Cette page a été traduite automatiquement. Lire l’original en anglais.

Component

core/security/nemo-guardrails

NVIDIA NeMo Guardrails' server as a guard in a model's path: a service speaking the OpenAI chat-completions API that checks each request against its input rail, forwards what it admits to the upstream model API in infra.upstream_endpoint, and returns the upstream's answer, streamed or not. A request the rail refuses is answered by the guard itself with an ordinary completion, and nothing of it reaches the upstream. The rail sends the latest user message to the detector wired on the detector socket and refuses the request when the detector reports one of values.blocked_labels at or above values.threshold; a detector that cannot be reached, or answers with an error, refuses the request too. It judges the latest user message alone. Its verdict is a probabilistic classifier's: it misses some attacks, refuses some legitimate requests and is known to be evadable. Only a flow in its configuration, redefining the product's own, keeps a request body from switching the rail off, and that closure has to be proved again for any image or flow it was not proved on. NVIDIA's NemoGuard models add rails on the content_safety and jailbreak_detection sockets; the jailbreak rail admits a request when its classifier is unreachable. It authenticates no caller, and with no credential attached to its upstream_key slot a caller can reach the upstream on a credential of its own. The Release supplies the image in infra.image and the catalog names none: TrustyAI's published build carries NLTK's punkt tokenizer model, whose licence is unresolved upstream, and is published only under a floating tag, whose superseded digests stop resolving. Nothing renders until the Release sets `infra.image`.

Provenance

OfficielSigné par une identité de signature que ce catalogue reconnaît comme éditeur officiel.
Empreinte du contenu
sha256:f871e845d92afd93e2d6d8c7fc6a9b33f86abbcfe1730cc89aab9d44fb3f2721
Publié
2026-11-10T10:13:39Z
Identité de signature
https://token.actions.githubusercontent.com/https://github.com/aggrega-ai/aggrega-hub/.github/workflows/catalog-publish-sign.yml@refs/heads/main

Signature

  • Officielapplication/vnd.dev.sigstore.bundle.v0.3+json

    https://token.actions.githubusercontent.com/https://github.com/aggrega-ai/aggrega-hub/.github/workflows/catalog-publish-sign.yml@refs/heads/main

    sha256:4a2b09d9b13c5c2efc29928cc9bd5abc2304b72043e199a4b9ffddab6eb8b06e

Référence à cette version

oci://hub.aggrega.ai/component/core/security/nemo-guardrails:1.0.0

Paramètres du développeur

Valeurs que le développeur définit dans l’Aggregate.

NomTypeValeur par défautObligatoireDescription
thresholdnumber (min 0, max 1)0.5nonProbability at or above which a label the detector reports refuses the request. Measured with the detector's default model, which scores most texts close to 0 or close to 1: at 0.5 the guard refuses plain attempts to override a model's instructions; at 0.001 it also refuses an override respelled in leetspeak, which the model scores near 0.008, and with it some ordinary requests the model scores low but not near zero. At either value it refuses the ordinary requests the model scores close to 1, and serves attacks it scores close to 0. Those two numbers hold for that model only, and are measured again for any other.
blocked_labelsarray (min 1)INJECTIONnonLabels of the detector's classifier that refuse a request when reported at or above the threshold. The default is the label the detector's default model gives an attempt to override a model's instructions; another model labels by its own vocabulary, which this list follows. It may not be empty, because a guard refusing on no label forwards everything.

Dimensionnement de l’opérateur

Dimensionnement de l’infrastructure que l’opérateur ajuste pour chaque environnement.

