Skip to content

Encoding profiles

The models rebasis knows without being told. Values come from each model card's official retrieval instructions, not from guesswork.

Why this table exists. Many retrieval models encode a query differently from a document — usually with a prefix. Getting that wrong produces no error and no warning; it only lowers quality, which is the hardest kind of failure to attribute. So the prefix lives in the profile, the profile is fingerprinted, and the fingerprint blocks an adapter from loading against an index it was not built for.

A model not listed here still works. Pass --dim, and --query-prefix / --document-prefix if it is asymmetric. rebasis will not guess a prefix: being asked once is cheaper than debugging a quality drop.

Asymmetric marks the models where the two prefixes differ. For those, auto measures both a shared adapter and a query-specific one and keeps whichever scores better on the held-out set — M0 measured the mean difference at -0.003, with the sign varying by model pair.

Model Dim Asymmetric Query prefix Document prefix Pooling
BAAI/bge-base-en-v1.5 768 yes Represent this sentence for searching relevant passages: cls
BAAI/bge-large-en-v1.5 1024 yes Represent this sentence for searching relevant passages: cls
BAAI/bge-m3 1024 no cls
BAAI/bge-small-en-v1.5 384 yes Represent this sentence for searching relevant passages: cls
intfloat/e5-base-v2 768 yes query: passage: mean
intfloat/e5-large-v2 1024 yes query: passage: mean
intfloat/e5-small-v2 384 yes query: passage: mean
intfloat/multilingual-e5-base 768 yes query: passage: mean
jinaai/jina-embeddings-v2-base-en 768 no mean
minishlab/potion-base-8M 256 no mean
mixedbread-ai/mxbai-embed-large-v1 1024 yes Represent this sentence for searching relevant passages: cls
nomic-ai/nomic-embed-text-v1.5 768 yes search_query: search_document: mean
sentence-transformers/all-MiniLM-L12-v2 384 no mean
sentence-transformers/all-MiniLM-L6-v2 384 no mean
sentence-transformers/all-mpnet-base-v2 768 no mean
thenlper/gte-base 768 no mean