config.yaml

# Global configuration for the AutoInterp Agent Framework

# Provider-specific configurations
providers:
  anthropic:
    api_version: "2023-06-01"
  openai:
    api_version: "2020-10-01"  # Current REST API version
    organization_id: null       # Must be set via env var OPENAI_ORGANIZATION_ID
  openrouter:
    base_url: "https://openrouter.ai/api/v1"

# Framework settings
framework:
  name: "AutoInterp Agent Framework"
  version: "0.1.0"
  log_level: "INFO"
  log_file: "autointerp.log"
  log_message_limit: 500  # Maximum length for log messages
  execute_full_workflow: true  # Whether to run the complete workflow automatically

# Task configuration
task:
  description: ""  # Task description will be provided by user or auto-generated

# Model to analyze
model:
  name: "meta-llama/Llama-3.2-1B-Instruct"  # HuggingFace model ID
  tokenizer: "meta-llama/Llama-3.2-1B-Instruct"
  cache_dir: "hf_models"  # Local cache directory


# Analysis settings
analysis:
  execution:
    timeout: 120  # seconds
    max_memory: 4000  # MB
    sandbox: true
  executor_type: "sandbox"
  output_format: "json"
  timeout_per_analysis: 300
  max_iterations: 3      # Maximum number of analysis cycles per question
  confidence_threshold: 0.85  # Continue analysis if confidence_threshold > confidence

# Execution settings
execution:
  mode: "sandbox"
  timeout: 300
  max_retries: 5
  max_memory_mb: 4096
  force_clean_venv: false  # Force clean virtual environment on startup (disabled when using existing venv)
  use_system_python: false  # Set to true to use system Python instead of virtual environment
  use_existing_venv: false  # Set to true to use an existing virtual environment (requires existing_venv_path)
  existing_venv_path: ""  # Path to existing virtual environment (or use --venv CLI flag)
  fail_on_max_retries: false  # If true, shutdown entire system when max_retries exceeded; if false, continue to next analysis
  docker:
    image: "python:3.10-slim"  # Default container image for sandbox execution
    use_gpu: false             # Set true to pass host GPUs into the container (requires nvidia-container-runtime)
    cache_dir: "~/.autointerp/docker_cache"  # Persistent cache for installed packages
    extra_args: []             # Additional docker run arguments
    env: []                    # Additional environment variables to pass through (e.g., ["CUSTOM_ENV"])

# Visualization settings
visualization:
  default_format: "png"
  dpi: 300
  save_dir: "projects/{project_id}/visualizations"
  attention_maps: true
  attention_flow: true
  neuron_activations: true
  comparative_analysis: true

# Reporting settings
reporting:
  default_format: "markdown"  # Options: markdown, jupyter, html, pdf
  save_dir: "projects/{project_id}/reports"

# Computational resources
resources:
  max_gpu_memory: 8000  # MB
  batch_size: 8
  precision: "fp16"  # Options: fp32, fp16, bf16

# Default paths
paths:
  data: "data"
  models: "models"
  projects: "projects"
  create_missing: true
  cleanup_old: false


# Default output directories - all should use project-specific paths
output_directories:
  analysis_scripts: "projects/{project_id}/analysis_scripts"
  analysis_results: "projects/{project_id}/analysis_results"
  visualizations: "projects/{project_id}/visualizations"
  reports: "projects/{project_id}/reports"
  evaluation_results: "projects/{project_id}/evaluation_results"



# Default LLM configuration
llm:
  provider: "anthropic"
  model: "claude-sonnet-4-5"
  temperature: 0.7
  max_tokens: 3000
  timeout: 120

# Agent-specific configurations
agents:
  # Question Generation Agent
  question_generator:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.8
      max_tokens: 1000
      timeout: 60
  
  # Question Prioritization Agent
  question_prioritizer:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.667
      max_tokens: 1000
      timeout: 60
  
  # Analysis Planning Agent
  analysis_planner:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.7
      max_tokens: 3000
      timeout: 90
      
  # Analysis Generation Agent
  analysis_generator:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.7
      max_tokens: 12000
      timeout: 120

  # Visualization Planning Agent
  visualization_planner:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.7
      max_tokens: 3000
      timeout: 120

  # Visualization Generation Agent
  visualization_generator:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.7
      max_tokens: 12000
      timeout: 120

  # Analysis Evaluation Agent
  evaluator:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.7
      max_tokens: 31000
      timeout: 180

  # Visualization Evaluation Agent
  visualization_evaluator:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.3
      max_tokens: 2000
      timeout: 120

  # Report Generation Agent
  reporter:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.7
      max_tokens: 8000
      timeout: 120

  # Title Generation Agent
  title_generator:
    llm:
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      temperature: 0.7
      max_tokens: 200
      timeout: 30

# Provider configurations are now dynamically applied at runtime
# Use the interactive provider selection menu or set --config with overrides