config.example.yaml

# Red Team AI Benchmark Configuration Example
# Copy to config.yaml and customize

# LLM Provider Settings
provider:
  name: ollama # ollama, lmstudio, openrouter
  endpoint: http://localhost:11434
  # api_key_env: OPENROUTER_API_KEY  # For OpenRouter
  # default_model: llama3.1:8b

# Scoring System Configuration
scoring:
  method: keyword # keyword, semantic, hybrid, llm_judge

  # Semantic scoring settings (requires sentence-transformers)
  semantic_model: Alibaba-NLP/gte-large-en-v1.5
  semantic_weight: 0.7
  keyword_weight: 0.3

  # LLM Judge settings (requires OpenRouter API key)
  llm_judge_model: anthropic/haiku-4.5

  # Hybrid scoring gray zone (when to use LLM for uncertain cases)
  gray_zone_low: 0.30
  gray_zone_high: 0.70
  use_llm_in_gray_zone: true

# Export Settings
export:
  formats:
    - json
    - csv
  output_dir: ./results
  include_response: true

# Prompt Optimization (for censored responses)
optimization:
  enabled: false
  optimizer_model: llama3.3:70b
  # optimizer_endpoint: http://localhost:11434  # Optional custom endpoint
  max_iterations: 3
  strategies:
    - role_playing
    - technical
    - few_shot
    - cve_framing

# Benchmark Settings
questions_file: benchmark.json
answers_file: answers_all.txt
rate_limit_delay: 1.5
max_tokens: 1024
temperature: 0.2

# Langfuse Observability
langfuse:
  enabled: false  # Set to true to enable tracing
  # secret_key: sk-lf-xxx
  # public_key: pk-lf-xxx
  host: http://localhost:3000