Chat API Reference

Chat is the high-level interface that wraps messages and Template to render prompts and build tokenized inputs. This is the interface that we recommend users to use.

Minimal Example

from chat_bricks import Chat, get_template
from transformers import AutoTokenizer

tools = [{
    "type": "function",
    "name": "get_weather",
    "description": "Retrieves current weather for the given location.",
    "parameters": {
        "type": "object",
        "properties": {
            "location": {
                "type": "string",
                "description": "City and country e.g. Bogotá, Colombia"
            },
            "units": {
                "type": "string",
                "enum": ["celsius", "fahrenheit"],
                "description": "Units the temperature will be returned in."
            }
        },
        "required": ["location", "units"]
    }
}]

chat = Chat(
    template="qwen3",
    messages=[{"role": "user", "content": "Hello!"}],
)

# Render prompt text
prompt = chat.prompt(tools=tools)
print(prompt)

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
# Tokenize for training or inference
inputs = chat.tokenize(
    tokenizer,
    add_generation_prompt=True,   # keep generation token for inference
    tools=tools                   # optional tool definitions
)

chat_bricks.chat.Chat

__init__(template, messages=None, tools=None, skills=None, tokenizer=None, ignore_tool_calls=False)

Parameters:

• template (str | Template | HFTemplate) –

  The name of the template to use.

• messages (List[Dict], default: None ) –

  The messages to use for the chat.

• tools –

  The tools to use for the chat.

• skills –

  Optional list of skill objects/dicts to advertise via the template's {skills} placeholder (requires skills_template on the template).

• tokenizer (PreTrainedTokenizer, default: None ) –

  The tokenizer to use for the chat.

preprocess_messages(messages)

Preprocess the messages for the chat. Since we don't change message content, we don't copy it here.

prompt(add_generation_prompt=False, tools=None, skills=None, **kwargs)

Get the prompt for the chat.

Parameters:

• add_generation_prompt –

  Whether to add the generation prompt.

• tools –

  The tools to use for the chat.

• skills –

  Optional list of skill objects/dicts (see Chat.__init__).

• **kwargs –

  Additional keyword arguments to pass to the template render method.

Returns:

• str –

  The string formatted prompt for the messages after applying the chat template.

prompt_with_mask(add_generation_prompt=False, tools=None, skills=None, **kwargs)

Get the prompt for the chat with highlight on the masked parts.

Parameters:

• add_generation_prompt –

  Whether to add the generation prompt.

• tools –

  The tools to use for the chat.

• skills –

  Optional list of skill objects/dicts (see Chat.__init__).

• **kwargs –

  Additional keyword arguments to pass to the template render method.

Returns:

• str –

  The string formatted prompt for the messages after applying the chat template with highlight on the masked parts.

set_messages(messages)

Set the messages for the chat.

tokenize(tokenizer=None, add_generation_prompt=False, tools=None, skills=None, processor=None, train_on_last_turn_only=False, **kwargs)

Tokenize the messages.

Parameters:

• tokenizer (PreTrainedTokenizer, default: None ) –

  The tokenizer to use for the chat.

• add_generation_prompt –

  Whether to add the generation prompt.

• tools –

  The tools to use for the chat.

• skills –

  Optional list of skill objects/dicts (see Chat.__init__).

• processor –

  The processor to use for the chat.

Returns:

• inputs ( dict ) –

  Inputs for helping training. - input_ids - attention_mask - labels - action_mask - multi_modal_inputs