Fix OpenShift Dashboard Playground OpenWebUI Connection (#634)

* fix(dashboard): add OpenShift hostname detection for OpenWebUI playground

Apply PlaygroundPage.tsx patch to detect OpenShift route hostnames and
automatically construct the correct OpenWebUI URL for the playground iframe.

This fixes the 'localhost refused to connect' error when accessing the
playground through the dashboard on OpenShift deployments.

Changes:
- Update PlaygroundPage.tsx to detect OpenShift hostname patterns
- Dynamically construct OpenWebUI route URL based on dashboard hostname
- Add BuildConfig for building custom dashboard in OpenShift
- Fallback to /embedded/openwebui/ for local development

Signed-off-by: szedan <[email protected]>

* feat: enhance OpenShift deployment with one-click observability stack

This commit improves the OpenShift deployment workflow by implementing
a streamlined, fully automated deployment process with comprehensive
observability support.

Key changes:

**Dashboard Build Support:**
- Add dashboard/Dockerfile for OpenShift binary builds
- Multi-stage build: Node.js frontend + Go backend
- Enables building dashboard from local source without Docker daemon
- Includes PlaygroundPage fix for dynamic OpenWebUI URL detection

**Deployment Script Enhancements:**
- Add --no-observability flag for minimal core-only deployments
- Implement full observability stack deployment (Dashboard, OpenWebUI, Grafana, Prometheus)
- Add inline OpenWebUI PVC creation (fixes missing PVC issue)
- Improve deployment feedback with component-specific success messages
- Add help flag with usage documentation

**Cleanup Script Improvements:**
- Add comprehensive observability resource cleanup
- Include dashboard-custom buildconfig and imagestream cleanup
- Update resource lists to match new deployment structure

**Cross-Cluster Portability:**
- Remove hardcoded hostname from openwebui/route.yaml
- Enable automatic route hostname generation per cluster
- Improve cluster portability for multi-cluster deployments

**Documentation Cleanup:**
- Remove outdated README files (dashboard, openwebui, observability)
- Remove deprecated build-custom-dashboard.sh script
- Remove obsolete PlaygroundPage.tsx.patch file
- Update main README.md with one-click deployment instructions
- Add binary build workflow documentation

**Testing:**
- End-to-end deployment verified on OpenShift 4.x
- Cleanup and redeploy tested successfully
- All pods running and routes accessible

Fixes: Dashboard playground "localhost refused to connect" error
Related: PlaygroundPage now uses dynamic hostname detection

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>
Signed-off-by: szedan <[email protected]>

* fix: add blank lines around lists in OpenShift README for markdown linting (MD032)

Resolves markdown linting errors by adding blank lines before list sections
at lines 254, 261, and 268 in deploy/openshift/README.md as required by
markdownlint MD032 rule.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>
Signed-off-by: szedan <[email protected]>

* fix(openshift): make vLLM endpoint IPs dynamic and lower classification threshold

This commit addresses two critical issues for OpenShift deployment portability
and query handling:

**Dynamic vLLM Endpoint Configuration:**
- Replace hardcoded ClusterIPs with dynamic placeholders (DYNAMIC_MODEL_A_IP, DYNAMIC_MODEL_B_IP)
- Update deploy script to substitute placeholders with actual discovered ClusterIPs
- Remove fallback logic that relied on old hardcoded IPs
- Improve error handling when IP substitution fails
- Makes deployments truly portable across different OpenShift clusters

**Classification Threshold Adjustment:**
- Lower category classifier threshold from 0.6 to 0.45
- Allows simple queries like "what is cat?" (confidence ~0.45) to be classified
- Previously, queries below 0.6 threshold resulted in empty category and routing failures
- Add default_model configuration for fallback routing

**Files Changed:**
- deploy/openshift/config-openshift.yaml: Replace hardcoded IPs with DYNAMIC_* placeholders, lower threshold, add default_model
- deploy/openshift/deploy-to-openshift.sh: Simplify sed patterns for IP replacement, improve error handling

**Testing:**
- Verified dynamic IP discovery during deployment
- Tested queries successfully routed to vLLM backends
- All pre-commit hooks passing

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>
Signed-off-by: szedan <[email protected]>

* feat(openshift): enable GPU for vLLM models and lower classification threshold

Changes:
- Enable CUDA GPU for both vLLM model deployments (Model-A and Model-B)
  - Changed --device from "cpu" to "cuda"
  - Added nvidia.com/gpu: 1 resource requests and limits
- Lower category classification threshold from 0.45 to 0.35
  - Fixes issue where queries like "what is physics?" (confidence 0.3581) were failing classification
  - Empty category results in no model routing, causing request failures
  - New threshold allows more queries to be successfully classified and routed

Tested on OpenShift cluster:
- Both models successfully loading on CUDA
- Classification now working for previously failing queries

