics. Demo: Agent doesn't guess calibers or free-generate SQL; calls controlled deterministic tools for asset discovery, semantic parsing, query compilation, execution. Returns result plus semantic basis, data sources, tool call chain — explainable, reproducible, auditable. Ambiguity/unsafe joins trigger clarification or refusal. Semantic Generator uses metadata, query history, domain knowledge to produce auditable proposals; combined with deterministic validation, human review, drift detection, regression testing. Closing line: "Insights trustworthy, then operational actions dare to automate."

9. 奇虎360 — Half-Baked Cold Start, Grow While Running

Speaker: 郭朝阳 (Lakehouse Data Agent Evolution Lead). Reality: business won't wait 6 months for perfect semantic layer. Most enterprises only have scattered table docs and tribal knowledge. Core shift: "model writes SQL" → "engine writes SQL"; uncertain → model, certain → engine. But new problem: semantic layer incomplete. Solution: "half-baked cold start, grow while running". Dual-path hybrid routing: semantic-complete → deterministic compilation engine (LLM only extracts intent); only tables/docs → RAG fallback (hybrid retrieval: embedding + keyword + foreign-key graph). Routing decision: PreMatch + confidence scoring. Every result carries confidence score — user knows "this is computed, that is guessed". Confidence enables production use in high-stakes decisions (e.g., product ops analyzing retention impact of feature change): >0.95 auto-include in reports; <0.65 system prompts "uncertain, verify manually". Skill system: thin API → thick Skill → ontology metadata. Aligns with Snowflake Cortex Agent layers; end state: Skill degrades to pure orchestrator, knowledge fully from ontology. Three-level evaluation: L1 programmatic result-set comparison → L2 LLM Judge semantic judgment → L3 human fallback. Data flywheel: user likes filtered, auto-sampled into regression test set; 9 metrics tracked continuously. Supports internal big data cluster ops, S3/PoleFS storage ops, cross-dept (Ops, doc search, membership).

10. 小米 — From BIRD Global 3rd to Production: Semantic Layer as Evolution Fuel

Speaker: 方彪 (Senior Algorithm Engineer). Team placed 3rd on BIRD Text2SQL benchmark. Talk focuses not on rank but on making capability continuously improve in real business. Answer: turn years of BI semantic layer into AI evolution fuel — low-cost domain evaluation sets, drive effect closed-loop iteration, leverage verified eval sets to boost both accuracy and latency. This year opened CLI for users to complete knowledge precipitation and effect iteration via natural language.

11. 悦点科技 — Above Semantic Layer: Make Enterprise Tacit Experience AI-Executable Assets

Speaker: 任鑫琦 (Founder/CEO). 12 years, ~100 enterprises (manufacturing, auto, energy). Core distinction: domain ontology ≠ renamed warehouse semantic layer; it's an independent enterprise business world model organizing business objects, relations, rules, executable actions. Many firms have data platforms + semantic layers + LLM apps, but projects stall: AI can query/answer but cannot enter core business flows because business concepts, metric calibers, constraints, expert judgment logic remain tacit in docs/heads, never delivered to AI. Lessons: inconsistent terminology/calibers; "big and comprehensive" ontologies obsolete fastest; business/algorithm/data tri-party cognition misalignment. Methodology: scenario-first lightweight ontology modeling + "LLM-assisted extraction + expert review" closed-loop iteration. Case: manufacturing quality traceability across CRM/MES/ERP/PLM — from multi-person days to single-person 5 minutes, yield up 5‰. Quote: "LLM is expression/interaction layer; ontology carries enterprise unique business knowledge; together they form long-term AI assets."

12. 每日互动 — OntoOS: Ontology-Driven Agentic Data Analysis & Strategy Execution

Speaker: 书酉 (Algorithm Expert/AI Architect). Enterprise data analysis evolving from BI and single-turn NL2SQL to Agentic analysis covering query, diagnosis, decision, execution. Core challenge: not just executable SQL, but continuous accurate, explainable, auditable results under complex models, calibers, rules, permissions. OntoOS: enterprise knowledge & business semantic ontology OS — unified management of business objects, metrics, relations, rules, permissions, Evidence. Agent handles intent understanding, semantic alignment, query planning, SQL/graph query orchestration, result validation, analysis explanation. Deterministic engine handles data calc, permission control, rule execution, audit trails. Beyond query: metric decomposition, dimension drill-down, correlation analysis, business knowledge reasoning → verifiable attribution hypotheses → identify key segments → form content/benefit/channel/timing operational strategies. Strategies human-reviewed → audience selection, task orchestration, delivery. Post-activity: response/conversion/cost data → cohort performance, strategy effect, attribution hypothesis review → feedback to next analysis/strategy generation. Background: 个推 AIBI topped "NL2SQL hardest global leaderboard" — accuracy global 4th, efficiency global 6th.

Synthesis

The 85%→50% gap is not in the model but in the unfinished semantic layer. Practical answers to "can we start before semantic layer is done?": 360 — dual-path routing with confidence scores; Gaode — architecture shift to Skill with on-demand retrieval + self-reflection; Ant — infrastructure approach (100+ models, 5,000+ metrics, define once consume everywhere); Zhihu — most complete path: MetricFlow → knowledge base → lightweight KG (1,400+ metrics, 300+ wide tables, 25,000+ SQL templates, 65% penetration). Final takeaway from 悦点科技: "LLM is expression/interaction layer; ontology carries enterprise unique business knowledge; together they form long-term AI assets."