Audit AI apps to meet compliance goals

Copy link

Deploy a financial audit system with insights into the performance, cost, and usage of your large language models in real time.

Table of Contents

Copy link

• Detailed description
  • Architecture
  • Components
• Requirements
• Installation
• Advanced Usage
  • MaaS Integration (Optional)
  • Uninstall
• References

Detailed description

Copy link

This quickstart features a Financial Services Industry (FSI) compliance demo that demonstrates auditability and traceability for AI-powered financial applications.

As AI workloads become more complex, organizations need robust telemetry and observability of Large Language Model (LLM) infrastructure. This quickstart demonstrates those needs by providing:

• AI observability for model performance, resource utilization, and distributed tracing
• Standardized deployment patterns for consistent, scalable AI service delivery
• Enterprise-grade monitoring using OpenShift-native observability tools

This repository provides helm charts for both the monitoring infrastructure and AI service deployments needed to run Llama Stack reliably in OpenShift AI.

Architecture

Copy link

The proposed observability & telemetry architecture:

This architecture demonstrates the complete observability ecosystem, from AI workload telemetry collection through distributed tracing to comprehensive monitoring dashboards.

Components

Copy link

All components are organized by dependency layers in the ./helm/ directory:

Phase 1: Operators (./helm/01-operators/)

Copy link

• cluster-observability-operator - PodMonitor/ServiceMonitor CRDs and UI plugins
• grafana-operator - Grafana operator for visualization and dashboard management
• otel-operator - Red Hat Build of OpenTelemetry operator
• tempo-operator - Distributed tracing backend operator

Phase 2: Observability Infrastructure (./helm/02-observability/)

Copy link

• tempo - Distributed tracing backend with S3-compatible storage
• otel-collector - OpenTelemetry collector configurations for telemetry collection and processing
• grafana - Visualization and dashboard management with pre-built dashboards
• uwm - User Workload Monitoring with PodMonitors for VLLM and AI workloads
• distributed-tracing-ui-plugin - OpenShift console integration for tracing

Phase 3: AI Services (./helm/03-ai-services/)

Copy link

• llama-stack-instance - Complete Llama Stack Instance with configurable endpoints
• llama3.2-3b - Llama 3.2 3B model deployment on vLLM
• llama-stack-playground - Interactive Llama-Stack web interface for testing
• llama-guard - Llama Guard 1B for providing Safety / Shield mechanisms

Phase 4: MCP Servers (./helm/04-mcp-servers/)

Copy link

• mcp-weather - MCP weather service
• hr-api - MCP HR API demonstration service
• openshift-mcp - MCP OpenShift/Kubernetes operations service

Observability in Action

Copy link

The telemetry and observability stack provides comprehensive visibility into AI workload performance and distributed system behavior.

Distributed Tracing Examples

Copy link

End-to-End Request Tracing: Complete visibility into AI inference request flows through the Llama Stack infrastructure.

Create Agent from LlamaStack Tracing: Detailed trace view showing complex interactions between different services in the AI stack.

These traces provide insights into:

• Request latency and service dependencies
• Error tracking and performance bottlenecks
• Load distribution across model endpoints

Requirements

Copy link

Minimum hardware requirements

Copy link

• CPU: 8+ cores recommended for full stack deployment
• Memory: 16GB+ RAM for monitoring stack, additional memory based on AI workload requirements
• Storage: 100GB+ for observability data retention
• GPU: NVIDIA GPU required for AI model inference (varies by model size)

Minimum software requirements

Copy link

• OpenShift 4.12+ or Kubernetes 1.24+
• OpenShift AI 2.19 onwards
• Helm 3.8+ for chart deployment
• oc CLI or kubectl for cluster management

Required user permissions

Copy link

• Cluster Admin - Required for operator installation and observability stack setup
• GPU Access - Required for AI workload deployment

Deploy

Copy link

# step 0 - git clone
https://github.com/rh-ai-quickstart/lls-observability.git && \ 
cd lls-observability

Quick Start - Automated Installation

Copy link

Option 1: Complete Stack (Recommended)

# Run the full installation script
./scripts/install-full-stack.sh

Option 2: Phase-by-Phase Installation

# Phase 1: Install operators
./scripts/install-operators.sh

# Phase 2: Deploy observability infrastructure
./scripts/deploy-observability.sh

# Phase 3: Deploy AI workloads
./scripts/deploy-ai-workloads.sh

Option 3: Using Makefile (Optional)

# Install everything in one command
make install-all

# Or phase-by-phase
make install-operators      # Phase 1
make deploy-observability   # Phase 2  
make deploy-ai              # Phase 3

Advanced Usage

Copy link

Manual Step-by-Step Installation

Copy link

For users who prefer to understand each step or need to customize the installation.

