LightRAG/k8s-deploy/lightrag-minimal/README.md

LightRAG Minimal Helm Chart

This Helm chart deploys a production-ready LightRAG setup with PostgreSQL and pgvector support for Kubernetes environments. It has been tested and validated with complete teardown/rebuild cycles.

Configuration

This chart provides a comprehensive LightRAG deployment with:

• PostgreSQL with pgvector: For vector storage, KV storage, and document status using pgvector/pgvector:pg16 image
• NetworkX: For graph storage (local, no external database required)
• Persistent Storage: For data persistence across pod restarts
• Health Checks: Automated health monitoring
• API Endpoints: Document upload, query, and management
• Conservative Concurrency: Optimized OpenAI API usage to prevent rate limiting

Prerequisites

• Kubernetes 1.19+ (tested with Minikube)
• Helm 3.0+ with Bitnami repository
• OpenAI API key
• Storage class that supports ReadWriteOnce (standard storage class works)
• Minimum resources: 2 CPU cores, 4Gi memory available

Validated Installation Steps

Deploying the Chart

cd lightrag-minimal
helm repo add bitnami https://charts.bitnami.com/bitnami
helm repo update
helm dependency update

# (Optional) create a copy of values.yaml and customize it for your environment
cp values.yaml my-values.yaml
# edit my-values.yaml as needed (OpenAI keys, storage class, Postgres password, etc.)

helm install lightrag-minimal . \
  -f my-values.yaml \
  --namespace lightrag \
  --create-namespace

kubectl wait --namespace lightrag \
  --for=condition=ready pod \
  -l app.kubernetes.io/name=postgresql \
  --timeout=120s

kubectl wait --namespace lightrag \
  --for=condition=ready pod \
  -l app.kubernetes.io/name=lightrag-minimal \
  --timeout=120s

# Optional: expose the service locally
kubectl port-forward --namespace lightrag svc/lightrag-minimal 9621:9621 &

Configuration Options

Validated Environment Configuration

env:
  # OpenAI API Configuration (REQUIRED)
  LLM_BINDING: "openai"
  LLM_BINDING_HOST: "https://api.openai.com/v1"
  EMBEDDING_BINDING: "openai"
  EMBEDDING_BINDING_HOST: "https://api.openai.com/v1"
  EMBEDDING_MODEL: "text-embedding-ada-002"
  EMBEDDING_DIM: "1536"
  
  # Conservative concurrency (prevents API errors)
  MAX_ASYNC: "4"
  MAX_PARALLEL_INSERT: "2"
  
  # LLM Configuration
  ENABLE_LLM_CACHE: "true"
  ENABLE_LLM_CACHE_FOR_EXTRACT: "true"
  TIMEOUT: "240"
  TEMPERATURE: "0"
  MAX_TOKENS: "32768"

PostgreSQL Configuration

postgresql:
  # CRITICAL: Use pgvector image for vector support
  image:
    registry: docker.io
    repository: pgvector/pgvector
    tag: pg16
  auth:
    password: "your-secure-password"

Development vs Production

Setting	Development	Production
Resources	1 CPU, 2Gi RAM	4 CPU, 8Gi RAM
Storage	5Gi	100Gi
Replicas	1	2-10 (autoscaling)
Ingress	Disabled	Enabled with TLS
Storage Class	Default	fast-ssd

Accessing LightRAG

Development Access

# Port forward (included in installation steps above)
kubectl port-forward --namespace lightrag svc/lightrag-minimal 9621:9621 &

# Access URLs
echo "Web UI: http://localhost:9621/webui"
echo "API Docs: http://localhost:9621/docs"
echo "Health Check: http://localhost:9621/health"

Verify Deployment

# Check health
curl http://localhost:9621/health

# Expected response:
{
  "status": "healthy",
  "configuration": {
    "llm_model": "gpt-4o",
    "kv_storage": "PGKVStorage",
    "vector_storage": "PGVectorStorage",
    "graph_storage": "NetworkXStorage"
  }
}

Production (Ingress)

Production uses ingress with TLS (see values-prod.yaml):

ingress:
  enabled: true
  className: "nginx"
  hosts:
    - host: lightrag.yourdomain.com

Monitoring

Check Deployment Status

kubectl get pods -l app.kubernetes.io/name=lightrag-minimal
kubectl get services -l app.kubernetes.io/name=lightrag-minimal

View Logs

kubectl logs -l app.kubernetes.io/name=lightrag-minimal -f

Health Checks

The deployment includes health checks on /health endpoint.

Scaling

For production workloads, consider enabling autoscaling:

autoscaling:
  enabled: true
  minReplicas: 2
  maxReplicas: 10
  targetCPUUtilizationPercentage: 70

Upgrading

helm upgrade lightrag-minimal ./lightrag-minimal -f values-prod.yaml

Uninstalling

helm uninstall lightrag-minimal

Note: This will delete all data unless you have persistent volumes with a retain policy.

