Create OpenAI client for optillm proxy with local inference

Lazy initialization of OpenAI client

Create a chat completion with OpenAI-compatible parameters

Create completion with appropriate timeout handling

Evaluate if the current research is complete or needs more information Returns (is_complete, list_of_missing_aspects)

Final check to determine if consensus has been reached

Generate text using MLX

Generate focused search queries to address missing aspects

Extract only the content after </think> tag Args: text: Text with thinking tags Returns: Content after </think> tag, or

Get the answer type for a problem

Get the expected answer for a problem

Get the index of the last processed problem.

Get math test messages for reasoning validation

Get the current metrics.

Get list of adapter IDs loaded for a specific model.

Categorize modules by their primary focus.

Get specific modules by their IDs.

Get all steering vectors for a specific reasoning pattern. Args: pattern: The reasoning pattern

Find strategies similar to a query using TF-IDF similarity.

Get statistics about conversation logging

Async wrapper for solution improvement

Determine if a query is high complexity. Args: text: The query text threshold: Confidence threshold

Check if MLX is available (macOS only)

Check if model should use MLX

Check if adding this token creates a thought switch sequence.

Run iterative improvement on solutions that failed verification

Log an error using the global logger instance

Log a provider call using the global logger instance

Mock batch processor that returns simple responses

Helper function to normalize ordered tuples/lists of numbers.

Anonymize the input text.

Optimize generation parameters based on prompt characteristics

Process messages with AutoThink's controlled thinking. Args: messages: List of message dictionaries

Process a batch of prompts with all optimizations

Process a batch of requests using true batching when possible Args: batch_requests: List of request data dic

Reset internal state for new trace.

Deep Research plugin implementing TTD-DR (Test-Time Diffusion Deep Researcher) This plugin orchestrates web search, URL fetching, and me

Main entry point for the Deep Think plugin. Combines SELF-DISCOVER reasoning structure discovery with uncertainty-routed chain-of-t

Generic majority voting implementation.

Main plugin execution function.

Main Chain of Code execution function.

Main entry point for the GenSelect plugin. Generates multiple candidate solutions and uses LLM comparison to select the best one.

Main proxy plugin entry point. Supports three usage modes: 1. Standalone proxy: model="proxy-gpt-4" 2. Wrapping approach: extra_

Plugin entry point for System Prompt Learning. Args: system_prompt: The system prompt initial_query: The user's query

Web search plugin that uses Chrome to search Google and return results Args: system_prompt: System prompt for the conversation

Main plugin execution function called by OptILLM Args: system_prompt: System prompt initial_query: User query cl

Executes the full LongCePO multi-stage pipeline to answer a complex query from long context. The pipeline includes: - Normalizing the

Run a performance comparison between sequential and batch processing This function can be called separately for benchmarking

Perform a search using the managed browser session with automatic recovery

Log and forward notifications