Foundational Models

In this chapter, you'll get an overview of Cohere's foundation models.

The foundational models available on the Cohere platform consist of two types: generative and representative models.

Generative models take a text input (the “prompt”) and generate text as output (the “response”). This is useful for question-answering, copywriting, text summarization, and similar use cases.

Representative models take a text input and generate its numerical representation (an “embedding”), which captures its contextual meaning. This is useful for semantic search, text classification, topic modeling, and similar use cases.

Text Generation

Command

command is Cohere’s flagship generative model. Example use cases enabled by this model include:

  • Copy generation: Draft marketing copy, emails, blog posts, product descriptions, docs, and more.
  • Chat: Create chatbots to brainstorm, problem-solve, answer questions, and more. Integrate with a search system to create grounded chat powered by retrieval augmented generation.
  • Text summarization: Summarize articles, transcripts, doctor’s notes, internal documents, and more.

Use the command model if you are optimizing for accuracy.

Command-light

command-light is the smaller and faster version of command. Use command-light if you are optimizing for latency.

Command-nightly-*

There are also nightly versions of command and command-light available. These models, command-nightly and command-light-nightly, are experimental versions of their default counterparts and are released on a more regular basis.

Visit the documentation and the API reference for more information about the text generation models.

Text Representation

Embed-english

embed-english is Cohere’s most performant English text representation model that converts text into vector embeddings.

Example use cases enabled by this model include:

  • Semantic search: Search call transcripts or internal knowledge sources
  • Text classification: Classify intent in customer support chat logs
  • Text clustering: Identify common topics in customer reviews or news data
  • Recommendation: Represent podcast descriptions as a numerical feature for use in a recommendation model

embed-english-light is the smaller and faster version of embed-english. Use embed-english-light if you are optimizing for latency.

Embed-multilingual

Embed-multilingual is the state-of-the-art multilingual embedding model that can convert text in over 100 languages into vector embeddings. It can be applied to the same use cases as Cohere’s English embedding models.

Visit the documentation and the API reference for more information about the embedding models.

Rerank-english

Rerank-english is used to improve the relevance of search systems by performing a second-stage ranking of search results. It can be applied to an existing search system without changing its infrastructure.

Example use cases enabled by this model include:

  • Chatbots: Refine results for Retrieval-Augmented-Generation-based chatbots (or RAG, for short)
  • Enterprise search: Improve question-answering search on enterprise knowledge bases
  • Customer support: Improve the ability for users to self-serve answers from customer support documents

Rerank multilingual

Rerank multilingual is used to improve the relevance of search systems by performing a second-stage ranking of search results in 100+ languages. It can be applied to an existing search system in 100+ languages without changing the existing infrastructure. In addition, it can be applied to the same use cases as Cohere’s English Rerank models.

Visit the documentation and the API reference for more information about the reranking models.


What’s Next

Next, learn about the endpoints of the Cohere API.