NVIDIA NeMo Guardrails v0.24.0
NVIDIA NeMo Guardrails v0.24.0 expands IORails to support 59 action-backed input and output surfaces using a unified RailOutcome contract, manifests, and shared outbound HTTP, alongside several breaking API updates.
Why this signal matters
NVIDIA NeMo Guardrails v0.24.0 substantially aligns IORails and LLMRails by allowing IORails to execute 59 of the 67 library action-backed surfaces. This unification is driven by three architectural additions: the engine-neutral RailOutcome contract (allow, block, transform), typed rail manifests for feature and privacy discovery, and a centralized outbound HTTP client managing pooling, retries, and metrics. The release adds F5 Guardrails integration, /v1/health endpoints, multiple namespaced self-check rails, and output-checking via /v1/checks. It also enforces fail-closed behavior on rail action streaming failures, distinguishes policy blocks from execution failures, and improves error mappings for OpenAI-compatible server interfaces. Significant breaking changes affect custom action decorators, IORails call signatures, dependency extras, /v1/checks invocation, and strict OpenAI message schema validation.
Actionable summary
NeMo Guardrails 0.24.0 introduces RailOutcome (allow/block/transform), typed rail manifests, and a shared HTTP client. It includes breaking changes for custom actions (removal of output mapping), IORails parameter handling, /v1/checks configuration loading, Colang 1 action naming, and OpenAI chat completions payload validation.
- Agent usefulness
- 89/100
- Confidence
- 96%
- Canonical data
- JSON + Markdown
What builders should check
- Refactor custom actions to return an explicit RailOutcome and remove the deprecated @action(output mapping=...) argument.
- Update IORails generate/generate async calls to pass message lists using the keyword argument messages= instead of positionally, and handle GenerationResponse return types when generation options are provided.
- Install transformers and torch directly if utilizing local classifiers, as the hf-classifier extra has been removed.
- Update /v1/checks requests to reference a server-loaded config id or rely on the default configuration instead of passing inline configurations.