Skip to content

README.md

Latest commit

 

History

History
472 lines (411 loc) · 10.4 KB

README.md

File metadata and controls

472 lines (411 loc) · 10.4 KB

Language Detector Guardrails

In this demo, we will use papluca/xlm-roberta-base-language-detection, a sequence classification model trained to detect 20 different lanaguages as a guardrail to block inputs and outputs in languages that are not allowed.

Prerequisites

  1. Ensure you have the Red Hat Openshift AI operator with the TrustyAI component enabled.
  2. GPU is recommended.
  3. A namespace/project is created and you have access to it.

1) Apply the service account and role bindings

oc apply -f service_account.yaml

2) Set up the detector model storage container

oc apply -f detector_storage.yaml

3) Deploy detector model

oc apply -f detector_deployment.yaml

Note that SAFE_LABELS environment variable is set to '[3, "pl"]' in the detector_deployment.yaml. Subsequently, this means that only Polish language should be allowed by the guardrail.

Generally, SAFE_LABELS is a list that can contain either string or integer labels (or a mixture of both). The labels should correspond to the labels used by the underlying sequence classification model; these can be usually found in the model card. For example, for the xlm-roberta-base-language-detection model used in this demo, the labels can be found here (the mapping is in the id2label field).

If SAFE_LABELS is not set inside the detector_deployment.yaml, then the default behaviour is to set first labels as safe (i.e. '[0]').

4) Set up the generative llm model storage container

oc apply -f llm_storage.yaml

5) Deploy the generative llm model

oc apply -f llm_deployment.yaml

6) Deploy the guardrails orchestrator

oc apply -f guardrails_auto_config.yaml

Sample requests -- Detector API

Get the route:

DETECTOR_ROUTE=$(oc get routes language-detector-route -o jsonpath='{.spec.host}')

Perform the health check

curl -v -H "Authorization: Bearer $(oc whoami -t)" https://$DETECTOR_ROUTE/health | jq

this should return:

"ok"

Send requests that should be allowed

These request should be allowed by the guardrail as their language is in the safe list of labels (Polish)

curl -v -X POST \
  "https://$DETECTOR_ROUTE/api/v1/text/contents" \
  -H "Authorization: Bearer $(oc whoami -t)" \
  -H 'accept: application/json' \
  -H 'detector-id: language_detector' \
  -H 'Content-Type: application/json' \
  -d '{
    "contents": ["Czesc - czy mozes mi pomoc?", "Jaki AI model jest najlepszy?"],
    "detector_params": {}
  }' | jq

this should return:

[
  [],
  []
]

Send requests that should be blocked

These requests should be blocked by the guardrail as their language is not in the safe list of labels (Polish)

curl -v -X POST \
  "https://$DETECTOR_ROUTE/api/v1/text/contents" \
  -H "Authorization: Bearer $(oc whoami -t)" \
  -H 'accept: application/json' \
  -H 'detector-id: language_detector' \
  -H 'Content-Type: application/json' \
  -d '{
    "contents": [
        "Hello - can you help me?",
        "Which AI model is the best?",
        "Bonjour - pouvez-vous m'"'"'aider?",
        "Quel modèle d'"'"'IA est le meilleur?",
        "Hola - ¿puedes ayudarme?",
        "¿Cuál es el mejor modelo de IA?",
        "Hallo - kannst du mir helfen?",
        "Welches KI-Modell ist das beste?",
        "Ciao - puoi aiutarmi?",
        "Qual è il miglior modello di intelligenza artificiale?"
    ],
    "detector_params": {}
  }' | jq

this should return:

[
  [
    {
      "start": 0,
      "end": 24,
      "text": "Hello - can you help me?",
      "detection": "single_label_classification",
      "detection_type": "en",
      "score": 0.9464484453201294,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 27,
      "text": "Which AI model is the best?",
      "detection": "single_label_classification",
      "detection_type": "en",
      "score": 0.9105632901191711,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 30,
      "text": "Bonjour - pouvez-vous m'aider?",
      "detection": "single_label_classification",
      "detection_type": "fr",
      "score": 0.9945297837257385,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 33,
      "text": "Quel modèle d'IA est le meilleur?",
      "detection": "single_label_classification",
      "detection_type": "fr",
      "score": 0.9915143251419067,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 24,
      "text": "Hola - ¿puedes ayudarme?",
      "detection": "single_label_classification",
      "detection_type": "es",
      "score": 0.9750027656555176,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 31,
      "text": "¿Cuál es el mejor modelo de IA?",
      "detection": "single_label_classification",
      "detection_type": "es",
      "score": 0.9921737909317017,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 29,
      "text": "Hallo - kannst du mir helfen?",
      "detection": "single_label_classification",
      "detection_type": "de",
      "score": 0.9949387311935425,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 32,
      "text": "Welches KI-Modell ist das beste?",
      "detection": "single_label_classification",
      "detection_type": "de",
      "score": 0.9937286376953125,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 21,
      "text": "Ciao - puoi aiutarmi?",
      "detection": "single_label_classification",
      "detection_type": "it",
      "score": 0.9950234889984131,
      "evidences": [],
      "metadata": {}
    }
  ],
  [
    {
      "start": 0,
      "end": 54,
      "text": "Qual è il miglior modello di intelligenza artificiale?",
      "detection": "single_label_classification",
      "detection_type": "it",
      "score": 0.9948535561561584,
      "evidences": [],
      "metadata": {}
    }
  ]
]

