Deploy a privacy-focused AI assistant

Copy link

Build a healthcare AI assistant that ensures your large language model has multiple layers of protection, including PII detection and content moderation.

Detailed description

Copy link

This quickstart includes a healthcare AI assistant demo showing how guardrails could help protect HIPAA-compliant applications.

The demo tests a patient services AI with four protection layers:

1. PII Detection - Protects Social Security Numbers and medical IDs
2. Content Moderation - Blocks inappropriate language
3. Prompt Injection Protection - Prevents system manipulation
4. Gibberish Detection - Filters out nonsense queries

For example, here's how PII detection works in action:

Explore the complete interactive demo in docs/healthcare-guardrails.ipynb.

The LLM Guardrails quickstart is a quick-start template for deploying multiple layers of protection for LLM applications using TrustyAI's orchestrator and specialized detector services.

This quickstart includes a Helm chart for deploying:

• A Llama 3.2 3B Instruct model with GPU acceleration.
• Multiple AI safety detectors: gibberish detection, prompt injection detection, and hate/profanity detection.
• TrustyAI GuardrailsOrchestrator for coordinating safety checks.
• Configurable detection thresholds and routing policies.

Architecture diagrams

Copy link

Requirements

Copy link

Recommended hardware requirements

Copy link

• GPU required for main LLM: +24GiB vRAM
• CPU cores: 12+ cores total (4 for LLM + 8 for detectors)
• Memory: 24Gi+ RAM total
• Storage: 10Gi

Minimum hardware requirements

Copy link

• GPU required for main LLM: 1 x NVIDIA GPU with 24GiB vRAM
• CPU cores: 8+ cores total
• Memory: 16Gi+ RAM total
• Storage: 5Gi

Minimum software requirements

Copy link

• Red Hat OpenShift 4.19.9
• Red Hat OpenShift Service Mesh 2
• Red Hat OpenShift AI 2.23.0
  • KServe needs to be enabled

Please note before you start

This example was tested on Red Hat OpenShift 4.19.9 & Red Hat OpenShift AI 2.23.0.

Required user permissions

Copy link

• Cluster admin permissions are required

Install

Copy link

Clone the repository

Copy link

git clone https://github.com/rh-ai-quickstart/guardrailing-llms.git && cd guardrailing-llms/

Create a new project

Copy link

PROJECT="guardrails-demo"

oc new-project ${PROJECT}

Install with Helm

Copy link

helm install ${PROJECT} helm/ --namespace ${PROJECT} 

Wait for the pods to be ready

Copy link

oc get pod -n ${PROJECT}

You should see an output similar to:

NAME                                                         READY   STATUS      RESTARTS   AGE
gibberish-detector-predictor-578fc59776-www4s                2/2     Running     0          25h
gorch-sample-5f95f587fd-wmk4x                                3/3     Running     0          51m
guardrails-workbench-0                                       2/2     Running     0          93m
guardrails-workbench-clone-repo-96jhd                        0/1     Completed   0          93m
ibm-hate-and-profanity-detector-predictor-846758cfb5-wnlnd   2/2     Running     0          25h
llama-32-3b-instruct-predictor-c8d55bd58-lctjn               2/2     Running     0          18m
prompt-injection-detector-predictor-7d784957f9-f2x5g         2/2     Running     0          25h

Test

Copy link

You can get the OpenShift AI Dashboard URL by:

oc get routes rhods-dashboard -n redhat-ods-applications

Once inside the dashboard, navigate to Data Science Projects -> guardrails-demo (or what you called your ${PROJECT} if you changed from default).

Inside the project you can see Workbenches, open up the one for guardrails-workbench.

Open the workbench, inside of the Jupyter Notebook folder, you'll see the guardrailing-llms repository already cloned, go to docs/healthcare-guardrails.ipynb and follow the instructions.

Enjoy!

Delete

Copy link

helm uninstall ${PROJECT} --namespace ${PROJECT} 

References

Copy link

• Red Hat documentation