ecoguard.yaml

Rama main · réplica de lectura del repositorio original.

# Eco-Guard GitOps Configuration
# Place this file at the root of your project.
# It hot-reloads every 10 seconds — no restart needed.
# git push to reconfigure your inference gateway.

# Model routing rules: prompt pattern → model path
routes:
  "code": "/app/models/codellama.gguf"
  "translate": "/app/models/tinyllama.gguf"
  "math": "/app/models/mistral.gguf"

# Model pricing (USD per 1000 tokens)
models:
  llama-3-8b:
    input_cost_per_1k: 0.00006
    output_cost_per_1k: 0.00006
  gpt-4o:
    input_cost_per_1k: 0.0025
    output_cost_per_1k: 0.01
  custom-model:
    input_cost_per_1k: 0.001
    output_cost_per_1k: 0.002

# Preloaded API keys (hashed at startup)
api_keys: []

# Rate limiting configuration
rate_limits:
  requests: 100
  window_seconds: 60

# Guardrails (choose which to enable)
guardrails:
  enabled:
    - pii              # Redact PII from prompts
    - injection        # Block prompt injection attempts
    - content_safety   # Flag harmful content

# Fallback chain (used when a model fails)
fallback:
  steps:
    - model: "llama-3-8b"
      backend_url: "http://localhost:8080"
      timeout_seconds: 30
    - model: "gpt-4o"
      backend_url: "http://openai-proxy:8001"
      timeout_seconds: 15
    - model: "llama-3-8b"
      timeout_seconds: 10