Why is cosine similarity preferred over Euclidean distance for text embeddings? · Zenaique

Attempt it

Your answer

**Cosine similarity** measures the angle between two vectors. **Euclidean distance** measures the straight line distance between their tips. For text embeddings, angle is what carries semantic meaning, and the straight line distance gets polluted by vector magnitude. Text embedding models do not always produce vectors of the same length. Longer documents, denser passages, and slight differences in tokenisation can change the magnitude of the resulting vector even when the meaning is the same. If you use Euclidean distance, two vectors describing the same concept but with different magnitudes will appear far apart because their tips end up in different places in space. Cosine ignores this because two vectors pointing in the same direction always have an angle of zero between them, regardless of how long they are. A useful mental picture: cosine asks where on the unit sphere a vector points. All vectors get projected onto the sphere, and similarity is measured there. Two passages on the same topic should point to nearby spots on the sphere, even if one is a sentence and the other is a paragraph. Many modern embedding APIs normalise vectors to unit length before returning them. When all vectors are unit length, cosine and Euclidean ranking are equivalent, because the only difference between two unit vectors is their angle. But you cannot assume this is happening unless the documentation says so.

**Cosine similarity** wins for text embeddings because direction in embedding space carries semantic meaning, while magnitude carries mostly uninteresting variation like document length and token density. The definitions diverge on the magnitude axis. Cosine similarity is one formula. Euclidean distance is another. $$\text{cos}(u, v) = \frac{u \cdot v}{\lVert u \rVert \, \lVert v \rVert}$$ $$d_{\text{euc}}(u, v) = \lVert u - v \rVert$$ Cosine divides out both norms. Two vectors pointing in the same direction always score 1 regardless of length. Euclidean distance grows with the gap in magnitudes even when the directions match. Since longer passages tend to produce vectors with larger norms in many embedding models, Euclidean conflates length with semantic difference. In practice the distinction collapses when vectors are pre-normalised to unit length, because two unit vectors differ only by angle. Most modern embedding APIs from OpenAI, Voyage, Cohere, and Nomic return unit-norm vectors, which means cosine and Euclidean rank identically. But the conceptual reason cosine is the right default is the magnitude-invariance argument. That argument is what matters when you build the index or pick a similarity metric.

Quick reference

TL;DR

Text embedding norms vary with content length and density, so cosine measures direction (semantic meaning) while ignoring magnitude. Euclidean distance would conflate length with semantic difference.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine arrows on a giant map, all starting from the centre and pointing outward. Each arrow stands for a sentence. We want to measure how alike two sentences are. One way: ask if the arrows point in the same direction, no matter how long they are. Another way: ask how close the tips of the arrows are. The first way only cares about direction. The second cares about direction and length together. For sentences, the direction is what carries meaning. The length usually just reflects boring things like how many words were in the sentence. If we measured by tip position, two paragraphs about the same topic but with different lengths would look unrelated. By only looking at direction, they line up the way a person would expect: same idea, same heading on the map.

Key concepts

Embeddings Vector Databases

Detailed answer & concept explanation

~4 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

Open with direction vs magnitude as the conceptual axis. State why direction carries meaning and magnitude carries length-and-density noise. Note that modern APIs normalise, which makes cosine and Euclidean rank equivalently. Close with the contrastive-training argument tying the inference metric to the training objective.

Real-world examples

Real products, models, and research that use this idea.

OpenAI's `text-embedding-3-large` returns unit-normalised vectors, so dot-product equals cosine and either is the natural index metric.
Voyage AI's `voyage-3` and Cohere's `embed-english-v3` both follow the same unit-norm convention for the same reason.
pgvector supports cosine, L2, and inner product distances and lets the user pick per index, but the documentation recommends cosine for sentence-embedding workloads.
Pinecone's documentation explicitly recommends cosine for text and L2 for image embeddings, reflecting how the training objective shapes the right metric.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy does the contrastive training objective most modern embedding models use favour cosine at inference?

Sign in or go Pro to see the approach for this follow-up.

QHow does Matryoshka representation learning interact with the choice of distance metric?

Sign in or go Pro to see the approach for this follow-up.

QWhat goes wrong if you compute cosine similarity on vectors that include large mean offsets per dimension?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Assuming cosine and Euclidean rank results identically. They agree only when all vectors are normalised to unit length. Modern embedding APIs often already L2-normalise, hiding the distinction in practice.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.