How OceanBase Builds a Data Foundation for AI Agents

Introduction

In the AI era, the rapid growth of AI agents and the emergence of massive multimodal data create new requirements for data isolation, sandboxing, and mixed‑modality search. OceanBase responds by evolving its core engine into an AI Database and launching the OceanBase AI Lake to meet these challenges.

Challenges from AI Agents

AI agents now perform substantive enterprise work, generating thousands of apps that need independent data and compute isolation. They require long‑term memory (e.g., medical history) and trial‑and‑error cycles that can affect production systems if not sandboxed.

Engineering Solutions

Logical Tenant

OceanBase implements logical tenants that provide flow‑control and isolation at the compute layer while sharing storage resources (schemas, tables, indexes) at the storage layer. When an app becomes inactive, its schema and indexes can be off‑loaded to OSS, reducing long‑term storage cost.

Layered Decoupling Architecture

Active applications run on high‑performance NVMe SSDs for millisecond‑level response, whereas inactive apps are moved to object storage, freeing compute resources and lowering cost.

Hybrid Search Capability

To support long‑term memory, OceanBase offers mixed search across four modalities:

Scalar filtering (e.g., range queries on medication time)

Keyword text search

Vector similarity search

Knowledge‑graph traversal

Benchmark results show strong performance: OceanBase leads in TPC‑C and TPC‑H tests, and in Vector DB Bench it competes favorably against Milvus, Elasticsearch and pgvector. In many full‑text scenarios its performance surpasses Elasticsearch, with further improvements planned.

Database Branching for Millisecond Sandboxes

Database Branching creates a full‑data branch from the production database in milliseconds. Failed trials can be deleted without affecting production; successful branches can become the new production instance. This capability is already used in Haier’s agent scenarios.

Multimodal Data Governance

Non‑structured data now accounts for 80‑90% of enterprise data and grows faster than structured data. OceanBase redesigns large‑object handling:

Small objects (e.g., images) stored inline.

Medium objects stored in dedicated tables.

Very large objects stored in object storage.

This approach leverages transactional consistency and point‑in‑time rollback while controlling storage cost.

Model in DB and AI Columns

Model‑in‑DB stores model metadata and prompts as column definitions. AI columns generate structured or semi‑structured data from unstructured inputs (images, audio, video). The inference process is asynchronous, versioned, and retryable, avoiding the pitfalls of simple JSON storage. The generated data resides in wide tables, enabling both traditional SQL processing and hybrid search.

Typical Deployments

OceanBase’s AI capabilities have been applied to:

25 million “flash” apps in the Vibe Coding ecosystem.

Taobao’s “Universal Search” memory store.

GAC’s intelligent driving scenario where users query rain‑day images.

Sales‑team audio analysis: from raw recordings to transcribed text, model‑derived structure, and SQL analytics—all within OceanBase.

These cases demonstrate that multimodal data can become a queryable, governed, and decision‑making asset.

Conclusion

OceanBase’s multimodal database integrates metadata management, compute orchestration, and storage optimization, introducing the AI Lake to turn multimodal data into a unified, searchable, and governed asset. The system’s innovations—logical tenants, layered decoupling, hybrid search, database branching, and AI columns—address the core technical challenges of AI‑agent workloads.