DeepSeek

DeepSeekV3

DeepSeekV3(model_name: str, skill_sources: str | None = None, system_extension: str | None = None, system_template: str = SYSTEM_TEMPLATE, execution_output_template: str = EXECUTION_OUTPUT_TEMPLATE, execution_error_template: str = EXECUTION_ERROR_TEMPLATE, api_key: str | None = None, **kwargs)

Bases: LiteLLM

Code action model class for DeepSeek V3.

Parameters:

Name	Type	Description	Default
`model_name`	`str`	The LiteLLM-specific name of the model.	required
`skill_sources`	`str \| None`	Skill modules source code to be included into `system_template`.	`None`
`system_extension`	`str \| None`	Domain- or environment-specific extensions to the `system_template`.	`None`
`system_template`	`str`	Prompt template for the system instruction that guides the model to generate code actions. Must define a `{python_modules}` placeholder for the `skill_sources` and an `{extensions}` placeholder for the `system_extension`.	`SYSTEM_TEMPLATE`
`execution_output_template`	`str`	A template for formatting successful code execution output. Must define an `{execution_feedback}` placeholder.	`EXECUTION_OUTPUT_TEMPLATE`
`execution_error_template`	`str`	A template for formatting code execution errors. Must define an `{execution_feedback}` placeholder.	`EXECUTION_ERROR_TEMPLATE`
`api_key`	`str \| None`	Provider-specific API key. If not provided, reads from `DEEPSEEK_API_KEY` environment variable.	`None`
`**kwargs`		Default completion kwargs.	`{}`

Source code in freeact/model/deepseek/model.py

def __init__(
    self,
    model_name: str,
    skill_sources: str | None = None,
    system_extension: str | None = None,
    system_template: str = v3.SYSTEM_TEMPLATE,
    execution_output_template: str = v3.EXECUTION_OUTPUT_TEMPLATE,
    execution_error_template: str = v3.EXECUTION_ERROR_TEMPLATE,
    api_key: str | None = None,
    **kwargs,
):
    system_instruction = system_template.format(
        python_modules=skill_sources or "",
        extensions=system_extension or "",
    )
    super().__init__(
        model_name=model_name,
        execution_output_template=execution_output_template,
        execution_error_template=execution_error_template,
        system_instruction=system_instruction,
        api_key=api_key or os.getenv("DEEPSEEK_API_KEY"),
        **kwargs,
    )

DeepSeekR1

DeepSeekR1(model_name: str, skill_sources: str | None = None, instruction_extension: str | None = EXAMPLE_EXTENSION, instruction_template: str = INSTRUCTION_TEMPLATE, execution_output_template: str = EXECUTION_OUTPUT_TEMPLATE, execution_error_template: str = EXECUTION_ERROR_TEMPLATE, api_key: str | None = None, **kwargs)

Bases: LiteLLM

Code action model class for DeepSeek R1.

Does not set a system instruction. All instructions go into the first user message formatted with instruction_template.

Parameters:

Name	Type	Description	Default
`model_name`	`str`	The LiteLLM-specific name of the model.	required
`skill_sources`	`str \| None`	Skill modules source code to be included into `system_template`.	`None`
`instruction_extension`	`str \| None`	Domain- or environment-specific extensions to the `instruction_template`.	`EXAMPLE_EXTENSION`
`instruction_template`	`str`	Prompt template that guides the model to generate code actions. Must define a `{user_query}` placeholder for the user query, an `{extensions}` placeholder for the `instruction_extension` and a `{python_modules}` placeholder for the `skill_sources`.	`INSTRUCTION_TEMPLATE`
`execution_output_template`	`str`	A template for formatting successful code execution output. Must define an `{execution_feedback}` placeholder.	`EXECUTION_OUTPUT_TEMPLATE`
`execution_error_template`	`str`	A template for formatting code execution errors. Must define an `{execution_feedback}` placeholder.	`EXECUTION_ERROR_TEMPLATE`
`api_key`	`str \| None`	Provider-specific API key. If not provided, reads from `DEEPSEEK_API_KEY` environment variable.	`None`
`**kwargs`		Default completion kwargs.	`{}`

Source code in freeact/model/deepseek/model.py

def __init__(
    self,
    model_name: str,
    skill_sources: str | None = None,
    instruction_extension: str | None = r1.EXAMPLE_EXTENSION,
    instruction_template: str = r1.INSTRUCTION_TEMPLATE,
    execution_output_template: str = r1.EXECUTION_OUTPUT_TEMPLATE,
    execution_error_template: str = r1.EXECUTION_ERROR_TEMPLATE,
    api_key: str | None = None,
    **kwargs,
):
    default_kwargs = {
        "temperature": 0.6,
        "max_tokens": 8192,
    }

    super().__init__(
        model_name=model_name,
        execution_output_template=execution_output_template,
        execution_error_template=execution_error_template,
        system_instruction=None,
        api_key=api_key or os.getenv("DEEPSEEK_API_KEY"),
        **(default_kwargs | kwargs),
    )

    self.instruction_template = instruction_template
    self.instruction_kwargs = {
        "python_modules": skill_sources or "",
        "extensions": instruction_extension or "",
    }

extract_code

extract_code(response: LiteLLMResponse) -> str | None

Extracts the last code block from response.text.

DeepSeek-R1 often produces code blocks during thinking but usually only the last code block in the actual response is relevant.

Source code in freeact/model/deepseek/model.py

def extract_code(self, response: LiteLLMResponse) -> str | None:
    """Extracts the last code block from `response.text`.

    DeepSeek-R1 often produces code blocks during thinking but usually
    only the last code block in the actual response is relevant.
    """
    return code_block(response.text, -1)