Embeddings API: OpenAI-compatible vectors, processed in the EU

POST /v1/embeddings takes a model and an input, which is either one string or an array of strings, and returns one vector per input. The response has a data array in which each item carries an index matching the position of the input, and an embedding, a list of floats. usage reports the prompt_tokens used.

Multilingual models such as qwen3-embedding-8b place a German, an English and a French sentence with the same meaning close to each other, so a question in one language finds documents in another. That matters for European companies whose contracts, tickets and wikis are rarely in one language.

Send batches rather than one string per request: an array of inputs is processed together and is far more efficient. Each input must fit within the model's maximum input length, so split long documents into chunks first.

dimensions shortens the vector on models trained to support it, such as the Qwen3 embedding family. Shorter vectors take less storage and make search faster, with a small loss of accuracy; test 512 or 1024 against the full size on your own data. encoding_format can be float or base64; base64 makes large responses smaller over the network. Always use the same model and dimension for documents and queries.

A typical retrieval pipeline converts documents to clean text, splits them into chunks of a few hundred tokens with some overlap, embeds each chunk and stores the vectors with the source text in a vector database such as pgvector, Qdrant or Weaviate. At query time, embed the question, fetch the 20 to 50 nearest chunks by cosine similarity, and pass them to a reranker to keep the best five or so.

Then give those chunks to a chat model as context. For Qwen3 embeddings, prefix queries with a short task instruction, as the model card recommends, and embed documents without it. Run initial indexing of large collections through the batch API.

An input longer than the model's limit returns a 400 error rather than being silently cut, so chunk before you send. Empty strings are rejected too; filter them out. Sending a chat model id to this endpoint returns an error that names the model type, and an unknown id returns 404.

Vectors from different models, or from the same model at different dimensions, cannot be compared. If you change either, re-embed the whole collection. Most models return normalised vectors, so cosine similarity and dot product give the same ranking. Retry 429 responses with backoff, and spread large indexing jobs over time or use batches.