Document Loading

After successful deployment, load your documentation using the included loader. The loader supports two reference modes:

Reference Modes

Files Mode (Default): Uses file paths in citations

# Install dependencies (if needed)
pip install httpx

# Load documents with file path references
python ../../../load_docs.py /path/to/your/docs --endpoint http://localhost:9621

# Example with relative path
python ../../../load_docs.py ../docs --endpoint http://localhost:9621

URLs Mode: Uses website URLs in citations (recommended for public documentation)

# Load Apolo documentation with URL references
python ../../../load_docs.py ../apolo-copilot/docs/official-apolo-documentation/docs \
  --mode urls --base-url https://docs.apolo.us/index/ --endpoint http://localhost:9621

# Load custom documentation with URL references  
python ../../../load_docs.py /path/to/docs \
  --mode urls --base-url https://your-docs.example.com/docs/ --endpoint http://localhost:9621

Benefits of URL Mode

• Clickable References: Query responses include direct links to source documentation
• Better User Experience: Users can easily navigate to original content
• Professional Citations: References point to live documentation sites

⚠️ Important: File Structure Requirements for URL Mode

Your local file structure must match your documentation site's URL structure:

# Example: GitBook documentation site
docs/
├── getting-started/
│   ├── installation.md          → https://docs.example.com/getting-started/installation
│   └── first-steps.md           → https://docs.example.com/getting-started/first-steps
├── administration/
│   ├── README.md                → https://docs.example.com/administration
│   └── setup.md                 → https://docs.example.com/administration/setup
└── README.md                    → https://docs.example.com/

Quick Setup Guide:

1. Analyze your docs site: Visit URLs and note the path structure
2. Create matching directories: mkdir -p docs/{section1,section2,section3}
3. Organize markdown files: Place files to match URL paths (remove .md from URLs)
4. Verify mapping: Test a few URLs manually before loading

URL Mapping Rules:

• .md extension is removed from URLs
• README.md files map to their directory URL
• Subdirectories become URL path segments
• File and folder names should match URL slugs exactly

Expected Output

Both modes produce similar output with different reference formats:

🚀 Loading Documentation into LightRAG
============================================================
📁 Documentation path: /path/to/docs
🔧 Reference mode: urls
🌐 Base URL: https://docs.apolo.us/index/
🌐 LightRAG endpoint: http://localhost:9621

✅ LightRAG is healthy: healthy
📚 Found 58 markdown files
🔧 Mode: urls
🌐 Base URL: https://docs.apolo.us/index/
📊 Total content: 244,400 characters
📊 Average length: 4,287 characters

🔄 Starting to load documents...
✅ Loaded: Document Title
📈 Progress: 10/58 (10 success, 0 failed)
...
✅ Loading complete!
📊 Successful: 58
📊 Failed: 0
✅ Query successful!

Query Response Examples

Files Mode References:

### References
- [DC] getting-started/installation.md
- [KG] administration/cluster-setup.md

URLs Mode References:

### References
- [DC] https://docs.apolo.us/index/getting-started/installation
- [KG] https://docs.apolo.us/index/administration/cluster-setup

Troubleshooting

Common Issues

Issue: UnsupportedProtocol: Request URL is missing protocol

• Solution: Ensure LLM_BINDING_HOST and EMBEDDING_BINDING_HOST are set to https://api.openai.com/v1

Issue: Document processing failures with API connection errors

• Solution: Reduce concurrency with MAX_ASYNC: "4" and MAX_PARALLEL_INSERT: "2"

Issue: pgvector extension missing

• Solution: Ensure using pgvector/pgvector:pg16 image, not standard PostgreSQL

Validation Commands

# Check all pods are running
kubectl get pods --namespace lightrag

# Verify API connectivity
kubectl exec --namespace lightrag \
  $(kubectl get pod -l app.kubernetes.io/name=lightrag-minimal --namespace lightrag -o jsonpath='{.items[0].metadata.name}') \
  -- python -c "import requests; print(requests.get('https://api.openai.com/v1/models', headers={'Authorization': 'Bearer ' + open('/dev/null').read()}, timeout=5).status_code)"

# Check document processing status
curl http://localhost:9621/documents | jq '.statuses | to_entries | map({status: .key, count: (.value | length)})'

Clean Teardown and Rebuild

For testing or redeployment:

# Complete teardown
helm uninstall lightrag-minimal --namespace lightrag
kubectl delete namespace lightrag

# Rebuild (repeat installation steps above)
# This process has been validated multiple times

Validated Features

✅ Pure Helm Deployment - No manual kubectl apply commands needed
✅ PostgreSQL with pgvector - Automatic extension creation via proper image
✅ Environment Flexibility - Separate dev/prod configurations
✅ Document Loading - Working API with file_source parameter
✅ Conservative Concurrency - Prevents OpenAI API rate limiting
✅ Health Monitoring - Comprehensive health checks and status endpoints
✅ Persistent Storage - Data survives pod restarts and cluster updates

Comparison with Docker Compose

Feature	Docker Compose	Helm Chart
PostgreSQL	pgvector/pgvector:pg16	Same image via subchart
Concurrency	MAX_ASYNC=4	Same settings
API Configuration	.env file	Environment variables
Scaling	Single container	Kubernetes autoscaling
Persistence	Local volumes	PersistentVolumeClaims
Monitoring	Manual	Kubernetes native

This chart maintains the same conservative, working configuration as the Docker Compose setup while adding Kubernetes-native features for production deployment.