Embed 5 Pro
Neural Network
Embed 5 Pro is a Cohere multimodal embedding model for semantic search and RAG on text and images.
Specifications
- Input
- text
- Output
- vectors (embeddings)
- Context size
- 128 000 tokens
- Vector dimensions
- 2 048
Prompt cost
(per 1M tokens)
How it works Embed 5 Pro?
Embed 5 Pro is the highest-accuracy tier of the Cohere Embed 5 family, with about 2.4 billion parameters. The model converts text, images, and mixed content into a unified vector space, enabling semantic search rather than keyword matching. Through BotHub, you can connect without a VPN or foreign card, pay in rubles as you go, and access over 250 models in one window to compare on your data; corporate teams get contracts, invoices, and limits. Scenarios: hybrid semantic search over knowledge bases, RAG over Qdrant or pgvector, clustering and deduplication of multimodal catalogs, and finding similar products by photo and description.Frequently asked questions about Embed 5 Pro
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.
embed-5-pro turns text into vectors—embeddings—which are useful for searching for similar fragments by meaning rather than keyword matching. This forms the basis for semantic search, RAG systems, knowledge bases, clustering, and thematic document classification. A context window of up to 128k tokens allows processing large text fragments in their entirety.
embed-5-pro does not generate text, images, or files: its output is a vector, a set of numbers of fixed dimensionality that describes the meaning of the input text. This vector is passed to a vector database or search index, such as Qdrant or pgvector, and used to compare fragments by proximity. Check the provider's documentation for vector dimensionality.
We have not observed a separate reasoning step before outputting results for embed-5-pro. It is not a chat model: it receives text as input and immediately outputs a vector for search or clustering, without intermediate explanations or action plans. If you need a chain of thought, it should be built on the application side or using another model.
We have not observed function calling or structured JSON output in our data. By design, embed-5-pro is not a conversational model: it returns a numerical vector, not a text response, so tools typical for chat models do not apply. If you need JSON or function calling, it should be generated by the application using the retrieved vectors.
We have no confirmation of image, audio, or video input in our data: we have only observed text processing, with embeddings as output. We will not claim that such capabilities do not exist, but you should not rely on them for design. Check the provider's documentation for the current list of supported input formats.
Our data does not contain an assessment of its performance with the Russian language, so we cannot make specific promises. Embedding models are usually evaluated by how closely related Russian fragments are placed in the vector space. You should test this with your own query and document pairs before implementing it in search or RAG.