E5-small-v2

E5-small-v2 is the compact member of Microsoft's E5 embedding family (intfloat/e5-small-v2), trained with weakly-supervised contrastive pre-training on large-scale web text pairs (CCPairs) plus supervised fine-tuning on MS MARCO and NQ. At ~33M parameters producing 384-dimensional vectors over 512 tokens, it became the default lightweight retrieval baseline across RAG stacks, vector-DB examples, and MTEB leaderboard comparisons — the model everything else in its weight class gets measured against.

The E5 family was the first to beat the BM25 baseline on BEIR zero-shot without labeled data, and the v2 training refresh improved on the original release. The honest picture: E5 models require the quirky 'query:' / 'passage:' instruction prefixes at inference — skip them and quality degrades noticeably, which trips up first-time users. At 512 tokens max and 2022-era training data, it is outgunned by modern Matryoshka and LLM-backbone embedders, but nothing this small and this free is easier to run everywhere.