Open AI Embeddings API

The OpenAI Embeddings API is a hosted endpoint that converts text into vectors with a single HTTP call. You send one string or a batch of strings to the /v1/embeddings endpoint with a model name (the text-embedding-3-small and text-embedding-3-large family), and you get back a float array per input. There’s nothing to host or fine-tune — it’s the fastest path from “I have text” to “I have embeddings.”

It matters because it’s the default choice in a huge share of production RAG systems, so you’ll encounter it whether or not you pick it yourself. It also introduced a genuinely useful feature: the dimensions parameter, which lets you shorten vectors (say, from 3072 down to 256) while preserving most of their quality — cutting vector-database storage and search cost without switching models. Pricing is per token, and batching inputs into a single request is dramatically cheaper and faster than looping one string at a time.

In practice you’ll call it through the official SDK — client.embeddings.create(model="text-embedding-3-small", input=[...]) — and read vectors from response.data[i].embedding. You’ll batch documents at index time, embed queries one at a time at search time, and always use the same model and dimension setting for both. Watch the input token limit per string, chunk long documents before embedding, and normalize your cost math to tokens, not calls.