FastEmbed
A lightweight, fast Python library for generating text, sparse and image embeddings using the ONNX runtime instead of PyTorch, with no GPU required.
FastEmbed is a lightweight Python library for embedding generation, built to avoid downloading gigabytes of PyTorch dependencies by running models through the ONNX Runtime instead. Its default text embedding model is a Flag Embedding model that supports separate query and passage prefixes and appears on the MTEB leaderboard.
It requires no GPU, uses data parallelism to encode large datasets faster, and covers several embedding types beyond dense text: sparse text embeddings such as SPLADE++, late-interaction models such as ColBERT, and image embeddings through models such as a CLIP vision encoder. New models outside its default supported list can also be added to dense text embedding.
It installs with pip, with an optional GPU-enabled package, and is released under the Apache 2.0 license.