OpenAI adds an embeddings endpoint to its API

OpenAI added an embeddings endpoint to its API, offering models that convert text or code into numerical vectors rather than generating further text. The release comprised three families of models, each tuned for a different task — text similarity, text search, and code search — built on the same underlying GPT-3 series and available at several sizes with a corresponding cost and quality trade-off.

The stated purpose was to make a set of previously research-grade techniques into an API call: semantic search, clustering, recommendation, anomaly detection, and classification, all of which had until then typically required training or fine-tuning a bespoke model. A vector produced by the endpoint placed semantically similar pieces of text near each other in a high-dimensional space, so that comparing two vectors — rather than the underlying words — could stand in for comparing meaning.

OpenAI presented the models as improving on prior embedding approaches on standard text-similarity and search benchmarks, while acknowledging the models were not fine-tuned for every downstream task and that performance could vary by domain. The endpoint’s pricing and rate limits made it substantially cheaper to call than generating completions, positioning it as infrastructure rather than a headline product.

The release is a minor entry on its own terms, but it mattered downstream: turning documents into vectors and using nearest-neighbour lookups to feed relevant context into a language model — the technique that became known as retrieval-augmented generation — depended on serviceable, affordable embeddings being available as a commodity API call. OpenAI later replaced the January 2022 family with a single consolidated model, text-embedding-ada-002, at the end of the year.