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>
Signed-off-by: szedan <[email protected]>

---------

Signed-off-by: szedan <[email protected]>
Co-authored-by: Claude <[email protected]>

Author: szedan-rh

dashboard/Dockerfile

@@ -0,0 +1,25 @@
+# Build frontend
+FROM node:18-alpine AS frontend-builder
+WORKDIR /app/frontend
+COPY frontend/package*.json ./
+RUN npm ci
+COPY frontend/ ./
+RUN npm run build
+
+# Build backend
+FROM golang:1.21-alpine AS backend-builder
+WORKDIR /app/backend
+COPY backend/go.* ./
+RUN go mod download
+COPY backend/ ./
+RUN CGO_ENABLED=0 GOOS=linux go build -o dashboard-server .
+
+# Final image
+FROM alpine:3.18
+RUN apk add --no-cache ca-certificates
+WORKDIR /app
+COPY --from=backend-builder /app/backend/dashboard-server .
+COPY --from=frontend-builder /app/frontend/dist ./frontend
+ENV DASHBOARD_STATIC_DIR=./frontend
+EXPOSE 8700
+CMD ["./dashboard-server"]

dashboard/frontend/src/pages/PlaygroundPage.tsx

@@ -2,7 +2,15 @@ import React, { useState, useEffect } from 'react'
 import styles from './PlaygroundPage.module.css'
 
 const PlaygroundPage: React.FC = () => {
-  const [openWebUIUrl] = useState('http://localhost:3001')
+  // Detect OpenWebUI URL based on current hostname
+  // Assumes openwebui and dashboard have matching hostname patterns
+  const getOpenWebUIUrl = () => {
+    const hostname = window.location.hostname
+    const openwebuiHost = hostname.replace('dashboard', 'openwebui')
+    return `${window.location.protocol}//${openwebuiHost}`
+  }
+
+  const [openWebUIUrl] = useState(getOpenWebUIUrl())
   const [currentUrl, setCurrentUrl] = useState('')
 
   // Auto-load on mount

deploy/openshift/README.md

@@ -9,26 +9,54 @@ This directory contains OpenShift-specific deployment manifests for the vLLM Sem
 - OpenShift cluster access
 - `oc` CLI tool configured and logged in
 - Cluster admin privileges (or permissions to create namespaces and routes)
+- Local source code (for dashboard build)
 
-### Automated Deployment (Recommended)
+### One-Click Full Deployment (Recommended)
 
-The deployment script automatically handles everything including dynamic IP configuration:
+Deploy the complete stack including semantic-router, vLLM models, and all observability components:
 
 ```bash
 cd deploy/openshift
 ./deploy-to-openshift.sh
 ```
 
-This script will:
+This script will deploy:
+
+**Core Components:**
 
 - ✅ Build the llm-katan image from Dockerfile
 - ✅ Create namespace and PVCs
-- ✅ Deploy vLLM model services (model-a and model-b)
+- ✅ Deploy vLLM model services (Model-A and Model-B)
 - ✅ Auto-discover Kubernetes service ClusterIPs
 - ✅ Generate configuration with actual IPs (portable across clusters)
-- ✅ Deploy semantic-router with Envoy proxy
+- ✅ Deploy semantic-router with Envoy proxy sidecar
 - ✅ Create OpenShift routes for external access
 
+**Observability Stack:**
+
+- ✅ Dashboard (built from local source with PlaygroundPage fix)
+- ✅ OpenWebUI playground for testing models
+- ✅ Grafana for metrics visualization
+- ✅ Prometheus for metrics collection
+
+### Minimal Deployment (Core Only)
+
+If you only want the core semantic-router and vLLM models without observability:
+
+```bash
+cd deploy/openshift
+./deploy-to-openshift.sh --no-observability
+```
+
+This deploys only the core components without Dashboard, OpenWebUI, Grafana, and Prometheus.
+
+### Command Line Options
+
+| Flag | Description |
+|------|-------------|
+| `--no-observability` | Skip deploying Dashboard, OpenWebUI, Grafana, and Prometheus |
+| `--help`, `-h` | Show help message |
+
 ### Manual Deployment (Advanced)
 
 If you prefer manual deployment or need to customize:
@@ -53,23 +81,67 @@ If you prefer manual deployment or need to customize:
 
 4. **Note:** You'll need to manually configure ClusterIPs in `config-openshift.yaml`
 