Set these environment variables before running the installation commands:

export OBSERVABILITY_NAMESPACE="observability-hub"
export UWM_NAMESPACE="openshift-user-workload-monitoring"
export AI_SERVICES_NAMESPACE="llama-serve"

Launch this instructions:

# 1. Create required namespaces
oc create namespace ${OBSERVABILITY_NAMESPACE}
oc create namespace ${UWM_NAMESPACE}
oc create namespace ${AI_SERVICES_NAMESPACE}

# 2. Install required operators
helm install cluster-observability-operator ./helm/01-operators/cluster-observability-operator
helm install grafana-operator ./helm/01-operators/grafana-operator
helm install otel-operator ./helm/01-operators/otel-operator
helm install tempo-operator ./helm/01-operators/tempo-operator

# 3. Wait for operators to be ready
oc wait --for=condition=Ready pod -l app.kubernetes.io/name=cluster-observability-operator -n openshift-cluster-observability-operator --timeout=300s
oc wait --for=condition=Ready pod -l app.kubernetes.io/name=observability-operator -n openshift-cluster-observability-operator --timeout=300s

# 4. Deploy observability infrastructure
helm install tempo ./helm/02-observability/tempo -n ${OBSERVABILITY_NAMESPACE}
helm install otel-collector ./helm/02-observability/otel-collector -n ${OBSERVABILITY_NAMESPACE}
helm install grafana ./helm/02-observability/grafana -n ${OBSERVABILITY_NAMESPACE}

# 5. Enable User Workload Monitoring for AI workloads
helm template uwm ./helm/02-observability/uwm -n ${OBSERVABILITY_NAMESPACE} | oc apply -f-

# Verify UWM setup
oc get configmap user-workload-monitoring-config -n ${UWM_NAMESPACE}
oc get podmonitors -n ${OBSERVABILITY_NAMESPACE}

# 6. Deploy AI workloads
# Deploy MCP servers in AI services namespace
helm install mcp-weather ./helm/04-mcp-servers/mcp-weather -n ${AI_SERVICES_NAMESPACE}
helm install hr-api ./helm/04-mcp-servers/hr-api -n ${AI_SERVICES_NAMESPACE}

# Milvus vector database is configured inline within llama-stack-instance
# No external Milvus deployment needed

# Deploy AI services in AI services namespace  
helm install llama3-2-3b ./helm/03-ai-services/llama3.2-3b -n ${AI_SERVICES_NAMESPACE} \
  --set model.name="meta-llama/Llama-3.2-3B-Instruct" \
  --set resources.limits."nvidia\.com/gpu"=1

helm install llama-stack-instance ./helm/03-ai-services/llama-stack-instance -n ${AI_SERVICES_NAMESPACE} \
  --set 'mcpServers[0].name=weather' \
  --set 'mcpServers[0].uri=http://mcp-weather.${AI_SERVICES_NAMESPACE}.svc.cluster.local:80' \
  --set 'mcpServers[0].description=Weather MCP Server for real-time weather data'

helm install llama-stack-playground ./helm/03-ai-services/llama-stack-playground -n ${AI_SERVICES_NAMESPACE} \
  --set playground.llamaStackUrl="http://llama-stack-instance.${AI_SERVICES_NAMESPACE}.svc.cluster.local:80"

helm install llama-guard ./helm/03-ai-services/llama-guard -n ${AI_SERVICES_NAMESPACE}

# 7. Enable tracing UI
helm install distributed-tracing-ui-plugin ./helm/02-observability/distributed-tracing-ui-plugin

MaaS Integration (Optional)

Copy link

Model-as-a-Service (MaaS) Remote Models

The Llama Stack can optionally integrate with external MaaS (Models as a Server) to access remote models without local GPU requirements.

Configuration Options:

maas:
  enabled: true                    # Enable MaaS provider
  apiToken: "your-api-token"      # Authentication token
  url: "https://endpoint.com/v1"  # MaaS endpoint URL
  maxTokens: 200000               # Maximum token limit
  tlsVerify: false                # TLS verification
  modelId: "model-name"           # Specific MaaS model identifier

Uninstall

Copy link

Complete Stack Uninstall

# Uninstall everything in reverse order
make clean

References

Copy link

Documentation

Copy link

• OpenShift Distributed Tracing
• OpenShift Observability
• User Workload Monitoring

Related Projects

Copy link

• Llama Stack
• vLLM
• OpenTelemetry
• Grafana
• Tempo

Community

Copy link

• Red Hat AI Quickstarts
• OpenShift AI Documentation