Cohere Command A in the catalog: a model for RAG and document work

The Yenisei catalog includes Cohere's flagship, Command A (command-a). Cohere has traditionally built models for enterprise tasks, above all RAG and knowledge base work. If your task is "answer from our documents, not from general knowledge", this model is the first one to look at.

The compact Command R7B is not in the catalog; below is what to take instead.

Command A

• 256K context, comfortable for working with sets of documents.
• Built for RAG. The model sticks to the attached sources well and is less inclined to invent facts that are not in the context.
• Citations. The Command lineup has historically been strong at showing which fragment an answer came from, which is critical where verifiability is required.
• Multilingual, and it handles Russian fine.

How to call

curl https://api.yenisei.ru/v1/chat/completions \
  -H "Authorization: Bearer $YENISEI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "command-a",
    "messages": [
      {"role": "system", "content": "Answer only on the basis of the attached documents. Cite the source."},
      {"role": "user", "content": "What is the warranty period under the contract? Documents: ..."}
    ]
  }'

Instead of Command R7B

R7B was a compact model for volume: rough fragment selection, simple answers, preprocessing before calling the flagship. In the Yenisei catalog this role is covered by:

• Mistral Nemo (mistral-nemo), light and fast, with tool calling;
• Granite 4.1 8B (granite-8b), IBM's compact model for utility tasks;
• GPT-OSS 20B (gpt-oss-20b), OpenAI's open model, when you need a bit more reasoning at a low price.

Who it suits

Take Command A if you are building an assistant over an internal knowledge base: policies, support, legal documents, technical documentation. Its strength is discipline: sticking to the source and not making things up. For creative writing or heavy reasoning there are more interesting models, but in its genre Command A is very predictable.

Put the lightweight model on volume inside the same RAG system: it can take the initial filtering, and Command A is called only for the final answer. Searching the base also needs embeddings, for example BGE-M3 (bge-m3).

Prices are on the plans and pricing page, billed in rubles, 1 credit = 1 ₽.