zvec: The SQLite of AI Vector Databases

Vector databases are all the rage these days, but most of them feel like overkill for small to medium projects. Do you really need a dedicated cluster just to store a few thousand embeddings?

That's where zvec comes in. Built by Alibaba, it's a lightweight, embedded vector database that feels a lot like SQLite for AI. It's designed to be simple, fast, and zero-config. No servers, no orchestration, no cloud dependencies. Just drop it into your app and start running similarity searches.

What It Does

zvec is a compact vector database that stores and retrieves high-dimensional vectors (like embeddings from AI models) using approximate nearest neighbor (ANN) search. Think of it as a library you embed into your application process, not a separate service you connect to over the network.

It supports:

• Adding, searching, and deleting vectors
• Cosine similarity and Euclidean distance metrics
• Basic persistence to disk
• A simple, clean API (C/C++ with bindings for Python and Rust)

The core idea is that it's embedded, meaning your app and the database live in the same memory space. No network calls, no socket overhead.

Why It's Cool

1. Zero infrastructure. Most vector databases spin up a server, open ports, require a config file. zvec is just a library. Include it, call a function, done. This is huge for prototyping, CI/CD pipelines, or running AI features on edge devices.

2. Fast and minimal. Because everything is in-process, you avoid the serialization and network latency of client-server setups. It's built for speed when your dataset fits in memory (which it does, for most hobby or small-scale projects).

3. SQLite vibes. If you've ever used SQLite for local storage, zvec feels familiar. You call zvec_init(), zvec_insert(), zvec_search(). No daemons, no connection pools, no mental overhead.

4. Production at Alibaba. It's not just a side project. Alibaba uses this internally for real workloads, so it's battle-tested. But the interface stays refreshingly simple.

How to Try It

The quickest way to get started is with the C example or the Python bindings.

Python (via pip):

pip install zvec

Then in code:

import zvec

# Initialize an in-memory index
index = zvec.Index(dim=128, metric="cosine")

# Add some vectors
index.add([(0, [0.1]*128), (1, [0.2]*128)])

# Search for top 5 nearest neighbors
results = index.search([0.1]*128, top_k=5)
print(results)

C example:

Clone the repo, build with cmake, then run the sample:

git clone https://github.com/alibaba/zvec
cd zvec
mkdir build && cd build
cmake .. && make
./example/simple_search

Full docs and more examples are in the README.

Final Thoughts

zvec isn't trying to replace Milvus or Pinecone for massive, sharded workloads. It's for the other 90% of use cases where you just need to quickly search a few thousand vectors without spinning up a server. It's the kind of tool every developer should have in their toolbox, especially if you're building AI-powered features into desktop apps, mobile prototypes, or CI testing.

If you've been avoiding vector databases because they felt heavyweight, give zvec a look. Sometimes the simplest tool really is the best.

Shared by @githubprojects