Chat

Generates a text response to a user message and streams it down, token by token. To learn how to use the Chat API with streaming follow our Text Generation guides.

Follow the Migration Guide for instructions on moving from API v1 to API v2.

Request

This endpoint expects an object.

streamfalseRequiredDefaults to false

Defaults to false.

When true, the response will be a SSE stream of events.

Streaming is beneficial for user interfaces that render the contents of the response piece by piece, as it gets generated.

modelstringRequired

The name of a compatible Cohere model or the ID of a fine-tuned model.

messageslist of objectsRequired

A list of chat messages in chronological order, representing a conversation between the user and the model.

Messages can be from User, Assistant, Tool and System roles. Learn more about messages and roles in the Chat API guide.

toolslist of objectsOptional

A list of tools (functions) available to the model. The model response may contain ‘tool_calls’ to the specified tools.

Learn more in the Tool Use guide.

documentslist of strings or objectsOptional

A list of relevant documents that the model can cite to generate a more accurate reply. Each document is either a string or document object with content and metadata.

citation_optionsobjectOptional

Options for controlling citation generation.

response_formatobjectOptional

Configuration for forcing the model output to adhere to the specified format. Supported on Command R, Command R+ and newer models.

The model can be forced into outputting JSON objects by setting { "type": "json_object" }.

A JSON Schema can optionally be provided, to ensure a specific structure.

Note: When using { "type": "json_object" } your message should always explicitly instruct the model to generate a JSON (eg: “Generate a JSON …”) . Otherwise the model may end up getting stuck generating an infinite stream of characters and eventually run out of context length.

Note: When json_schema is not specified, the generated object can have up to 5 layers of nesting.

Limitation: The parameter is not supported when used in combinations with the documents or tools parameters.

safety_modeenumOptional

Used to select the safety instruction inserted into the prompt. Defaults to CONTEXTUAL. When OFF is specified, the safety instruction will be omitted.

Safety modes are not yet configurable in combination with tools and documents parameters.

Note: This parameter is only compatible newer Cohere models, starting with Command R 08-2024 and Command R+ 08-2024.

Note: command-r7b-12-2024 and newer models only support "CONTEXTUAL" and "STRICT" modes.

Allowed values:

max_tokensintegerOptional

The maximum number of output tokens the model will generate in the response. If not set, max_tokens defaults to the model’s maximum output token limit. You can find the maximum output token limits for each model in the model documentation.

Note: Setting a low value may result in incomplete generations. In such cases, the finish_reason field in the response will be set to "MAX_TOKENS".

Note: If max_tokens is set higher than the model’s maximum output token limit, the generation will be capped at that model-specific maximum limit.

stop_sequenceslist of stringsOptional

A list of up to 5 strings that the model will use to stop generation. If the model generates a string that matches any of the strings in the list, it will stop generating tokens and return the generated text up to that point not including the stop sequence.

temperaturedoubleOptional

Defaults to 0.3.

A non-negative float that tunes the degree of randomness in generation. Lower temperatures mean less random generations, and higher temperatures mean more random generations.

Randomness can be further maximized by increasing the value of the p parameter.

seedintegerOptional>=0<=18446744073709552000

If specified, the backend will make a best effort to sample tokens deterministically, such that repeated requests with the same seed and parameters should return the same result. However, determinism cannot be totally guaranteed.

frequency_penaltydoubleOptional

Defaults to 0.0, min value of 0.0, max value of 1.0. Used to reduce repetitiveness of generated tokens. The higher the value, the stronger a penalty is applied to previously present tokens, proportional to how many times they have already appeared in the prompt or prior generation.

presence_penaltydoubleOptional

Defaults to 0.0, min value of 0.0, max value of 1.0. Used to reduce repetitiveness of generated tokens. Similar to frequency_penalty, except that this penalty is applied equally to all tokens that have already appeared, regardless of their exact frequencies.

kintegerOptional>=0<=500Defaults to 0

Ensures that only the top k most likely tokens are considered for generation at each step. When k is set to 0, k-sampling is disabled. Defaults to 0, min value of 0, max value of 500.

pdoubleOptional

Ensures that only the most likely tokens, with total probability mass of p, are considered for generation at each step. If both k and p are enabled, p acts after k. Defaults to 0.75. min value of 0.01, max value of 0.99.

logprobsbooleanOptional

Defaults to false. When set to true, the log probabilities of the generated tokens will be included in the response.

tool_choiceenumOptional

Used to control whether or not the model will be forced to use a tool when answering. When REQUIRED is specified, the model will be forced to use at least one of the user-defined tools, and the tools parameter must be passed in the request. When NONE is specified, the model will be forced not to use one of the specified tools, and give a direct response. If tool_choice isn’t specified, then the model is free to choose whether to use the specified tools or not.

