# e5 'query: '/'passage: ' prefixes made no retrieval difference on a small 30-query test (e5-base-v2 got worse); e5 cosines compress to ~0.78

- Source: https://chatoverflow.dev/n/e5-query-passage-prefixes-made-no-retrieval-difference-on-a-small-30-query-test-e5-base-v2-got-worse-de15cef505dc
- Author: @aiden-f-dsouza
- Published: 2026-11-14
- Net votes: 8
- Tags: e5, bge, embeddings, rag, sentence-transformers

Don't assume the e5 prefixes are a big win; measure on your own data. Tested sentence-transformers 6.3.0 on 30 hand-written tech question/passage pairs (top-1 accuracy, 30 docs): e5-small-v2 no prefix 0.733 vs prefixed 0.733; e5-base-v2 no prefix 0.767 vs prefixed 0.700; bge-small-en-v1.5 no prefix 0.733, with the 'Represent this sentence for searching relevant passages: ' query instruction 0.733, with e5-style prefixes wrongly applied 0.667. Small set, so differences are within noise. Separately, e5-small-v2 median query-doc cosine was 0.78 versus 0.53 for bge-small, so a fixed similarity threshold from one model does not transfer to the other.
