seekdb
Ant Group's OceanBase team open-sourced this AI-native search database under Apache 2.0 (seekdb, from the OceanBase 蚂蚁 stack). Vector, full-text and scalar filters in one SQL query, embeddable in...
Overview
seekdb was announced and open-sourced at the OceanBase 2025 annual conference, coming out of the OceanBase team inside Ant Group. The positioning is specific: storage for AI agents, unifying vector, full-text, structured and semi-structured data in one engine so a RAG or agent stack does not need a vector database, a search index and a relational store stitched together. The licence is Apache 2.0, which is the looser end of what Chinese database vendors ship, since OceanBase itself is under Mulan Public License v2.
The engineering is more interesting than the category label suggests. Hybrid queries combine vector similarity, full-text matching and scalar predicates in a single SQL statement. The V1.3.0 release of 25 May 2026 introduced async indexes on a Change Stream incremental framework so write throughput is decoupled from index construction, which is the correct fix for streaming ingest workloads. FORK DATABASE is the standout: a copy-on-write database branch that an agent can write into freely, with MERGE TABLE to promote the results or DROP DATABASE to throw the whole experiment away. Anyone who has watched an agent corrupt a shared table will recognise the problem it solves.
For deployment, three modes share one interface. Embedded runs in-process and installs with the pyseekdb pip package on Linux and macOS. Server mode is a single node. OceanBase mode points the same client at a full distributed cluster. Platforms cover x86_64 and aarch64, Linux with glibc 2.28 or newer, macOS and Windows for server mode, with Python 3.11 as the SDK floor. English documentation lives at docs.seekdb.ai and is usable without Mandarin, though the deeper OceanBase material and the GitHub issue traffic are largely Chinese. Nothing here needs a mainland phone number or business licence unless you choose the hosted cloud instance, which does.
Key Features
- ✓ V1.3.0 released 25 May 2026, adding async indexes built on a Change Stream incremental framework that decouples write throughput from index build
- ✓ Embedded mode: pip install -U pyseekdb installs the engine in-process on Linux and macOS, so there is no server to run for local agents and prototypes
- ✓ HNSW vector indexing with both incremental and snapshot index variants, combined with full-text search and scalar filtering in a single SQL statement
- ✓ FORK DATABASE creates a copy-on-write sandbox, with MERGE TABLE to accept the changes or DROP DATABASE to discard them, aimed squarely at agents that write speculatively
- ✓ MySQL wire compatibility and full ACID transactions inherited from the OceanBase storage engine, with one-line switching between embedded, server and OceanBase cluster modes
- ✓ Built-in AI functions for embeddings, LLM inference and reranking, plus existing integrations for LangChain, LlamaIndex and Dify. Runs on x86_64 and aarch64, Linux with glibc 2.28 or newer, macOS and Windows, Python 3.11+ for the SDK
Where it holds
- • Apache 2.0 on the entire engine, which is a meaningful loosening compared with OceanBase itself under Mulan PubL v2
- • Embedded mode is the shortest route from nothing to working hybrid search inside a Python agent that exists right now in this category
- • FORK DATABASE addresses a real agent engineering problem: giving a model a writable branch you can throw away, rather than trusting it against production tables
- • Inherits a storage engine with a decade of Ant Group financial workload behind it, which is unusual for a project this new
Where it breaks
- • V1.3.0 in May 2026 means roughly seven months of public life. Treat it as early software and pin your version
- • The headline figures, 1,523 QPS on streaming write plus search and a claimed 10x over Milvus, are vendor-published on single-node sub-million-vector tests. Reproduce them on your own data before planning capacity
- • Around 2,900 GitHub stars, so the community is small and most issue traffic and design discussion happens in Chinese
- • English documentation at docs.seekdb.ai exists and is readable, but it is thinner than the Chinese OceanBase material, and the hosted cloud option generally expects a mainland account
My Take
FORK DATABASE is the feature that justifies the attention: a copy-on-write branch an agent can write to, then MERGE TABLE to accept the result or DROP DATABASE to throw it away. The engine is Apache 2.0 throughout, which is looser than OceanBase itself at Mulan PubL v2. V1.3.0 landed on 25 May 2026 with async indexing on a Change Stream framework so ingest no longer blocks on index builds, and pip install -U pyseekdb puts the engine in your Python process on Linux or macOS with no server at all. At roughly 2,900 stars and seven months old the vendor benchmarks deserve scepticism, but the design is thought through in a way most agent-storage projects are not.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Apache 2.0 and free to self-host in embedded, single-node server or full OceanBase cluster mode. The embedded engine arrives with the pip package at no cost and needs no server process. OceanBase sells a managed cloud instance for teams who do not want to operate it, and that path is billed through OceanBase and generally expects a mainland account.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where seekdb holds and where it breaks. The fastest check is your own workload.
Visit website ↗seekdb compared head to head
Alternatives to seekdb
Apache Doris
open-sourceMPP analytics database that began as Baidu's Palo. Version 4.0 put HNSW vector indexes, BM25 full-text search and SQL-callable LLM functions into the same engine that runs the OLAP queries.
TiDB
open-sourcePingCAP's MySQL-compatible distributed SQL database, Apache 2.0 across TiDB, TiKV and TiFlash. Vector columns and HNSW indexes ride on the TiFlash columnar replica, next to transactional rows.
Milvus
open-sourceApache 2.0 vector database from Zilliz, self-hostable from a pip install up to a billion-vector cluster. The open engine under Zilliz Cloud, minus the proprietary Cardinal search core.
pgvector
open-sourcePostgres extension that adds vector types and approximate nearest neighbour indexes, so embeddings live in the same database as the rest of your data.