Component
core/security/trustyai-huggingface-detector
TrustyAI's Hugging Face detector: a service that runs one text classifier from the Hugging Face Hub and reports what it detects in the texts a caller sends. For each text posted to its content route with a threshold, it answers with every label the classifier scores at or above it, never the classifier's first label, read as the one for harmless text. It changes nothing and forwards nothing: what to do about a detection is the caller's decision, and core/security/nemo-guardrails is one such caller. The classifier is values.model at values.model_revision. The default, protectai/deberta-v3-base-prompt-injection-v2 (Apache-2.0), labels each text SAFE or INJECTION, and its own model card states that the project which published it is archived and no longer maintained. Another model labels and scores differently, so a caller's labels and thresholds follow it. The verdict misses some attacks and flags some ordinary requests. The weights are fetched where the service runs, never at build time. It authenticates no caller and every text reaches it as sent, so it is published on an in-cluster address only. Its own metrics, in the Prometheus format at /metrics on the same port, are published as `metrics_endpoint`. The Release supplies the image in infra.image and the catalog names none: TrustyAI's published build carries NVIDIA's proprietary CUDA runtime, whose licence agreement limits its use to systems with NVIDIA GPUs, and is published only under a floating tag, whose superseded digests stop resolving. Nothing renders until the Release sets `infra.image`.
Provenance
- Empreinte du contenu
- sha256:8f985fb76602947f3d871a7625d6614d76375528aa12e151d26a51559254d630
- Publié
- 2026-11-10T10:13:59Z
- 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:c34f4d9c71225ccd5e22a7d802c196dda140050af22eb57aa716bedbe73525f8
Référence à cette version
oci://hub.aggrega.ai/component/core/security/trustyai-huggingface-detector:1.0.0Paramètres du développeur
Valeurs que le développeur définit dans l’Aggregate.
| Nom | Type | Valeur par défaut | Obligatoire | Description |
|---|---|---|---|---|
| model | string | protectai/deberta-v3-base-prompt-injection-v2 | non | Hugging Face model id of the sequence-classification model the service runs. The default is a prompt-injection classifier labelling each text SAFE or INJECTION; another model labels by its own vocabulary, which a caller's refusal labels must then follow. |
| model_revision | string | 90c9989b1a342275dd0d1a95aad283c04e075671 | non | Revision of values.model to fetch — a commit id, a tag or a branch. The default is the commit of the default model whose scores a caller's thresholds are measured against; set it together with values.model, and main follows the repository's newest revision. |
Dimensionnement de l’opérateur
Dimensionnement de l’infrastructure que l’opérateur ajuste pour chaque environnement.
| Nom | Type | Valeur par défaut | Obligatoire | Description |
|---|---|---|---|---|
| port | integer (min 1024, max 65535) | 8000 | non | Port the service 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. |
| image | string | — | oui | Image the service and its weights-pull step run, supplied by the Release; the catalog names none. The Recipes run the published build's own server command, log configuration and Hugging Face command-line tool, so the image is TrustyAI's Hugging Face detector or a copy of it. That build carries NVIDIA's proprietary CUDA runtime, whose licence agreement limits its use to systems with NVIDIA GPUs, 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. |
| compute | object | — | non | — |
| model_cache | object | — | non | The volume the pull step writes the weights to and the server reads them from. |
Contrat de sortie
Valeurs que ce composant publie vers les consommateurs via la topologie uses.
| Nom | Type | Valeur par défaut | Obligatoire | Description |
|---|---|---|---|---|
| host | string | — | oui | In-cluster DNS name of the Service fronting the detector. Every caller reaching this name can submit text, so it is an address for the component that calls the service rather than one to publish. |
| port | integer | — | oui | TCP 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. |
| endpoint | string | — | oui | Base URL of the detector's HTTP API, with the host and the port already composed and no path. A caller posts the texts to check to the content route under it. |
| model | string | — | oui | Hugging Face model id of the classifier the service loaded, so a caller can record which model gave a verdict. |
| protocol | string | — | non | URL scheme to reach `host` and `port` over, for a consumer composing its own address rather than reading `endpoint`. |
| metrics_endpoint | string | — | non | The full address Prometheus scrapes this component's metrics at — scheme, host, port and path, `http://<host>:<port>/metrics` on the port callers reach. Empty where the component serves none reachable. |
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.
Aucun socket uses déclaré.
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.
| Emplacement | Sélecteur | Obligatoire | Injection | Sensible | Description |
|---|---|---|---|---|---|
| model_pull_token | feature.aggrega.ai/kind=secret | non | env | oui | Optional Hugging Face Hub access token for a gated model. When attached, the weights-pull step receives it as an environment variable, and the server itself never does; unattached, weights are pulled anonymously, which the default model allows. |
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
Cette entité n’utilise rien du catalogue.
Utilisé par
utilise le socket « auth_core »
utilise le socket « auth_core »
utilise le socket « auth_core »
utilise le socket « auth_core »
utilise le socket « upstream »
utilise le socket « forward_target »
utilise le socket « scrape_target »
utilise le socket « store »
utilise le socket « detector »
utilise le socket « detector »
utilise le socket « jailbreak_detection »
s’applique à
s’applique à
s’applique à
s’applique à
s’applique à
s’applique à
s’applique à
implémente
implémente