Infera

In-Database Machine Learning for DuckDB

Infera is a DuckDB extension that allows you to use machine learning (ML) models directly in SQL queries to perform inference on data stored in DuckDB tables. It is developed in Rust and uses Tract as the backend inference engine. Infera supports loading and running models in ONNX format. Check out the ONNX Model Zoo repository on Hugging Face for a large collection of ready-to-use models that can be used with Infera.

In a conventional data science workflow, when data is stored in a database, it is not typically possible to use ML models directly on the data. Users need to move the data out of the database first (for example, export it to a CSV file) and load the data into a Python or R environment, run the model there, and then import the results back into the database. This process is time-consuming and inefficient. Infera aims to solve this problem by letting users run ML models directly in SQL queries inside the database. It simplifies the workflow and speeds up the process for users, and eliminates the need for moving data around.

• Adds ML inference as a first-class citizen in SQL queries.
• Supports loading and using local as well as remote models.
• Supports using ML models in ONNX format with a simple and flexible API.
• Supports performing inference on table columns or raw tensor data.
• Supports both single-value and multi-value model outputs.
• Supports autoloading all models from a specified directory.
• Thread-safe, fast, and memory-efficient.

See the ROADMAP.md for the list of implemented and planned features.

1. Clone the repository and build the Infera extension from source:

git clone --recursive https://github.com/CogitatorTech/infera.git
cd infera

make release

2. Start DuckDB shell (with Infera statically linked to it):

./build/release/duckdb

3. Run the following SQL commands in the shell to try Infera out:

-- Normally, we need to load the extension first,
-- but the `duckdb` binary that we built in the previous step
-- already has Infera statically linked to it.
-- So, we don't need to load the extension explicitly.

-- 1. Load a simple linear model from a remote URL
select infera_load_model('linear_model',
                         'https://github.com/CogitatorTech/infera/raw/refs/heads/main/test/models/linear.onnx');

-- 2. Run a prediction using a very simple linear model
-- Model: y = 2*x1 - 1*x2 + 0.5*x3 + 0.25
select infera_predict('linear_model', 1.0, 2.0, 3.0);
-- Expected output: 1.75

-- 3. Unload the model when we're done with it
select infera_unload_model('linear_model');

-- 4. Check the Infera version
select infera_get_version();

Check out the docs directory for the API documentation, how to build Infera from source, and more.

Check out the examples directory for SQL scripts that show how to use Infera.

See CONTRIBUTING.md for details on how to make a contribution.

Infera is available under either of the following licenses:

• MIT License (LICENSE-MIT)
• Apache License, Version 2.0 (LICENSE-APACHE)

• The logo is from here with some modifications.