sentence-transformers: paraphrase-MiniLM-L3-v2 vectors are not unit-norm (~4-7) while all-MiniLM-L6-v2/bge-small are, so swapping models breaks dot-product indexes
encode() does not normalize by default (normalize_embeddings=False). Whether you get unit vectors depends on whether the model's module list ends in Normalize, so swapping embedding models in a RAG pipeline that uses a dot-product index (faiss.IndexFlatIP, pgvector inner product <#>) silently changes scores.
Modules and vector norms (sentence-transformers 6.3.0): - all-MiniLM-L6-v2: Transformer, Pooling, Normalize -> norm 1.0 - BAAI/bge-small-en-v1.5: Transformer, Pooling, Normalize -> norm 1.0 - paraphrase-MiniLM-L3-v2: Transformer, Pooling -> norms 4.6-5.7 (7.2 for a one-word text)
All three report model.similarity_fn_name == 'cosine', so model.similarity() looks fine while the index computes a raw dot product. With the L3 model, dot scores land around 2-27 (breaks any 0-1 similarity threshold), and longer-norm short texts get boosted: for 'what is a cat?' the doc 'cat' (norm 7.17) outranked 'A cat.' on dot product though cosine ranks 'A cat.' first.
Fix: normalize both indexed docs and queries at encode time:
emb = model.encode(texts, normalize_embeddings=True) # float32
index = faiss.IndexFlatIP(emb.shape[1]); index.add(emb)Check a model once with any(type(m).__name__ == 'Normalize' for m in model).
sentence-transformers 6.3.0, torch 2.15.0, transformers 5.23.0, Python 3.12, CPU.