+## How Dashboard Build Works
+
+The deployment script uses OpenShift's **binary build** approach for the dashboard:
+
+1. Creates a BuildConfig with Docker strategy
+2. Uploads the local `dashboard/` directory as build source
+3. Builds the image inside OpenShift (no local Docker required)
+4. Pushes to OpenShift internal registry
+5. Deploys using the built image
+
+### Why Binary Build?
+
+- ✅ No local Docker daemon required
+- ✅ Works on any machine with `oc` CLI
+- ✅ Builds with your local code changes (including PlaygroundPage fix)
+- ✅ Automatically integrated with OpenShift registry
+- ✅ Works across different OpenShift clusters
+
+### Updating Dashboard
+
+If you make changes to the dashboard code, rebuild and redeploy:
+
+```bash
+# Rebuild dashboard image from local source
+cd dashboard
+oc start-build dashboard-custom --from-dir=. --follow -n vllm-semantic-router-system
+
+# Restart deployment to use new image
+oc rollout restart deployment/dashboard -n vllm-semantic-router-system
+```
+
 ## Accessing Services
 
-After deployment, the services will be accessible via OpenShift Routes:
+After deployment, the script will display URLs for all services. Routes are automatically generated with cluster-appropriate hostnames.
 
 ### Get Route URLs
 
 ```bash
-# Classification API (HTTP REST)
+# Core Services
 oc get route semantic-router-api -n vllm-semantic-router-system -o jsonpath='{.spec.host}'
-
-# gRPC API
 oc get route semantic-router-grpc -n vllm-semantic-router-system -o jsonpath='{.spec.host}'
-
-# Metrics
 oc get route semantic-router-metrics -n vllm-semantic-router-system -o jsonpath='{.spec.host}'
+
+# Observability (if deployed)
+oc get route dashboard -n vllm-semantic-router-system -o jsonpath='{.spec.host}'
+oc get route openwebui -n vllm-semantic-router-system -o jsonpath='{.spec.host}'
+oc get route grafana -n vllm-semantic-router-system -o jsonpath='{.spec.host}'
+oc get route prometheus -n vllm-semantic-router-system -o jsonpath='{.spec.host}'
+```
+
+### Dashboard Playground
+
+Access the OpenWebUI playground through the dashboard:
+
+```bash
+DASHBOARD_URL=$(oc get route dashboard -n vllm-semantic-router-system -o jsonpath='{.spec.host}')
+echo "Playground: https://$DASHBOARD_URL/playground"
 ```
 
+The playground automatically detects the OpenWebUI URL by replacing `dashboard` with `openwebui` in the hostname - no configuration needed!
+
 ### Example Usage
 
 ```bash
@@ -133,20 +205,82 @@ curl https://$METRICS_ROUTE/metrics
 
 ## Cleanup
 
-Remove all resources:
+### Quick Cleanup
+
+Remove the entire namespace and all resources (recommended):
 
 ```bash
-oc delete -k deploy/openshift/
+cd deploy/openshift
+./cleanup-openshift.sh
 ```
 
-Or remove individual components:
+If not already logged in to OpenShift:
 
