LLM

LLM modules to connect to different LLM providers. Also extracts code, name and description.

class llamea.llm.DeepSeek_LLM(api_key, model='deepseek-chat', temperature=0.8, **kwargs)

Bases: OpenAI_LLM

A manager class for the DeepSeek chat models.

class llamea.llm.Dummy_LLM(model='DUMMY', **kwargs)

Bases: LLM

query(session_messages)

Sends a conversation history to DUMMY model and returns a random response text.

Args:

session_messages (list of dict): A list of message dictionaries with keys

“role” (e.g. “user”, “assistant”) and “content” (the message text).

Returns:

str: The text content of the LLM’s response.

class llamea.llm.Gemini_LLM(api_key, model='gemini-2.0-flash', **kwargs)

Bases: LLM

A manager class for handling requests to Google’s Gemini models.

query(session_messages, max_retries: int = 5, default_delay: int = 10)

Sends the conversation history to Gemini, retrying on 429 ResourceExhausted exceptions.

Args:

session_messages (list[dict]): [{“role”: str, “content”: str}, …] max_retries (int): how many times to retry before giving up. default_delay (int): fallback sleep when the error has no retry_delay.

Returns:

str: model’s reply.

class llamea.llm.LLM(api_key, model='', base_url='', code_pattern=None, name_pattern=None, desc_pattern=None, cs_pattern=None, logger=None)

Bases: ABC

extract_algorithm_code(message)

Extracts algorithm code from a given message string using regular expressions.

Args:

message (str): The message string containing the algorithm code.

Returns:

str: Extracted algorithm code.

Raises:

NoCodeException: If no code block is found within the message.

extract_algorithm_description(message)

Extracts algorithm description from a given message string using regular expressions.

Args:

message (str): The message string containing the algorithm name and code.

Returns:

str: Extracted algorithm name or empty string.

extract_configspace(message)

Extracts the configuration space definition in json from a given message string using regular expressions.

Args:

message (str): The message string containing the algorithm code.

Returns:

ConfigSpace: Extracted configuration space object.

abstract query(session: list)

Sends a conversation history to the configured model and returns the response text.

Args:

session (list of dict): A list of message dictionaries with keys

“role” (e.g. “user”, “assistant”) and “content” (the message text).

Returns:

str: The text content of the LLM’s response.

sample_solution(session_messages: list, parent_ids: list | None = None, HPO: bool = False, base_code: str | None = None, diff_mode: bool = False)

Generate or mutate a solution using the language model.

Args:

session_messages: Conversation history for the LLM. parent_ids: Identifier(s) of parent solutions. HPO: If True, attempt to extract a configuration space. base_code: Existing code to patch when diff_mode is True. diff_mode: When True, interpret the LLM response as a unified

diff patch to apply to base_code rather than full source code.

Returns:

tuple: A tuple containing the new algorithm code, its class name, its full descriptive name and an optional configuration space object.

Raises:

NoCodeException: If the language model fails to return any code. Exception: Captures and logs any other exceptions that occur during the interaction.

set_logger(logger)

Sets the logger object to log the conversation.

Args:

logger (Logger): A logger object to log the conversation.

class llamea.llm.Multi_LLM(llms: list[LLM])

Bases: LLM

query(session_messages: list)

Sends a conversation history to the configured model and returns the response text.

Args:

session (list of dict): A list of message dictionaries with keys

“role” (e.g. “user”, “assistant”) and “content” (the message text).

Returns:

str: The text content of the LLM’s response.

sample_solution(*args, **kwargs)

Generate or mutate a solution using the language model.

Args:

session_messages: Conversation history for the LLM. parent_ids: Identifier(s) of parent solutions. HPO: If True, attempt to extract a configuration space. base_code: Existing code to patch when diff_mode is True. diff_mode: When True, interpret the LLM response as a unified

diff patch to apply to base_code rather than full source code.

Returns:

tuple: A tuple containing the new algorithm code, its class name, its full descriptive name and an optional configuration space object.

Raises:

NoCodeException: If the language model fails to return any code. Exception: Captures and logs any other exceptions that occur during the interaction.

set_logger(logger)

Sets the logger object to log the conversation.

Args:

logger (Logger): A logger object to log the conversation.

class llamea.llm.Ollama_LLM(model='llama3.2', **kwargs)

Bases: LLM

query(session_messages, max_retries: int = 5, default_delay: int = 10)

Sends a conversation history to the configured model and returns the response text.

Args:

session_messages (list of dict): A list of message dictionaries with keys

“role” (e.g. “user”, “assistant”) and “content” (the message text).

Returns:

str: The text content of the LLM’s response.

class llamea.llm.OpenAI_LLM(api_key, model='gpt-4-turbo', temperature=0.8, **kwargs)

Bases: LLM

A manager class for handling requests to OpenAI’s GPT models.

query(session_messages, max_retries: int = 5, default_delay: int = 10)

Sends a conversation history to the configured model and returns the response text.

Args:

session_messages (list of dict): A list of message dictionaries with keys

“role” (e.g. “user”, “assistant”) and “content” (the message text).

Returns:

str: The text content of the LLM’s response.