Check for unused code in the parser.

Close the session and connection.

Copy GAIA-Text-103 tasks from GAIA validation logs Args: gaia_logs_dir: Directory containing GAIA validation logs

Call LLM to generate a response with optional tool call support. This is the main entry point for LLM interactions. It handles:

Detect the pass_at_k value used in the results directory

Determine conversion method based on provider Args: provider: LLM provider name Returns: 'oai' for OpenAI, 'non-oai' fo

Discover subdirectories inside results_dir that potentially contain a 'benchmark_results.jsonl'. We don't strictly require the subdir name to

End the current sub-agent session

Send end_of_workflow event. Args: workflow_id: The workflow ID to end

Evaluate pass@k accuracy (verification already done in run_single_task)

Example 1: Basic conversion using Python API

Extract evaluation details from a completed task log file. Returns: Dict with num_correct, num_expected, num_excessive, or empty dic

Extract failure experience summary from LLM response text. The text may contain: - <think>...</think> block (thinking content) - Mai

Summarize content using an LLM API. Args: content (str): The content to summarize info_to_extract (str): The specific types

Extract task information from a task file

Extract tool call information from LLM response

Filter google search organic results to remove unnecessary information

Filter google search result content based on environment variables. Args: result_content: The JSON string result from google search

Filter message to remove unnecessary information

Find JSON files from input paths Args: input_paths: List of file paths, directories, or patterns Returns: List of JSON

Find all task JSON files in the extraction directory

Format final summary information, including answers and token statistics. Args: final_answer_text: The final answer text

Generate final answer and handle fallback based on context management settings. Context Management (context_compress_limit > 0) is e

Generate final answer with retry mechanism. Args: system_prompt: System prompt for the LLM message_history:

Generates a summary of benchmark results by reading log files from a directory, calculating total and average trace data, both overall and gr

Get basic information of the task

Get message list for specified session

Get color code based on log level for better visual distinction

Generate a user-friendly summarize prompt for demo mode. This prompt is designed for better user experience, producing well-formatted Ma

Collect current configuration and environment information for logging. Gathers LLM settings, agent configuration, API key availability (mask

Get execution summary information

Get the extra information attached to this response.

Extract llm_provider from JSON file Args: json_file_path: Path to JSON file Returns: llm_provider value or 'unknown' if

Get performance summary information

Get the wrapped response object.

Process scrape result and truncate if too long. Args: result: Raw scrape result string (JSON or plain text) Ret

Get parameters for the specified server

Get spans statistical summary

Get step logs summary statistics

Collects the paths of successful log files from a dataset. This function extracts log file paths of successful records based on the valu

Grade all tasks with concurrent processing

Grade a single task using GAIA-Text-103 evaluator

Guess the MIME type and media category based on the file extension. Returns: Tuple of (mime_type, media_category) where media_catego

Guess the MIME type based on the file extension.

Handle failure when context_compress_limit > 0 (context management enabled). In this mode, the model has multiple chances to retry w

Handle fallback when context_compress_limit == 0 (no context management). In this mode, the model has only one chance to answer.

Check if this response has extra information attached.

Install dependencies (recommended to use uv)

Check if the given object is an ErrorBox instance.

Check if google_search result has empty organic results. This indicates a poor search query that should be retried. Args:

Check if a shape is a picture

Check if the given object is a ResponseBox instance.

Check if the scrape result is an error

Check if a shape is a table

Load the full MiroFlow configuration using Hydra, similar to how benchmarks work.

Load benchmark tasks from metadata.jsonl Args: limit: Maximum number of tasks to load (None for all) Returns:

Return True if the given path either: - exists and is a directory, OR - does not exist but looks like a directory (e.g., ends with '/

Main function

Main function

Main function

Main function

Main function

Main function

Compute the majority-vote prediction for a list of candidate predictions. Tie-breaking rules (deterministic): 1) Highest frequency win

Make HTTP request to Serper API with retry logic.

Make HTTP request to Serper API with retry logic.

Make request to Tencent Cloud SearchPro API with retry logic.