A Python library to generate fixed dimensional embeddings for multi-dimensional vectors based on Google Muvera.

pip install --upgrade muvfde

uv venv --python 3.12
uv pip install -e .
uv pip install -r integeration_test/requirements.txt
python integeration_test/benchmark.py

Check out integeration_test/benchmark.py for usage examples.

• The FDE algorithm is highly hyperparameter sensitive. The hyperparameters used in the benchmark are not optimal for all models.
• FDE quality is tightly linked to dimensionality; increase in dimensionality will generally result in quality gains.

Benchmark Results

• Google Muvera Blog
• MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encodings
• Reference Fixed Dimensional embeddings code