E5-base maintainers direct users to v2

The E5 family provides open text encoders for semantic search, retrieval-augmented generation, clustering, and related tasks. Its maintainers have added a migration notice to the original E5-base card, directing users to E5-base-v2 for better reported performance with the same usage pattern. Existing deployments can retain their input conventions and vector dimensions, but they must rebuild stored embeddings.

V2 keeps the 768-dimensional contract

Both checkpoints use 12-layer BERT-style encoders, produce 768-dimensional vectors, and carry an MIT license. They follow the contrastive training method described in the E5 paper, which trains the encoder to place related texts close together and unrelated texts farther apart.

E5-base-v2 retains the query: and passage: prefixes, attention-mask-aware average pooling, L2 normalization, and cosine similarity scoring. The new checkpoint creates a separate embedding space, so every indexed document needs a v2 embedding before the query encoder switches. Mixing v1 document vectors with v2 query vectors can produce unreliable rankings even though their dimensions match.

Published scores move upward

The model cards report higher v2 scores on the selected MTEB classification datasets below. MTEB covers several embedding tasks, including retrieval, classification, clustering, and semantic similarity; these rows represent its classification slice.