 ```bash
-oc delete -f deploy/openshift/routes.yaml
-oc delete -f deploy/openshift/service.yaml
-oc delete -f deploy/openshift/deployment.yaml
-oc delete -f deploy/openshift/pvc.yaml
-oc delete -f deploy/openshift/namespace.yaml
+oc login <your-cluster-url>
+./cleanup-openshift.sh
+```
+
+### Cleanup Options
+
+The cleanup script supports different cleanup levels:
+
+| Level | What Gets Deleted | What's Preserved |
+|-------|------------------|------------------|
+| `deployment` | Deployments, services, routes, configmaps, buildconfigs | Namespace, PVCs |
+| `namespace` (default) | Entire namespace and all resources | Nothing |
+| `all` | Namespace + cluster-wide resources | Nothing |
+
+**Examples:**
+
+```bash
+# Remove everything (default)
+./cleanup-openshift.sh
+
+# Keep namespace and PVCs, remove only deployments
+./cleanup-openshift.sh --level deployment
+
+# Dry run to see what would be deleted
+./cleanup-openshift.sh --dry-run
+
+# Force cleanup without confirmation
+./cleanup-openshift.sh --force
+```
+
+### What Gets Cleaned Up
+
+The cleanup script removes:
+
+**Core Components:**
+
+- semantic-router deployment
+- vLLM model deployments (Model-A, Model-B)
+- All services and routes
+- ConfigMaps (router config, envoy config)
+- BuildConfigs and ImageStreams (llm-katan, dashboard-custom)
+
+**Observability Stack:**
+
+- Dashboard deployment
+- OpenWebUI deployment
+- Grafana deployment
+- Prometheus deployment
+- All related services, routes, and configmaps
+
+**Storage (namespace level only):**
+
+- PVCs for models and cache
+
+### Manual Cleanup
+
+If you prefer manual cleanup:
+
+```bash
+# Delete entire namespace (removes everything)
+oc delete namespace vllm-semantic-router-system
+
+# Or delete specific components
+oc delete deployment,service,route,configmap,buildconfig,imagestream --all -n vllm-semantic-router-system
 ```
 
 ## Troubleshooting

deploy/openshift/cleanup-openshift.sh

@@ -61,8 +61,9 @@ OPTIONS:
     -h, --help               Show this help message
 
 CLEANUP LEVELS:
-    deployment    - Remove deployment, services, routes, configmap (keep namespace and PVC)
-    namespace     - Remove entire namespace and all resources (default)
+    deployment    - Remove deployments, services, routes, configmaps, buildconfigs (keep namespace and PVCs)
+                    Includes: semantic-router, vLLM models, dashboard, OpenWebUI, Grafana, Prometheus
+    namespace     - Remove entire namespace and all resources including PVCs (default)
     all           - Remove namespace and any cluster-wide resources
 
 EXAMPLES:
@@ -254,6 +255,18 @@ show_current_resources() {
     echo "=== ConfigMaps ==="
     oc get configmaps -n "$NAMESPACE" 2>/dev/null || echo "No configmaps found"
 
+    echo ""
+    echo "=== BuildConfigs ==="
+    oc get buildconfig -n "$NAMESPACE" 2>/dev/null || echo "No buildconfigs found"
+
+    echo ""
+    echo "=== ImageStreams ==="
+    oc get imagestream -n "$NAMESPACE" 2>/dev/null || echo "No imagestreams found"
+
+    echo ""
+    echo "=== Deployments ==="
+    oc get deployments -n "$NAMESPACE" 2>/dev/null || echo "No deployments found"
+
     echo ""
 }
 
@@ -270,10 +283,12 @@ confirm_cleanup() {
 
     case "$CLEANUP_LEVEL" in
         "deployment")
-            log "WARN" "Will delete: deployment, services, routes, configmaps (keeping namespace and PVCs)"
+            log "WARN" "Will delete: deployments, services, routes, configmaps, buildconfigs (keeping namespace and PVCs)"
+            log "WARN" "Components: semantic-router, vLLM models, dashboard, OpenWebUI, Grafana, Prometheus"
             ;;
         "namespace")
             log "WARN" "Will delete: entire namespace and all resources including PVCs"
+            log "WARN" "This removes ALL components: core + observability stack"
             ;;
         "all")
             log "WARN" "Will delete: namespace and any cluster-wide resources"
@@ -304,13 +319,43 @@ cleanup_deployment() {
 
     # Delete specific resources but keep namespace and PVCs
     local resources=(
+        # Core semantic-router resources
         "deployment/semantic-router"
+        "deployment/vllm-model-a"
+        "deployment/vllm-model-b"
         "service/semantic-router"
         "service/semantic-router-metrics"
+        "service/vllm-model-a"
+        "service/vllm-model-b"
         "route/semantic-router-api"
         "route/semantic-router-grpc"
         "route/semantic-router-metrics"
+        "route/envoy-http"
+        "route/envoy-admin"
         "configmap/semantic-router-config"
+        "configmap/envoy-config"
+        "buildconfig/llm-katan"
+        "imagestream/llm-katan"
+        "imagestream/python"
+
+        # Observability resources
+        "deployment/dashboard"
+        "deployment/openwebui"
+        "deployment/grafana"
+        "deployment/prometheus"
+        "service/dashboard"
+        "service/openwebui"
+        "service/grafana"
+        "service/prometheus"
+        "route/dashboard"
+        "route/openwebui"
+        "route/grafana"
+        "route/prometheus"
+        "configmap/dashboard-config"
+        "configmap/grafana-config"
+        "configmap/prometheus-config"
+        "buildconfig/dashboard-custom"
+        "imagestream/dashboard-custom"
     )
 
     for resource in "${resources[@]}"; do

deploy/openshift/config-openshift.yaml

@@ -32,11 +32,11 @@ prompt_guard:
 # This makes deployment portable across different OpenShift clusters
 vllm_endpoints:
   - name: "model-a-endpoint"
-    address: "172.30.64.134"  # PLACEHOLDER - will be replaced with actual ClusterIP
+    address: "DYNAMIC_MODEL_A_IP"  # model-a-ip - will be replaced with actual ClusterIP
     port: 8000
     weight: 1
   - name: "model-b-endpoint"
-    address: "172.30.116.177"  # PLACEHOLDER - will be replaced with actual ClusterIP
+    address: "DYNAMIC_MODEL_B_IP"  # model-b-ip - will be replaced with actual ClusterIP
     port: 8001
     weight: 1
 
@@ -59,7 +59,7 @@ classifier:
   category_model:
     model_id: "models/category_classifier_modernbert-base_model"
     use_modernbert: true
-    threshold: 0.6
+    threshold: 0.35  # Lowered from 0.45 to handle queries like "what is physics?" (confidence 0.3581)
     use_cpu: true
     category_mapping_path: "models/category_classifier_modernbert-base_model/category_mapping.json"
   pii_model: