Skip to content

Learn AI · Foundations

Embeddings Explained

Embeddings turn text into numbers that capture meaning. They power search, RAG, recommendations, and clustering — here's how.

6 min readPublished Mar 2025Updated Apr 2026
EmbeddingsVectorsSemantic Search
Edited by The AIKnowHub team · Editorial team

Key takeaways

  • 1Embeddings are vectors — typically 384, 768, 1024, 1536, or 3072 numbers.
  • 2Cosine similarity is the standard 'how close are these two?' metric.
  • 3Don't mix embedding models — vectors from different models aren't compatible.
  • 4Embeddings don't capture recency, negation, or numerical precision well.
  • 5For most apps in 2026, text-embedding-3-small or voyage-3 is a great starting point.

The idea in one image

An embedding is a list of numbers — typically 384, 768, 1536, or 3072 of them — that represents a piece of text as a point in high-dimensional space. Texts with similar meaning end up close together; unrelated texts end up far apart.

"How do I reset my password?" → [0.12, -0.43, 0.88, ...]

"I forgot my login credentials" → [0.14, -0.41, 0.85, ...] ← very close

"What's a good pasta recipe?" → [-0.61, 0.22, 0.04, ...] ← far away

Computing how close two embeddings are (cosine similarity) is the primitive that drives semantic search.

Where they come from

You feed text into an embedding model — text-embedding-3-small (OpenAI), voyage-3 (Voyage), embed-english-v3 (Cohere), or a local sentence-transformers model — and out comes the vector.

What they're used for

  • Semantic search — find documents by meaning, not keyword match.
  • RAG — retrieve relevant context for an LLM.
  • Clustering — group similar tickets, products, posts.
  • Classification — train a tiny model on top of embeddings instead of fine-tuning the LLM.
  • Recommendations — "items similar to this one."
  • Deduplication — find near-duplicate content.

Picking a model

Three things matter:

  1. Quality on retrieval benchmarks (MTEB is the standard leaderboard).
  2. Dimension — bigger isn't always better; 1024-dim often beats 3072-dim per dollar.
  3. Cost — embedding millions of documents adds up fast.

For most apps in 2026: text-embedding-3-small or voyage-3-lite is a great starting point. Local options like bge-small-en-v1.5 are free and surprisingly strong.

Cosine similarity in one line

The standard metric is the cosine of the angle between two vectors:

similarity = (a · b) / (|a| * |b|)

Most embedding APIs return normalized vectors (|v| = 1), so cosine similarity reduces to the dot product.

Gotchas

  • Don't mix models. Embeddings from model A and model B aren't comparable. If you switch, re-embed everything.
  • Embeddings don't know recency. "Latest news" and "news from 2010" can look identical in vector space.
  • Long documents lose signal. Embed chunks, not whole books.
  • Negation breaks them. "I love coffee" and "I don't love coffee" are dangerously close.
  • Numbers and dates get embedded as text — exact arithmetic is unreliable. Filter on metadata instead.

Embeddings are simple, fast, and shockingly useful. They're the single best return-on-investment primitive in applied AI.

Common misconceptions

The wrong-but-common takes worth correcting.

Myth

Bigger embeddings are always more accurate.

Reality

1024-dim often beats 3072-dim at half the cost. Pick by retrieval benchmark (MTEB), not by dimension count.

Myth

Embeddings understand negation.

Reality

'I love coffee' and 'I don't love coffee' end up dangerously close in vector space. If negation matters, you need additional logic.

Myth

Embeddings are domain-agnostic.

Reality

Models trained on general web text underperform on highly technical corpora (medicine, law, niche code). Specialized models or fine-tuning may be needed.

Real-world use cases

  • Semantic search

    Find documents by meaning instead of by keyword match.

  • RAG

    Retrieve the right chunks of context to give an LLM at query time.

    Open

  • Clustering

    Group similar tickets, posts, products by their embedding — useful for tagging, deduplication, and topic discovery.

  • Classification

    Train a tiny model on top of embeddings to classify text into your custom categories — cheaper than fine-tuning an LLM.

Frequently asked questions

A token is a piece of text. An embedding is a numerical vector representing meaning. LLMs work with both — tokens go in, intermediate embeddings flow through layers, and tokens come out.

Related on AIKnowHub