What are embeddings: the vectors behind semantic search

An embedding is a list of numbers that represents the meaning of a piece of text, an image or another input, so that a program can measure how related two things are. OpenAI's guide defines it as a vector of floating point numbers where the distance between two vectors measures their relatedness: small distances mean closely related, large distances mean unrelated. On October 23, 2026, embeddings are the quiet machinery behind semantic search, retrieval for chatbots, recommendations and clustering. This guide explains what they are, which models exist, what they cost and the mistakes that break them, based on the OpenAI, Google, Anthropic and Ollama documentation read on October 23, 2026.

What an embedding is

A language model reads text as tokens. An embedding model reads the same tokens but, instead of writing an answer, it returns a fixed-length vector: 1,536 numbers for OpenAI's text-embedding-3-small, for example. Texts with similar meaning land close together in that space even when they share no words, so "how do I reset my password" and "I forgot my login" end up near each other, while "password strength rules" lands a little further away.

You never read those numbers directly. You store them, and then you compare them. The usual comparison is cosine similarity, which OpenAI recommends for its models, and which for vectors normalized to length 1 is the same as a simple dot product.

What embeddings are used for

OpenAI lists the common uses: search (rank results by relevance to a query), clustering (group similar texts), recommendations, anomaly detection (find outliers), diversity measurement and classification. The best known today is retrieval for chatbots: you embed your documents once, embed each question, fetch the closest passages and give them to the model as context. That pattern is retrieval-augmented generation, and our comparison of RAG and fine-tuning explains when it beats retraining a model.

Embedding models compared

Table: embedding models from the official documentation (checked October 23, 2026)

Anthropic does not offer an embedding model of its own: its documentation points to Voyage. Gemini's embedding-2 also accepts images, audio, video and PDFs, within limits Google sets per request: up to 6 images, 180 seconds of audio, 120 seconds of video or one PDF of up to 6 pages.

Dimensions, normalization and similarity

More dimensions can capture more nuance, but they cost more storage and slower searches. Both OpenAI and Google let you shorten the vector: OpenAI with the dimensions parameter, Google with output_dimensionality between 128 and 3,072, thanks to a training technique Google calls Matryoshka Representation Learning. Voyage offers int8 and binary quantization that cut storage by 4 and 32 times.

Normalization is where silent bugs live. OpenAI's vectors come normalized to length 1, and Gemini's embedding-2 normalizes automatically, but with gemini-embedding-001 you have to normalize the shortened vectors yourself before comparing them. Skip it and cosine scores drift without any error message.

What embeddings cost

Embedding is one of the cheapest things you can buy from a model API. OpenAI quotes prices as pages per dollar, assuming about 800 tokens per page: 62,500 pages per dollar for text-embedding-3-small and 9,615 for text-embedding-3-large, roughly $0.02 and $0.13 per million tokens. Quality differs less than price: OpenAI reports 62.3% and 64.6% on the MTEB benchmark. Google offers its embeddings through batch mode at 50% of the standard embedding price, the same discount explained in our guide to the batch API. And a local model through Ollama costs nothing per token: our guide to running an LLM locally covers the setup, and Ollama's /api/embed endpoint accepts a list of inputs in one call.

Mistakes that break embeddings

• Mixing models. Vectors from different models live in different spaces. Ollama's documentation is explicit: use the same model for indexing and for querying. Changing models means re-embedding everything.
• Ignoring the input type. Voyage asks you to mark inputs as query or document, Gemini's embedding-001 takes a task_type and embedding-2 uses prompt prefixes instead. Using the wrong one lowers retrieval quality.
• Embedding whole documents. Inputs over the limit get rejected or cut, and a long text averaged into one vector loses detail. Split documents into passages first.
• Expecting world knowledge. OpenAI notes that its third-generation embedding models lack knowledge of events after September 2021. Embeddings capture meaning, not facts, so new product names may not land where you expect.