oceanbase/seekdb

🌊 seekdb – A State Store for AI Agents

What it is – seekdb is an open‑source, MySQL‑compatible database that combines traditional relational storage with vector and full‑text search. It is built on the OceanBase SQL engine and can run embedded in a process, as a single‑node server, or in a distributed OceanBase cluster.

Why it matters for AI – Modern AI agents need to:

• Write continuously (store observations, embeddings, logs) and read back within milliseconds.
• Branch and experiment safely (think of “sandbox” environments for planning).
• Retrieve using a mix of vector similarity, keyword search, and relational filters – all in a single SQL query.

seekdb provides these capabilities out of the box, letting developers focus on the agent logic rather than building glue code between a relational DB, a vector engine, and a full‑text index.

---

🔑 Core Features

Feature	What it does
Hybrid Search	One SQL statement can combine VECTOR similarity (l2_distance, cosine, etc.), FULLTEXT keyword matching, and scalar predicates. No client‑side merging required.
Async Index Pipeline (Change Stream)	Writes commit immediately; a background pipeline updates a delta HNSW index, so newly inserted vectors are searchable in the next query without a P99 latency spike.
Two‑Level HNSW	Incremental (delta) + snapshot indexes give fast writes and stable high‑throughput reads.
Copy‑on‑Write (COW) Sandboxes	FORK DATABASE creates an instant snapshot of the whole DB; agents can experiment, then MERGE TABLE or DROP DATABASE to accept or discard changes.
MySQL‑compatible protocol	Works with any MySQL driver, SQLAlchemy, LangChain, LlamaIndex, Dify, etc.
Deployment Flexibility	Embedded library, Docker image, binary install, or cloud‑hosted instance (zero‑install curl).
Performance	Benchmarks claim ~1,500 QPS with 21 ms P99 latency – roughly 10× Milvus and 3× Elasticsearch on the same workload.

---

🚀 Getting Started (quickest path)

# Python SDK – the recommended way for AI/ML projects
pip install -U pyseekdb

import pyseekdb
client = pyseekdb.Client(path="./agent_state.db")
mem = client.get_or_create_collection(name="episodic")
# write an observation
mem.upsert(ids=[1], documents=["User said hello"], embeddings=[[0.1, 0.2, …]])
# retrieve relevant context a few ms later
results = mem.query(query_texts=["hello"], n_results=5)
print(results)

For a full server deployment you can run the Docker image:

docker run -d -p 2881:2881 -p 2886:2886 \
  -v $(pwd)/data:/var/lib/oceanbase \
  oceanbase/seekdb:latest

All installation options (cloud instance, binary, Homebrew) are described in the README and the official docs.

---

📚 Typical Use Cases

• Agentic AI memory – continuous write‑then‑read loops for personal assistants, autonomous bots, or workflow automation.
• RAG / Knowledge bases – store documents with embeddings and run vector + keyword + filter queries in one go.
• Semantic code search – index code snippets as vectors, combine with text search, and isolate projects via sandbox databases.
• Enterprise AI layer – add a vector‑enabled AI front‑end to existing MySQL‑based applications without migrating data.
• Edge / on‑device AI – embed the binary in a device; the COW sandbox lets local agents experiment safely.

---

🤝 Ecosystem & Integrations

seekdb ships with a Python SDK (pyseekdb) and integrates with the major LLM‑app frameworks:

• LangChain, LlamaIndex, Dify, LangGraph, Coze, Spring‑AI‑Alibaba, FastGPT, Jina AI, Ragas, Instructor, Baseten, and more. These integrations are mostly PRs that add a VectorStore implementation, so developers can drop a seekdb client into existing pipelines.

---

📄 License & Community

• License: Apache License 2.0 (permissive, commercial‑friendly).
• Community: Discord, GitHub Discussions, issue tracker, and a contributor guide. The project is backed by OceanBase, the database powering large‑scale services at Alibaba, Alipay, DiDi, etc.

---

Bottom line – seekdb is a purpose‑built, high‑performance state store that lets AI agents treat a database like a mutable knowledge graph: write fast, query instantly, and experiment safely with sandboxed forks. It bridges the gap between traditional relational databases and modern vector search engines, all under a familiar MySQL interface.