Note: This parameter is only compatible with models Command-r7b and newer.

Allowed values:

thinkingobjectOptional

Configuration for reasoning features.

raw_promptingbooleanOptional

When enabled, the user’s prompt will be sent to the model without any pre-processing.

Compatible Deployments: Cohere Platform, Azure, AWS Sagemaker/Bedrock, Private Deployments

strict_toolsbooleanOptionalBeta

When set to true, tool calls in the Assistant message will be forced to follow the tool definition strictly. Learn more in the Structured Outputs (Tools) guide.

Note: The first few requests with a new set of tools will take longer to process.

Response

idstring

Unique identifier for the generated reply. Useful for submitting feedback.

finish_reasonenum

The reason a chat request has finished.

complete: The model finished sending a complete message.
max_tokens: The number of generated tokens exceeded the model’s context length or the value specified via the max_tokens parameter.
stop_sequence: One of the provided stop_sequence entries was reached in the model’s generation.
tool_call: The model generated a Tool Call and is expecting a Tool Message in return
error: The generation failed due to an internal error

Allowed values:

messageobject

A message from the assistant role can contain text and tool call information.

usageobject or null

logprobslist of objects or null

1	import cohere
2
3	co = cohere.ClientV2()
4
5	response = co.chat(
6	model="command-a-03-2025",
7	messages=[{"role": "user", "content": "hello world!"}],
8	)
9
10	print(response)

1	{
2	"id": "c14c80c3-18eb-4519-9460-6c92edd8cfb4",
3	"finish_reason": "COMPLETE",
4	"message": {
5	"role": "assistant",
6	"content": [
7	{
8	"type": "text",
9	"text": "LLMs stand for Large Language Models, which are a type of neural network model specialized in processing and generating human language. They are designed to understand and respond to natural language input and have become increasingly popular and valuable in recent years.\n\nLLMs are trained on vast amounts of text data, enabling them to learn patterns, grammar, and semantic meanings present in the language. These models can then be used for various natural language processing tasks, such as text generation, summarization, question answering, machine translation, sentiment analysis, and even some aspects of natural language understanding.\n\nSome well-known examples of LLMs include:\n\n1. GPT-3 (Generative Pre-trained Transformer 3) — An open-source LLM developed by OpenAI, capable of generating human-like text and performing various language tasks.\n\n2. BERT (Bidirectional Encoder Representations from Transformers) — A Google-developed LLM that is particularly good at understanding contextual relationships in text, and is widely used for natural language understanding tasks like sentiment analysis and named entity recognition.\n\n3. T5 (Text-to-Text Transfer Transformer) — Also from Google, T5 is a flexible LLM that frames all language tasks as text-to-text problems, where the model learns to generate output text based on input text prompts.\n\n4. RoBERTa (Robustly Optimized BERT Approach) — A variant of BERT that uses additional training techniques to improve performance.\n\n5. DeBERTa (Decoding-enhanced BERT with disentangled attention) — Another variant of BERT that introduces a new attention mechanism.\n\nLLMs have become increasingly powerful and larger in scale, improving the accuracy and sophistication of language tasks. They are also being used as a foundation for developing various applications, including chatbots, content recommendation systems, language translation services, and more.\nThe future of LLMs holds the potential for even more sophisticated language technologies, with ongoing research and development focused on enhancing their capabilities, improving efficiency, and exploring their applications in various domains."
10	}
11	]
12	},
13	"usage": {
14	"billed_units": {
15	"input_tokens": 5,
16	"output_tokens": 418
17	},
18	"tokens": {
19	"input_tokens": 71,
20	"output_tokens": 418
21	}
22	}
23	}

Headers

Request

Response

Errors