Ant Group's Financial Data Ontology: Automating Business Semantics with LangGraph

Ant Group faced a classic data governance problem: the same concept — "active user" — carried different definitions across Huabei, Jiebei, and consumer‑loan business lines. With thousands of tables and calibration logic buried in historical SQL code, manual harmonization was unsustainable.

The root cause was not computational but semantic: fields and metrics had never been translated into business objects such as "user", "product", or "behavior", so the system could not understand what the data actually meant.

Solution: Financial Data Knowledge Ontology

Ant built a Financial Data Knowledge Ontology that acts as a "textbook" teaching the system business semantics. Construction is automated through two pipelines:

LangGraph six‑node state machine —

InputParser → Comprehender → EntityResolver → Reconciler → Synthesizer → Verifier

— handles massive table‑structure perception, entity recognition, cross‑table contradiction resolution, and ontology synthesis with verification.

"Code as annotation" pipeline — reverse‑extracts field semantics and inter‑table relationships from ODPS historical SQL, turning existing code into training signals.

Core Principle: Mark Conflicts, Don't Force Merge

When business lines disagree on a metric's definition, the ontology records the conflict but preserves each line's original semantics . Agents then retrieve the relevant caliber on demand, avoiding premature standardization that would lose contextual nuance.

Three‑Layer Joint Retrieval Architecture

Analysis requests are served by a fused KnowledgePackage assembled from:

KG_cs — business caliber definitions

KG_fr — entity relationship graph

RC — raw document evidence

Results from the three layers are deduplicated, ranked, and packaged for the downstream Agent.

Measured Outcomes

Coverage: thousands of tables

Cross‑opportunity identification: reduced from 3–5 days to hours

Anomaly root‑cause analysis accuracy: >85%

Already supporting multiple marketing‑strategy deployments

Reusable Methodology Highlights

A seven‑layer progressive‑cost fusion strategy for automated ontology construction at massive scale

Layered joint retrieval (KG_cs / KG_fr / RC) that boosts business‑analysis efficiency

Practical pattern for multi‑business‑line caliber conflicts in finance: retain differences, retrieve contextually