Main entry point for enhanced audio understanding. Args: audio_path: Path to audio file (local or URL) quest

Clear all memory items

Placeholder for HTML to Markdown conversion function In the original class, this would call self._convert()

Estimate completion time based on overall progress rate from all completed tasks

Estimate token count for text

Format tool execution results to be fed back to LLM as user messages. Only includes necessary information (results or errors).

Get the log directory for the current benchmark and model.

- [Doc: llama.pdf, Page 34] - [vision_understanding_advanced: image2.jpg, reasoning] - [RAG-1] Returns: Dict with k

Args: Returns: Dict with all results

Configure logging for batch evaluation.

Update message history with tool calls data (llm client specific)

Add file information to FolderContents.

Calculates and adds average values to a summary block.

Call appropriate audio service based on provider. Args: audio_path: Path to audio file question: Optional qu

Call appropriate video service based on provider. Args: video_path: Path to video file question: Question ab

Call the VLLM provider API. Supports multiple providers: - openai: GPT-4V, GPT-4o, etc. - anthropic: Claude

Create LLM client with robust timeout and connection settings

Send message to OpenAI API. :param system_prompt: System prompt string. :param messages_history: Message history list.

Encode audio file to base64 and determine format.

Use tiktoken to estimate the number of tokens in text

Use tiktoken to estimate the number of tokens in text

Extract core identifiers (e.g. paper abbreviations) from a title.

Extract task ID from filename

Get audio duration in seconds. Tries to use wave (for .wav), then falls back to mutagen (for mp3, etc).

Determine the appropriate audio file extension from URL or content type. Args: url: The URL of the audio file content_type:

Get embedding for a single text with retry logic for rate limiting.

Get the embedding dimension from stored embeddings or use default.

Get the latest task file for each task ID in a run directory

Extract metadata from media files using exiftool Args: local_path: Path to the media file Returns: Dictionary of metada

Determine if LLM judge result indicates correct answer

Check if a task is completed based on its data

Load title list from useful_search.json or useful_search.jsonl.

Remove tool results from messages

Save retrieval result to the log file.

Check if we should block scraping of Hugging Face datasets/spaces. :param tool_name: The name of the tool being called :param

Send show_error event

Updates a summary block with data from a single result.

Validate if the given path is allowed for RAG operations. Returns: Tuple of (is_valid, error_message_or_corrected_path)

Verify claims against image content using Gemini multimodal model.

Configure the miroflow_agent logger with consistent formatting

Calculate how well the content matches the insight. Returns: - score: 0.0 to 1.0 - match_type: 'exact', 'partial', 'keyword', 'n

Check if this item can be further compressed

Close database connection.

Close the session and connection.

Define and return MCP server configuration list

Creates and initializes the core components of the agent pipeline. Args: cfg: The Hydra configuration object. Returns:

Search with document diversity - ensures results come from multiple documents. This method prioritizes getting results from

Check if current message_history + summary_prompt will exceed context If it will exceed, remove the last assistant-user pair and retu

Executes the full pipeline for a single task. Args: cfg: The Hydra configuration object. task_description: The description o

Extract text content from a message, handling different content formats.

Extract a JSON object from text, handling code blocks and truncation.

Extract text from LLM response, excluding <use_mcp_tool> tags. Stop immediately when this opening tag is encountered.

Find all run directories in the target path

Format final summary information, including answers and token statistics

Parse the JSON log file and generate a human-readable turn-by-turn log. Args: log_file_path: Path to the JSON log file o

Connect to all configured servers and get their tool definitions. Returns a list suitable for passing to the Prompt generator.

Get set of task numbers that have already been completed. Directory structure: result/<run_batch>/<dataset>/<context_size>/<model>/<task_numb

Get long-context content by title with fuzzy matching.

Determine the results directory based on folder path structure. For paths like 'data/bench_case1104/001', returns 'results/bench_case110