E5 Embeddings

E5 Embeddings is a family of text embedding models for developers building search and retrieval systems. It represents text as numerical vectors so applications can match queries with relevant passages on their own hardware.

• English models: Choose E5-small-v2, E5-base-v2 or E5-large-v2 to compare model size and retrieval performance within the same family.
• Multilingual models: Use multilingual-e5-small, multilingual-e5-base or multilingual-e5-large for multilingual text, with multilingual-e5-large-instruct as an instruction-based option.
• Mistral-based embeddings: E5-mistral-7b-instruct provides a separate instruction-based model supported by Hugging Face Transformers.
• Model evaluation: Use the project's BEIR retrieval and MTEB embedding evaluations to compare checkpoints, including multilingual evaluation.

Checkpoints are available on Hugging Face under the intfloat account. Models whose names end in -unsupervised use only unlabeled datasets for pretraining. The project's Python code is open source under the MIT license.

The BEIR evaluation scripts run on GPU machines and can use all available GPUs. Encoding a corpus can take substantial time, particularly with E5-mistral-7b-instruct.