Language Model (LLM) Utilities for ML Research Tools.

This module provides a set of utilities for interacting with Large Language Models:

  1. LLMClient Class: Complete client for LLM interactions with: - Configuration management (presets, tiers) - Automatic retries with exponential backoff - Result caching - Simple and chat-based interfaces

  2. Factory Function: Easy client creation through create_llm_client

Example

# Create client with default preset
client = create_llm_client()
response = client.simple_call(
    text="Summarize the following paper: [paper text]",
    system_prompt="You are a helpful academic assistant."
)

# Use a specific preset
client = create_llm_client(preset="premium")
response = client.simple_call(
    text="Explain this complex concept...",
    system_prompt="You are a helpful academic assistant."
)

# Chat interface
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Hello, how are you?"},
    {"role": "assistant", "content": "I'm doing well, thank you! How can I help you today?"},
    {"role": "user", "content": "Can you explain quantum computing?"}
]
client = create_llm_client(tier="premium")
response = client.call(messages=messages)

# For raw OpenAI client access
openai_client = create_llm_client().get_openai_client()

# Generate parameters for OpenAI API calls
params = generate_completion_params(
    config=config,
    messages=messages,
    stream=True
)
class ml_research_tools.core.llm_tools.Message[source]#

Bases: TypedDict

Type definition for a chat message.

role: Literal['system', 'user', 'assistant', 'tool']#
content: str#
ml_research_tools.core.llm_tools.get_llm_config(*, preset=None, tier=None, api_key=None, base_url=None, model=None, temperature=None, top_p=None, max_tokens=None, retry_attempts=None, retry_delay=None, config=None)[source]#

Get an LLM configuration by resolving preset/tier and applying overrides.

This factory function selects the appropriate LLM configuration based on: 1. Preset name or tier 2. Individual parameter overrides 3. Default configuration if nothing else is specified

Parameters:
Return type:

LLMConfig

Returns:

LLMConfig object with all parameters resolved

Raises:

ValueError – If no valid configuration can be determined

class ml_research_tools.core.llm_tools.LLMClient(*, preset=None, tier=None, api_key=None, base_url=None, model=None, temperature=None, top_p=None, max_tokens=None, retry_attempts=None, retry_delay=None, config=None, redis_cache=None)[source]#

Bases: object

Client for interacting with Language Models with preset configurations and caching.

This class provides a unified interface for making LLM API calls with: - Configuration management (presets, tiers, parameter overrides) - Automatic retries with exponential backoff - Result caching - Simple and chat-based interfaces

config#

The LLM configuration to use for API calls

Initialize an LLM client with the specified configuration.

Parameters:
__init__(*, preset=None, tier=None, api_key=None, base_url=None, model=None, temperature=None, top_p=None, max_tokens=None, retry_attempts=None, retry_delay=None, config=None, redis_cache=None)[source]#

Initialize an LLM client with the specified configuration.

Parameters:
property model#

Get the model name from the configuration.

get_openai_client()[source]#

Get the raw OpenAI client.

Return type:

OpenAI

Returns:

The underlying OpenAI client instance

simple_call(text, system_prompt, *, model=None, temperature=None, top_p=None, max_tokens=None, prefix='llm', use_cache=True)[source]#

Call an LLM with a simple system prompt + user text pattern.

Parameters:

  • text (str) – The user text to process

  • system_prompt (str) – Instructions for the LLM (system message)

  • model (Optional[str]) – Model name override

  • temperature (Optional[float]) – Temperature override

  • top_p (Optional[float]) – Top-p override

  • max_tokens (Optional[int]) – Max tokens override

  • prefix (str) – Prefix for cache keys

  • use_cache (bool) – Whether to use caching for this call

Return type:

str

Returns:

The LLM response text

call(messages, *, model=None, temperature=None, top_p=None, max_tokens=None, prefix='llm_chat', use_cache=True)[source]#

Call an LLM API with a complete chat history.

Parameters:

  • messages (List[Message]) – List of chat messages

  • model (Optional[str]) – Model name override

  • temperature (Optional[float]) – Temperature override

  • top_p (Optional[float]) – Top-p override

  • max_tokens (Optional[int]) – Max tokens override

  • prefix (str) – Prefix for cache keys

  • use_cache (bool) – Whether to use caching for this call

Return type:

str

Returns:

The LLM response text

ml_research_tools.core.llm_tools.create_llm_client(*, preset=None, tier=None, api_key=None, base_url=None, model=None, temperature=None, top_p=None, max_tokens=None, retry_attempts=None, retry_delay=None, config=None, redis_cache=None)[source]#

Create an LLMClient instance with the specified configuration.

This is a factory function that creates an LLMClient with the appropriate configuration.

Parameters:
Return type:

LLMClient

Returns:

An initialized LLMClient instance

ml_research_tools.core.llm_tools.generate_completion_params(*, llm_client, **additional_params)[source]#

Generate parameters for completion API calls based on configuration.

This function resolves LLM configuration based on presets/tiers and returns a dictionary of parameters suitable for passing to the OpenAI completion API calls.

Return type:

Dict[str, Any]

Returns:

Dictionary of parameters for OpenAI API calls

Parameters:
Contents