Cohere Command: RAG, Web Content and Citations

Command is Cohere’s family of language models for businesses. The models have no web access of their own. An application sends them text, and when it sends documents, the answer can cite them. So a Command answer cites your page only when an application fetched the page and passed it to the model.

Command models

Developers use Command models through the Cohere Chat API. Cohere’s models page lists these Command models as live in October 2026:

Model ID	Notes
command-a-plus-05-2026	Command A+, text and image input, 128K context
command-a-03-2025	Command A, 256K context
command-a-reasoning-08-2025	Reasoning model, 256K context
command-a-vision-07-2025	Image input
command-a-translate-08-2025	Translation in 23 languages
command-r7b-12-2024	Small, fast model
command-r-plus-08-2024 and command-r-08-2024	Command R+ and Command R updates from August 2024

The same page shows that Cohere deprecated the original command, command-light, command-r-03-2024 and command-r-plus-04-2024 models on September 15, 2025.

How an application gives Command web content

Retrieval-augmented generation (RAG) means that an application gives the model documents to answer from. Cohere’s RAG guide, written for version 2 of the API, describes this flow:

1. The model can write search queries from the question of the user.
2. The application runs those queries against its own data source or a search API and fetches the documents. Cohere does not describe this step, because it depends on the search API.
3. The application passes the documents in the documents parameter. Each document is a data object with an optional id.
4. The model answers from the documents and adds citations.

The guide builds the search step as a tool. In tool use, the model asks the application to run a function, and the application sends back the result.

The model does not choose web pages itself. If an application wants web content, it must fetch the pages and pass them to the model. The request to your site comes from that application, with the user agent that its developer set.

How Command cites sources

Cohere’s citations guide says that the Command family produces fine-grained citations. Each citation has a start and an end position in the answer, the cited text, and the sources that support it:

• In a normal response, the citations are in the message.citations field.
• In a streamed response, they arrive in citation-start events.
• If a document has an id, the API uses it as the citation reference. If not, the API makes an ID in the form doc:<id>.

For streamed responses, a developer can choose between two citation modes. The default, accurate, adds citations after the full answer. The fast mode, set with citation_options={"mode": "fast"}, adds citations while the answer is written, with slightly less precise matches.

A citation points to a document, not to a URL. The application decides whether to show a link to your page next to the answer. If you want a link, the application must keep the URL of each page with its document.

Which bots to allow or block

Cohere’s web crawlers page lists no crawler today, and Cohere says that it does not crawl the web to train its models at this time. Bot directories still name two Cohere tokens, cohere-ai and cohere-training-data-crawler. The cohere-ai guide explains what is known about them and how to block them.