MenxLi/tiny_vectordb

A small and fast Python JIT vector database

Tiny Vector Database

A lightweight vector database designed for small projects.

Features

• Just-in-time (JIT) compiling to optimize vector operations by setting the vector size at compile time.
• Accelerates vector operations using Eigen.
• Processes vectors using only Python lists, no need for any additional third-party data formats.
• Stores vectors as base-64 encoded strings in a SQLite database.

Performance
More than 10x Faster than numpy-based vector operations.

Development Status

It is currently compatible with g++ or clang++.
You may need to modify compile_config parameter in VectorDatabase initialization to inject your compile commands.
To make it work with other compiler, you may need to change tiny_vectordb.jit module.

Installation

pip install tiny_vectordb

Good to go!

Uninstallation

The package will emit some compiled files in the source directory, which may not be automatically removed using pip uninstall, so you need to run the following command manually if you want to uninstall the package comletely.

python -c "import tiny_vectordb; tiny_vectordb.cleanup()"

After that, you can safely uninstall the package with:

pip uninstall tiny_vectordb

Usage:

from tiny_vectordb import VectorDatabase

collection_configs = [
    {
        "name": "hello",
        "dimension": 256,
    },
    {
        "name": "world",
        "dimension": 1000,
    }
]
database = VectorDatabase("test.db", collection_configs)
collection = database["hello"]

# add vectors
collection.setBlock(
    ["id1", "id2"],             # ids
    [[1] * 256, [2] * 256]      # vectors
)

# search for nearest vectors
search_ids, search_scores = collection.search([1.9] * 256)  

For more usage, see example.py.

---

Designing Note:

1. No numpy array is used in the database, because I want it to be as lightweight as possible, and lists of numbers are eaiser to be converted into json for communication with http requests.

2. The data are always stored in contiguous memory to ensure the best searching performance.
   So the addition and deletion are preferred to be done in batches as they envolve memory reallocation.
   Here are some useful functions for batch operations:

class VectorCollection(Generic[NumVar]):
    def addBlock(self, ids: list[str], vectors: list[list[NumVar]]) -> None:
    def setBlock(self, ids: list[str], vectors: list[list[NumVar]]) -> None:
    def deleteBlock(self, ids: list[str]) -> None:
    def getBlock(self, ids: list[str]) -> list[list[NumVar]]: