How Ant Group Scaled Apache Ossie Semantic Layer from Zero to 5,000 Metrics

Large language models often produce hallucinations, unstable answers, and lack confidence when directly answering business data questions; for example, asking for last month’s East China GMV may yield fabricated numbers, inconsistent responses, or ignorance of changing definitions.

Ant Group addresses this gap by inserting a semantic layer between the model and raw data. The layer structures all business knowledge—metrics, dimensions, entities, processes, physical implementations, and SQL rules—so the AI operates within explicit constraints, generates SQL, and explains results instead of touching naked data.

Compared with text knowledge bases, RAG vector retrieval, and full‑metadata retrieval, the semantic layer offers three sharper advantages: clearer query constraints, explainable outcomes, and lower inference cost. It serves both human analysts and AI agents, delivering accurate answers for both.

After platformization, Ant unified more than 100 models and over 5,000 metrics under a single source of truth. The semantic layer also resolved long‑standing issues such as expired definitions, downstream desynchronization after table changes, inconsistent local versions, and scattered knowledge across teams by centrally managing definitions, versions, mapping relationships, and consumption governance.

The foundation is Apache Ossie, an open‑source semantic‑layer specification of which Ant is a core contributor. The presentation walks through the evolution from zero to 5,000 metrics, covering semantic modeling, semantic retrieval, ontology construction, semantic mapping, evaluation framework, MCP integration, and data‑governance practices—all in a single engineering pipeline.

For further details, the talk was delivered by Wang Xiaojun at DACon Beijing on 23‑24 Oct 2026, with a limited‑time 20% discount and registration via the provided link.