Skip to content

CLAUDE.md

Latest commit

 

History

History
80 lines (58 loc) · 2.67 KB

CLAUDE.md

File metadata and controls

80 lines (58 loc) · 2.67 KB

CLAUDE.md - NVIDIA NeMo Guardrails (TrustyAI fork)

Overview

NVIDIA NeMo Guardrails — toolkit for adding programmable guardrails to LLM apps. This is the trustyai-fork (trustyai-explainability/NeMo-Guardrails) with Red Hat / TrustyAI customisations.

Repository layout

nemoguardrails/              # main package
  rails/llm/llmrails.py      # LLMRails — core engine
  rails/llm/config.py        # RailsConfig
  server/api.py              # FastAPI server (+ fork-specific /v1/guardrail/checks endpoint)
  actions/llm/utils.py       # llm_call(), header forwarding
  context.py                 # context vars (request headers, runtime auth registry)
  logging/callbacks.py       # LangChain callback handler (invocation param logging)
  library/                   # built-in rails (30+ guardrail types)
  colang/v1_0/               # Colang v1
  colang/v2_x/               # Colang v2
tests/                       # pytest tests
scripts/                     # entrypoint, filter_guardrails, model discovery
Dockerfile.server            # UBI9 multi-stage build (fork-specific)

Git remotes

  • upstream — NVIDIA/NeMo-Guardrails (original)
  • trustyai-fork — trustyai-explainability/NeMo-Guardrails
  • midstream — opendatahub-io/NeMo-Guardrails
  • downstream — red-hat-data-services/NeMo-Guardrails

Default branch with the latest changes: develop

Build and install python package

# Install with Poetry (Python 3.10-3.13)
poetry install --all-extras

Running the server locally

# Start with a config directory (each subdirectory is a config_id)
nemoguardrails server --config /path/to/configs --port 8000

# With a default config and verbose logging
nemoguardrails server --config /path/to/configs --default-config-id my-config --verbose

Running tests

# All tests
pytest tests/

# Specific test file
pytest tests/test_header_forwarding.py -v

# Fork-specific tests
pytest tests/test_guardrail_checks_api.py -v

CI

PR workflows (.github/workflows/pr-tests.yml → _test.yml):

  1. poetry check --lock — lock file must be in sync with pyproject.toml
  2. pytest across Python 3.10, 3.11, 3.12, 3.13 (coverage on 3.11)

Fork-specific (security.yml): Trivy + Bandit scans on PRs and pushes to develop.

Key fork changes

  • /v1/guardrail/checks endpoint — check messages against rails without LLM generation
  • Header forwarding — X-* headers forwarded to LLM providers, auth token redaction in logs
  • Dockerfile.server — UBI9 multi-stage build with baked-in models
  • scripts/filter_guardrails.py — open/closed source guardrail filtering via profiles
  • requirements.txt — pinned deps via uv pip compile