Ontology-Driven DataAgents: OceanBase's Approach to Reliable Enterprise Analytics

Problem: Vector Similarity Misses Metrics and Cannot Explain Causes

When a bank executive asks "How is business performance?", a traditional RAG system retrieves only the most textually similar metric — often just net profit — missing loan balances, deposit balances, non‑performing loan ratios, and other KPIs that belong to the same business concept. Even when the five metrics are returned, the system cannot explain why a metric changed because the decomposition path (e.g., total deposits → personal demand deposits → specific customer/channel) lacks semantic links.

OceanBase identifies two failure modes: (1) relevant knowledge is not fully retrieved, and (2) after retrieving numbers the system cannot continue to explain causes. The root cause is that traditional RAG fragments enterprise knowledge into retrievable chunks but loses the types, hierarchies, and calculation rules that connect those chunks.

Solution: Ontology Semantic Layer Between Data Platforms and AI

OceanBase inserts an ontology semantic layer between offline/real‑time data warehouses, lakehouses, OLAP engines and upper‑layer AI applications. This layer organizes four elements into a single graph:

Business Objects — entities like Customer, Account, Deposit, Organization.

Business Associations — relationships such as "open", "hold", "belong to".

Business Attributes — categories, statuses, dimensions.

Business Metrics (Calibers) — calculation rules attached to objects.

Objects answer "who", associations answer "how to traverse", attributes answer "which category", and calibers answer "how to calculate". All four coexist in one graph so a business question can step down to concrete fields and filter conditions.

Example: Tracing a Deposit Decline

Query: "Which customers caused the deposit balance drop in East China?" The agent traverses: Customer → (open) → Account → (hold) → Deposit/Transaction → (belong to) → Organization (East China filter) → apply deposit balance caliber. Without objects and associations the AI cannot join "customer", "region", and "balance change" in one query; without calibers it might compute a different number.

Agent Autonomous Exploration on the Graph

Instead of stuffing the entire graph into context or retrieving only a few chunks, the agent explores incrementally:

Browse available business objects.

Expand attributes, associations, and calibers relevant to the current question.

Decide next step based on what it has read — backtrack, branch, or go deeper.

Exploration is not a fixed path. For "deposit balance" the agent may check both Personal Finance and Comprehensive Management domains. A wrong step only adds read cost; stopping too early risks missing the correct caliber.

Concrete Query Execution

For "a branch's personal demand deposit balance":

Locate Deposit object → expand calibers → identify "Personal Demand Deposit Balance" with subject code 2301.

Follow "hold" association to Account → follow "belong to" to Organization → bind branch as filter.

Only after object, caliber, and dimensions are confirmed does the information enter SQL generation.

Root‑Cause Analysis Example

Follow‑up: "Why did East China deposit balance drop?"

General deposit balance down 12% MoM .

Drill into deposit types, customer tiers, organizations, channels → Time deposits down 22% , Mobile banking transfers account for 68% .

Deeper: 68% of time‑deposit decline due to maturity non‑renewal , also via mobile banking.

Two distinct 68% figures correspond to different calibers (channel transfer vs. maturity non‑renewal), giving a concrete direction for business verification.

Lifecycle: Start Small, Validate, Then Converge

OceanBase does not build a company‑wide graph upfront. Instead:

Cold start in a well‑bounded domain (e.g., retail deposits).

Processing : auto‑extract objects, associations, attributes, calibers from table schemas, samples, existing SQL, BI dashboards, ETL lineage.

Growth : business users ask questions; errors surface missing rules (e.g., "exclude margin deposits") which are confirmed and added.

Convergence : only after a sub‑ontology is validated in production is it merged into a cross‑domain graph, preserving necessary distinctions (e.g., "Customer" in retail vs. corporate credit).

Validation is done by letting business users ask real questions and reviewing the agent's traversal path, not by expert node‑by‑node review. Rules that never appear in documents emerge only through repeated correction.

Productization and Long‑Term Goal

This approach powers OceanBase's DataPilot product. The lasting asset is not a single answer but the evolving knowledge structure that survives repeated business scrutiny. Each clarified caliber, corrected association, and traced anomaly becomes reusable context for the next analysis — moving DataAgent from occasional correctness to long‑term reliable understanding of enterprise business.