NomTypeValeur par défautObligatoireDescription
portinteger (min 1024, max 65535)8000nonPort the guard listens on and its Service accepts callers on. Most operators should omit this and let the Recipe assign the default (8000), the port the published image listens on unless told otherwise.
upstream_endpointstring—ouiBase URL of the OpenAI-compatible model API the guard forwards the requests it admits to, with its API path, for example http://model-server:8000/v1. A deterministic pointer string the guard resolves when it forwards; the compiler never contacts it. The request is forwarded under the model name the caller asked for.
imagestring—ouiImage the guard runs, supplied by the Release; the catalog names none. The Recipes run the published build's own start-up and paths, so the image is TrustyAI's NeMo Guardrails server or a copy of it. That build carries a data package whose licence is unresolved upstream and is tagged only as latest, under which the registry stops serving a build once the tag moves on, so a Release that must keep one build mirrors it and names the copy here. The guard's refusal of a request-body switch was proved on one build, and is proved again for any other image.
computeobject—non—

Contrat de sortie

Valeurs que ce composant publie vers les consommateurs via la topologie uses.

NomTypeValeur par défautObligatoireDescription
hoststring—ouiIn-cluster DNS name of the Service fronting the guard. Every caller reaching this name can send requests through the guard, so it is an address for the component routing model traffic through it rather than one to publish.
portinteger—ouiTCP port that Service accepts callers on. It follows the port an operator set in the Release, so a consumer reads it rather than assuming a default.
endpointstring—ouiBase URL of the OpenAI-compatible API the guard serves, with the host, the port and the API path already composed. A consumer hands it to a client library as its base URL, and sends chat completions to it.
protocolstring—nonURL scheme to reach `host` and `port` over, for a consumer composing its own address rather than reading `endpoint`.
thresholdnumber—ouivalues.threshold, the probability at or above which a label the detector reports refuses the request.

Sockets uses

Les sockets de dépendance par lesquels ce composant se connecte. La cible de chaque connexion câblée via un socket doit satisfaire son sélecteur d’étiquettes, et chaque connexion peut porter une attache dans les emplacements que le socket déclare. Un emplacement obligatoire doit être rempli par chaque connexion que l’Aggregate ou la Release câble via le socket.

detector

Sélecteur
security.aggrega.ai/text-detection=detector
Injection
—

Ce socket ne déclare aucun emplacement d’attache, donc une connexion qui le traverse ne porte aucune attache.

content_safety

Sélecteur
component.aggrega.ai/category=ai-model, model.aggrega.ai/api=openai
Injection
—

Ce socket ne déclare aucun emplacement d’attache, donc une connexion qui le traverse ne porte aucune attache.

jailbreak_detection

Sélecteur
component.aggrega.ai/category=security
Injection
—

Ce socket ne déclare aucun emplacement d’attache, donc une connexion qui le traverse ne porte aucune attache.

Emplacements d’attache

Points de liaison nommés auxquels une Feature s’attache à ce composant. Le sélecteur d’un emplacement restreint la Feature qui peut le remplir, et sa liste inject indique comment le matériel de la Feature peut être livré. Un emplacement obligatoire doit être lié sur chaque instance, par une attache sur l’instance ou à une portée supérieure, et un emplacement sensible est déclaré pour traiter du matériel secret. Une Feature peut aussi s’attacher sous un nom qu’aucun emplacement ne déclare.

EmplacementSélecteurObligatoireInjectionSensibleDescription
upstream_keyfeature.aggrega.ai/kind=secretnonenvouiOptional credential the guard presents to the upstream model API as a bearer token. Attached, every forwarded request carries it, whatever the caller sends. Unattached, the guard holds no upstream credential, which suits an upstream that asks for none: the product then sends a fixed placeholder in its place, which an upstream requiring a key refuses, or passes a caller's X-Authorization header on as the credential when the caller sends one.

Liste d’autorisation d’auto-injection

Aucun canal d’auto-injection déclaré.

Relations

Arêtes typées vers et depuis d’autres entités publiées. Une référence sans cible publiée s’affiche comme une arête rompue plutôt que comme un lien.

Utilise

Utilisé par

Référence du concept Component →

Le catalogue Aggrega — un index ouvert de composants, de fonctionnalités, de recettes et des plugins qui en assurent le rendu.

Conçu avec le système de design partagé d’Aggrega.

Confidentialité