Skip to content

configuration.md

Latest commit

 

History

History
157 lines (115 loc) · 5.69 KB

configuration.md

File metadata and controls

157 lines (115 loc) · 5.69 KB

Configuration Customization

AutoGluon Assistant uses YAML configuration files to control its behavior. This tutorial explains the configuration system and how to customize it for your specific needs.

Configuration Overview

The configuration system is based on hierarchical YAML files that control:

  1. General execution settings
  2. LLM provider settings
  3. Agent behaviors and parameters
  4. Resource utilization
  5. Data handling preferences

Basic Structure

A configuration file has this general structure:

# General settings
per_execution_timeout: 86400
create_venv: false
# ... other general settings

# Default LLM Configuration
llm: &default_llm  # The anchor defines reusable settings
  provider: bedrock
  model: "us.anthropic.claude-3-7-sonnet-20250219-v1:0"
  # ... other LLM settings

# Agent-specific configurations
coder:
  <<: *default_llm  # This merges all settings from default_llm
  multi_turn: True  # Override specific settings

Configuration Parameters

General Settings

Parameter Description Default
per_execution_timeout Maximum execution time (seconds) for code execution 86400
create_venv Whether to install additional packages in created conda environment false
condense_tutorials Whether to use condensed tutorials true
use_tutorial_summary Whether to use tutorial summary as retrieval key true
continuous_improvement Continue iterations after finding a valid solution false
optimize_system_resources Optimize resource usage during execution false
cleanup_unused_env Remove unused environments after execution true

Data Perception Settings

Parameter Description Default
max_file_group_size_to_show Minimum number of similar files to show as a group 5
num_example_files_to_show Number of example files to display for each type 1
max_chars_per_file Maximum characters to display per file 768
max_user_input_length Maximum length of user input to process 2048
max_error_message_length Maximum length of error messages to include 2048

Tutorial Settings

Parameter Description Default
num_tutorial_retrievals Number of tutorial segments to retrieve 30
max_num_tutorials Maximum number of tutorials to include 5
max_tutorial_length Maximum length of all tutorial contents 32768

LLM Settings

Parameter Description Default
provider LLM provider to use (bedrock, openai, anthropic, sagemaker) bedrock
model Specific model name for the selected provider
max_tokens Maximum token limit for model responses 65535
proxy_url Optional proxy URL for API requests null
temperature Controls randomness (0.0-1.0, lower = more deterministic) 0.1
top_p Nucleus sampling parameter for token selection 0.9
verbose Whether to log detailed information about LLM interactions true
multi_turn Whether to use multi-turn conversation with the LLM across different iterations false
template Optional custom prompt template null
add_coding_format_instruction Add explicit coding format instructions false

Agent-Specific Settings

AutoGluon Assistant uses specialized agents for different tasks. Each inherits the default LLM settings but can have custom overrides:

Coder Agent

Generates code based on requirements and context.

coder:
  <<: *default_llm
  multi_turn: True  # Enable multi-turn conversation across iterations for iterative coding

Executer Agent

Runs code and evaluates execution results.

executer:
  <<: *default_llm
  max_stdout_length: 8192  # Maximum length of stdout to capture
  max_stderr_length: 2048  # Maximum length of stderr to capture

Reader Agent

Analyzes and understands input data files.

reader:
  <<: *default_llm
  details: False  # Whether to include detailed file information

Task Descriptor Agent

Describes tasks based on input data.

task_descriptor:
  <<: *default_llm
  max_description_files_length_to_show: 1024         # Max length to show
  max_description_files_length_for_summarization: 16384  # Max length for summarization

Other Specialized Agents

  • error_analyzer: Analyzes execution errors and suggests fixes
  • retriever: Retrieves relevant tutorials
  • reranker: Re-ranks and selects top retrieved tutorials
  • description_file_retriever: Retrieves information from description files
  • tool_selector: Selects the appropriate ML Library based on requirements

Use a Custom Configuration

You can create and use a custom configuration file by:

  1. Create a new YAML file, e.g., my_custom_config.yaml
  2. Run with your custom config: mlzero -i <input_folder> -c my_custom_config.yaml

Best Practices

  1. Start Simple: Begin with minimal customizations and add more as needed
  2. Test Incrementally: Test changes one at a time to understand their impact
  3. Inheritance Issues: If you modify settings in the llm section, you must update agent references to it. The YAML anchor/alias inheritance (<<: *default_llm) is a one-time static reference, not dynamic. When you change the main llm config, agents won't automatically inherit these changes unless you explicitly update their references.

Need Help?

If you run into any issues:

  1. Check the API Reference for detailed documentation
  2. Browse the examples for common use cases (coming soon)
  3. Visit our GitHub repository for support