NVIDIA-NeMo/Guardrails v0.24.0 on GitHub

What's Changed

NeMo Guardrails 0.24.0 substantially expands the built-in rails available through IORails. IORails can now execute 59 of the library's 67 action-backed input and output surfaces, including community integrations and rails that transform content, while sharing the same underlying rail implementations with LLMRails.

Three related foundations make this possible:

RailOutcome: decide once, render many. A rail action returns one engine-neutral allow, block, or transform result that either LLMRails or IORails can enforce.
Rail manifests: declare, don't infer. Built-in rails describe their configuration, actions, execution surfaces, requirements, and privacy properties so both engines can discover and use them without maintaining separate support lists.
Shared outbound HTTP. Built-in integrations now use a common client boundary for connection pooling, retries, TLS, error handling, lifecycle management, and privacy-safe tracing and metrics.

This release also adds F5 Guardrails integration, server health endpoints, multiple namespaced self-check rails, output checking through /v1/checks, and extensive correctness improvements for streaming, provider errors, reasoning responses, and OpenAI-compatible server requests.

Breaking changes

Custom actions using @action(output_mapping=...) must remove that argument and return an explicit RailOutcome from the action.
Install transformers and torch directly instead of using the removed hf-classifier extra.
IORails calls that supply generation options now return GenerationResponse, and message lists must be passed with messages= rather than positionally.
Custom Colang 1 flows using the former space-separated Cleanlab, Fiddler, or GCP action names must use their snake_case names.
/v1/checks now selects a server-loaded configuration through config_id or the server default instead of accepting inline configuration.
Chat Completions requests must use supported role-specific OpenAI message shapes; internal event payloads, unexpected fields, and audio requests are rejected.

🚀 Features

(actions) Add the engine-neutral RailOutcome allow, block, and transform contract (#2150)
(rails) Add the typed rail manifest contract (#2157)
(http) Add the canonical outbound HTTP client and request lifecycle (#2209, #2210)
(http) Add outbound client tracing and metrics (#2219)
(iorails) Execute blocking rail outcomes directly (#2264)
(iorails) Apply transforming rail outcomes to input and output content (#2288)
(iorails) Add message checks (#2059)
(iorails) [breaking] Return GenerationResponse from non-streaming generate() and generate_async() calls when options are supplied; pass message lists with messages= instead of positionally (#2178)
(self-check) Run multiple self-check rails with per-rail namespaced task prompts (#1874, #2175)
(server) Add output-rail checking mode to /v1/checks (#2205)
(server) Add /v1/health and /healthz health-check endpoints (#2169)
(llm) Add shared model telemetry and an instrumented model decorator (#2214)
(library) Add F5 Guardrails integration (#2105)

🐛 Bug Fixes

(streaming) Fail closed when rail actions fail (#2152)
(library) [breaking] Make bundled manifest flows portable across Colang versions; custom Colang 1 flows using legacy space-separated Cleanlab, Fiddler, or GCP action names must switch to snake_case names (#2185)
(actions) Make RailDecision JSON serializable (#2194)
(iorails) Preserve normalized usage and provider metadata in include_metadata=True streams (#2198)
(iorails) Honor model-level default_headers in non-streaming and streaming HTTP requests (#2220)
(server) [breaking] Remove inline configuration from /v1/checks; select a server-loaded configuration with config_id or use the server default (#2228)
(server) Return OpenAI-compatible HTTP error envelopes while preserving provider status, code, parameter, and retry metadata (#1832)
(server) Return HTTP 400 for thread IDs without a datastore and preserve the configured main model API-key field during request model injection (#2240)
(server) Honor request-level stop parameters and reject chat messages that omit a role before dispatch (#2266)
(llm) Preserve streaming usage on terminal chunks (#2295)
(content-safety) Surface response parsing errors instead of silently allowing malformed results (#2294)
(checks) Reject unsatisfiable rail_types with HTTP 422 instead of silently passing (#2276)
(iorails) Propagate default_query parameters in model HTTP requests (#2296)
(server) Preserve configured main model fields when a request specifies a model (#2298)
(iorails) Map provider, timeout, connection, response, and streaming failures to safe client errors while preserving HTTP status and retry metadata (#2306)
(server) [breaking] Validate Chat Completions messages by role and reject internal event payloads, unexpected fields, and unsupported audio requests; send only supported OpenAI-compatible message shapes (#2311)
(server) Inline <think> tags when reasoning content is present (#2316)
(iorails) Handle reasoning-only non-streaming responses (#2317)
(iorails) Distinguish policy blocks from rail execution failures so callers can identify retryable provider outages (#2318)

💼 Other

(dependencies) [breaking] Remove the hf-classifier install extra and its packages from all; install transformers and torch directly for the local classifier backend (#2137)

🚜 Refactor

(actions) [breaking] Migrate built-in rail actions and streaming bypasses to RailOutcome and remove @action(output_mapping=...); custom rail actions using output_mapping must return explicit RailOutcome decisions (#2151)
(http) Migrate built-in integration request helpers and vendor actions to the canonical managed HTTP client lifecycle (#2211, #2212)
(iorails) Migrate IORails to manifest-compiled RailOutcome actions with shared model and HTTP dependencies (#2241, #2246, #2253, #2261, #2286)

📚 Documentation

(actions) Document RailOutcome-based rail actions (#2258)
(rails) Document rail manifests and action-backed surfaces (#2259)
(http) Document canonical outbound HTTP clients (#2260)
(iorails) Document per-surface LLMRails and IORails compatibility, fallback conditions, and configuration-dependent rail support (#2330)

New Contributors

@ciarancourtney made their first contribution in #2105
@fcanogab made their first contribution in #2192
@yixinh-nv made their first contribution in #2307

Full Changelog: v0.23.0...v0.24.0