Do I need to normalize vectors before similarity search?

It depends on which similarity measure you’re using.

Short answer

• Yes, normalize vectors if you want to use cosine similarity (or if your system uses inner product as a proxy for cosine).
• No, not necessarily if you’re using Euclidean distance and the vector lengths carry meaningful information.
• For dot-product / inner-product search, normalization is optional, but it changes the meaning a lot.

Common cases

1) Cosine similarity

Cosine similarity is based on direction, not magnitude.

• If your database/search library computes cosine similarity directly, you usually don’t need to manually normalize, because it may do it for you.
• If it uses inner product to approximate cosine, then you should normalize both query and indexed vectors to unit length.

Formula:

• Cosine similarity = ( \frac{x \cdot y}{|x||y|} )

If vectors are unit-normalized:

• cosine similarity = dot product

2) Dot product / inner product search

Here, vector magnitude matters.

• Normalize if you want ranking to depend only on direction.
• Don’t normalize if larger magnitudes should imply stronger match.

Example:

• In some embedding models, norm may encode confidence or importance.

3) Euclidean distance (L2)

Normalization is not required.

• Use normalization only if you explicitly want to remove length effects.
• Note: on unit-normalized vectors, L2 and cosine are closely related, so normalization can make L2 behave more like cosine search.

Rule of thumb

• Text embeddings / semantic search: usually normalize and use cosine similarity or inner product on normalized vectors.
• When magnitude matters: do not normalize.
• If unsure: normalize for semantic search, since that’s the most common choice.

Important caveat

Always check what your vector database/library expects:

• some systems have a cosine metric and normalize internally
• some have inner product and assume you already normalized
• some use L2 and should receive raw vectors

If you want, I can also explain how normalization affects ranking or how this works in common vector DBs like FAISS, Pinecone, Milvus, or pgvector.