Sample requests -- Orchestrator API

  1. Get the route of the health service
ORCHESTRATOR_HEALTH_ROUTE=$(oc get routes guardrails-orchestrator-health -o jsonpath='{.spec.host}')
  1. Get the route of the main guardrails orchestrator service
ORCHESTRATOR_ROUTE=$(oc get routes guardrails-orchestrator -o jsonpath='{.spec.host}')

Perform the info check

curl -v -H "Authorization: Bearer $(oc whoami -t)" https://$ORCHESTRATOR_HEALTH_ROUTE/info | jq

this should return:

{
  "services": {
    "built-in-detector": {
      "status": "HEALTHY"
    },
    "language-detector": {
      "status": "HEALTHY"
    },
    "openai": {
      "status": "HEALTHY"
    }
  }
}

Send a request to the detector that should be allowed

curl -v \
  -H "Authorization: Bearer $(oc whoami -t)" \
  -H "accept: application/json" \
  -H "Content-Type: application/json" \
  -d '{
    "detectors": {
      "language-detector": {}
    },
    "content": "Czesc - czy mozesz mi pomoc?"
  }' \
  "https://$ORCHESTRATOR_ROUTE/api/v2/text/detection/content" | jq

which should return:

{
  "detections": []
}

Send a request to the detector that should be blocked

curl -v \
  -H "Authorization: Bearer $(oc whoami -t)" \
  -H "accept: application/json" \
  -H "Content-Type: application/json" \
  -d '{
    "detectors": {
      "language-detector": {}
    },
    "content": "Hello - can you help me?"
  }' \
  "https://$ORCHESTRATOR_ROUTE/api/v2/text/detection/content" | jq

which should return:

{
  "detections": [
    {
      "start": 0,
      "end": 24,
      "text": "Hello - can you help me?",
      "detection": "single_label_classification",
      "detection_type": "en",
      "detector_id": "language-detector",
      "score": 0.9464484453201294
    }
  ]
}

Send a request to a guardrailed llm model that should be allowed

curl -v \
  -H "Authorization: Bearer $(oc whoami -t)" \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "llm",
    "messages": [
      {
        "content": "Czesc - czy mozesz mi pomoc",
        "role": "user"
      }
    ],
    "detectors": {
      "input": {
        "language-detector": {}
      },
      "output": {
        "language-detector": {}
      }
    }
  }' \
  "https://$ORCHESTRATOR_ROUTE/api/v2/chat/completions-detection" | jq

which should return

{
  "id": "chat-c8ef177dd7f349c9897c31ced017f788",
  "object": "chat.completion",
  "created": 1760357920,
  "model": "llm",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Cześć! Jak mogę Ci pomóc dalej?"
      },
      "logprobs": null,
      "finish_reason": "stop",
      "stop_reason": null
    }
  ],
  "usage": {
    "prompt_tokens": 39,
    "total_tokens": 53,
    "completion_tokens": 14
  },
  "prompt_logprobs": null
}

Send a request to a guardrailed llm model that should be blocked

curl -v \
  -H "Authorization: Bearer $(oc whoami -t)" \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "llm",
    "messages": [
      {
        "content": "Hello - can you help me?",
        "role": "user"
      }
    ],
    "detectors": {
      "input": {
        "language-detector": {}
      },
      "output": {
        "language-detector": {}
      }
    }
  }' \
  "https://$ORCHESTRATOR_ROUTE/api/v2/chat/completions-detection" | jq

which should return

{
  "id": "8f154a9f8c0c40378c4529468b18fd5f",
  "object": "",
  "created": 1760358109,
  "model": "llm",
  "choices": [],
  "usage": {
    "prompt_tokens": 7,
    "total_tokens": 0,
    "completion_tokens": 0
  },
  "prompt_logprobs": null,
  "detections": {
    "input": [
      {
        "message_index": 0,
        "results": [
          {
            "start": 0,
            "end": 24,
            "text": "Hello - can you help me?",
            "detector_id": "language-detector",
            "detection_type": "en",
            "detection": "single_label_classification",
            "score": 0.9464484453201294
          }
        ]
      }
    ]
  },
  "warnings": [
    {
      "type": "UNSUITABLE_INPUT",
      "message": "Unsuitable input detected. Please check the detected entities on your input and try again with the unsuitable input removed."
    }
  ]
}