Why it matters

E5 is Microsoft embedding family. Understanding shapes options.

Advertisement

The architecture

Contrastive pretraining on CCPairs.

Task prefixes: 'query:' and 'passage:'.

Multilingual + instruct variants.

E5 embedding familyContrastive pretrainCCPairsTask prefixesquery / passageMultilingual+ instructWang et al. 2022; strong on MTEB; task-conditioning via prefixes
E5 math.
Advertisement

How it works end to end

Wang et al. 2022.

E5-large + Multilingual E5 + Instruct E5.

Prefix conditioning.