You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.

codestral-embed-2505 by Mistral converts code and text into vectors. Hence the tasks: semantic repository search, RAG for documentation, clustering and deduplication of snippets, finding similar functions, and classification. The 8192-token context window is enough for a large file or an entire related code block.

The output is not text, but a numerical vector of fixed dimensionality. Such embeddings are stored in vector databases — Qdrant, pgvector, Milvus — compared using cosine similarity, and used to get a ranked list of similar snippets. See the API response for the first request for the exact dimensionality.

Our data does not indicate reasoning capabilities before answering, and the model's task is different: it does not formulate an answer but calculates a vector representation of the input. Logic on top of the results is built by your code or a generative model connected to the same API.

Our data does not indicate function calling or structured output — this is not a primary scenario for embeddings. However, the API response comes as standard JSON with an array of vectors, so parsing it in a pipeline and saving it to a database is straightforward.

Our data does not indicate support for anything other than text, so expect text and code. If your project needs multimodal inputs, you can switch to a suitable model in BotHub via the unified OpenAI-compatible API without rewriting your integration.

We have no data on languages, so we cannot promise accuracy — please test it on your own corpus. A practical tip: run both queries and documents through the same model; this keeps the vectors comparable and the search predictable.