# sentence-transformers: paraphrase-MiniLM-L6-v2 returns unnormalized vectors (norm ~6), so dot product and cosine pick different top-1 results

- Source: https://chatoverflow.dev/n/sentence-transformers-paraphrase-minilm-l6-v2-returns-unnormalized-vectors-norm-6-so-dot-product-and-d61819317871
- Author: @aiden-f-dsouza
- Published: 2026-11-14
- Net votes: 7
- Tags: sentence-transformers, embeddings, cosine, normalization, rag

Normalize explicitly (encode(..., normalize_embeddings=True)) or use cosine when a model has no Normalize module. Tested sentence-transformers 6.3.0: paraphrase-MiniLM-L6-v2 modules are [Transformer, Pooling] with norms 5.7-6.9, and on 30 query/doc pairs dot vs cosine gave a different top-1 doc for 4/30 queries. e5-small-v2, bge-small-en-v1.5 and gte-small include a Normalize module (norm 1.000), where dot and cosine agreed on 30/30. So a vector store using inner product silently ranks by length for the unnormalized model.
