guardex.yaml.example

# guardex.yaml - local ML engine settings for GuardEx in-process mode.
# Copy this file to guardex.yaml in your project root; it is auto-loaded
# from the current working directory. All fields have safe defaults.
#
# This is NOT the file for GuardExPolicy.from_yaml(). That call reads a
# flat policy file (pii_action, blocked_categories, topic_scope, ...) -
# see docs/guides/configuration.md.
#
# Environment variables (GUARDEX_* prefix) override this file.
# Example: GUARDEX_CACHE_DIR=/tmp/guardex

# ── Model selection ────────────────────────────────────────────────────────────
models:
  # HuggingFace repo for the ONNX safety classifier
  # Swap in your own fine-tuned model - must be a compatible ONNX classifier
  classifier: AtliQ-Technologies/toxicity-fast-onnx

  # HuggingFace repo for the GLiNER PII detector
  pii: nvidia/gliner-pii

  # Sentence-transformers model for topic scope + grounding
  embeddings: sentence-transformers/all-MiniLM-L6-v2

  # Ollama endpoint for optional LlamaGuard deep-safety layer
  # Requires: ollama pull llama-guard3:1b && ollama serve
  ollama_url: http://localhost:11434
  ollama_model: llama-guard3:1b

# ── Cache ──────────────────────────────────────────────────────────────────────
# Directory for downloaded model files.
# First-install footprint: ~250 MB (GLiNER + sentence-transformers + ONNX).
# Grounding adds ~700 MB on top, only when GUARDEX_GROUNDING_ENABLED=1.
# Override with: GUARDEX_CACHE_DIR=/path/to/cache
cache_dir: ~/.cache/guardex

# ── Policy defaults ────────────────────────────────────────────────────────────
# fail_open is the only policy key this file reads. Everything else (PII
# entities, action, threshold, blocked categories, topic scope) is set on
# GuardExPolicy in code or in a separate policy YAML loaded via
# GuardExPolicy.from_yaml() - see docs/guides/configuration.md.
policy:
  # Whether to silently pass through on ML errors instead of raising
  fail_open: false