GitHub - stephenc222/mysql_vss: MySQL plugin for storing and querying vector embeddings.

MySQL VSS

mysql_vss is a plugin designed for storing and searching vector embeddings using approximate nearest neighbor search, leveraging the Annoy library for fast lookups in high-dimensional spaces.

🚧 Experimental Stage: The plugin is experimental and not recommended for production use yet.

Setup

Prerequisites

• Install recent versions of gcc and g++.
• Ensure necessary build tools and libraries are installed. These include:
  • g++
  • gcc
  • libstdc++-static
  • cmake
  • openssl-devel
  • python-devel
  • ncurses-devel

Building the Plugin

1. Clone the repository:

git clone https://github.com/stephenc222/mysql_vss
cd mysql_vss

2. Initialize and update submodules:

git submodule update --init --recursive --progress

3. Compile mysql-server. This plugin requires a mysql-server build from source to link against:

cd src/vendor/mysql-server
mkdir build
cd build
cmake ..
make

*NOTE: mysql_vss uses mysql-server version 8.0, and this is pegged in the mysql-server git submodule. There will be additional packages you'll need to install, such as a modern version of bison and others. CMake's output is pretty helpful in determining what you need

4. Compile the mysql-vss plugin source:

The compiled output is a shared library named something like libmysql_vss_v0.0.1_AmazonLinux2023_x86_64.so, tailored to your operating system.

Deploying the Plugin

1. Deploy the shared library to MySQL's plugin directory.
2. Register the UDFs in MySQL:

CREATE FUNCTION vss_search RETURNS STRING SONAME 'libmysql_vss.so';
CREATE FUNCTION vss_version RETURNS STRING SONAME 'libmysql_vss.so';

Usage

Verification

Verify installation by checking the plugin version:

SELECT CAST(vss_version() AS CHAR);

Schema Setup

Set up the embeddings table as per the required schema (currently a manual process):

CREATE TABLE IF NOT EXISTS embeddings (
    ID INT PRIMARY KEY,
    vector JSON NOT NULL,
    original_text TEXT NOT NULL,
    annoy_index INT
);

Performing Searches

Use vss_search for querying similar embeddings:

SELECT e.original_text
FROM embeddings AS e
WHERE FIND_IN_SET(e.ID, CAST(vss_search('[0.01,0.02,0.03,...]') AS CHAR)) > 0
ORDER BY FIELD(e.ID, CAST(vss_search('[0.01,0.02,0.03,...]') AS CHAR)) DESC;

Known Issues and Future Enhancements

Embedding Dimensions

• The current version only supports 768-dimensional embeddings. mysql_vss was developed using the embedding model, gte-base, a top performing embedding model that is runnable on a wide variety of consumer hardware.

Annoy Index Management

• The Annoy Index loads once and requires a manual update for new embeddings.

Data Scaling

To enhance mysql_vss for larger data scales, some possible future development considerations include:

• External Process Management: Isolate the Annoy Index to manage memory separately from MySQL.
• Dynamic Index Reloading: Enable index updates without restarting the service.

Additionally, configurations that you can take as well:

• Containerization: Apply memory quotas to contain the Annoy Index within resource limits.
• MySQL Optimization: Tweak MySQL configurations for larger datasets (e.g., innodb_buffer_pool_size).
• Performance Benchmarks: Test against various data sizes to determine performance and stability.

Testing

• Expand testing to include unit and integration tests for robust validation.

Check out examples/app.py and the provided Dockerfile in the repository for demonstration and containerized deployment of mysql_vss.