RAG Deployment Strategies

Overview

Deploying RAG systems at scale requires careful consideration of containerization strategies, orchestration patterns, versioning approaches, and CI/CD integration. This page explores production-ready deployment patterns for enterprise RAG implementations on Azure Arc.

---

Container-Based RAG Deployment

RAG Component Architecture

A production RAG system consists of multiple containerized services:

graph TB
    subgraph App[Application Layer]
        UI[Chat Interface<br/>React/Vue]
    end

    subgraph Orch[Orchestration Layer]
        Orchestrator[RAG Orchestration Service<br/>Query routing<br/>Context assembly<br/>Response formatting]
    end

    subgraph Core[Core Services]
        LLM[LLM Service<br/>Ollama/VLLM<br/>Quantized Models]
        Vector[Vector Search<br/>Weaviate/Qdrant/Milvus]
        Data[Data Layer<br/>Postgres<br/>MongoDB<br/>Elastic]
    end

    UI --> Orchestrator
    Orchestrator --> LLM
    Orchestrator --> Vector
    Orchestrator --> Data

    style App fill:#E8F4FD,stroke:#0078D4,stroke-width:2px,color:#000
    style Orch fill:#FFF4E6,stroke:#FF8C00,stroke-width:2px,color:#000
    style Core fill:#F3E8FF,stroke:#7B3FF2,stroke-width:2px,color:#000

Kubernetes Deployment Manifest

Structure for RAG on AKS Arc:

apiVersion: v1
kind: Namespace
metadata:
  name: rag-system
---
# LLM Service Deployment
apiVersion: apps/v1
kind: Deployment
metadata:
  name: llm-service
  namespace: rag-system
spec:
  replicas: 3
  selector:
    matchLabels:
      app: llm-service
  template:
    metadata:
      labels:
        app: llm-service
    spec:
      containers:
      - name: ollama
        image: ollama:latest
        resources:
          requests:
            memory: "24Gi"
            nvidia.com/gpu: "1"
          limits:
            memory: "32Gi"
            nvidia.com/gpu: "1"
        volumeMounts:
        - name: model-cache
          mountPath: /root/.ollama
      volumes:
      - name: model-cache
        persistentVolumeClaim:
          claimName: llm-models-pvc
---
# Vector Database Deployment
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vector-db
  namespace: rag-system
spec:
  replicas: 3
  selector:
    matchLabels:
      app: vector-db
  template:
    metadata:
      labels:
        app: vector-db
    spec:
      containers:
      - name: weaviate
        image: weaviate:latest
        resources:
          requests:
            memory: "16Gi"
          limits:
            memory: "24Gi"
        volumeMounts:
        - name: vector-data
          mountPath: /var/lib/weaviate
      volumes:
      - name: vector-data
        persistentVolumeClaim:
          claimName: vector-db-pvc

Multi-Container Service Mesh

For large deployments, use service mesh for advanced traffic management:

graph TB
    Ingress[Ingress LoadBalancer]

    subgraph ServiceMesh[Istio/Linkerd Service Mesh]
        Control[Control Plane]
    end

    LLM1[LLM Service v1.0<br/>80% Traffic]
    LLM2[LLM Service v1.1<br/>20% Traffic - Canary]
    Vector[Vector DB Service]
    DataConn[Data Connector Service]

    Ingress --> Control
    Control --> LLM1
    Control --> LLM2
    Control --> Vector
    Control --> DataConn

    style Ingress fill:#E8F4FD,stroke:#0078D4,stroke-width:2px,color:#000
    style ServiceMesh fill:#FFF4E6,stroke:#FF8C00,stroke-width:2px,color:#000
    style LLM1 fill:#D4E9D7,stroke:#107C10,stroke-width:2px,color:#000
    style LLM2 fill:#FFFACD,stroke:#FF8C00,stroke-width:2px,stroke-dasharray: 5 5,color:#000

Benefits:

• Canary deployments (new models with 20% traffic)
• Circuit breaking (fail-safe degradation)
• Distributed tracing (latency visibility)
• Automatic retry policies

---

Kubernetes Orchestration Patterns

Pattern 1: Stateless LLM Service

Use Case: Horizontally scaled inference servers

Configuration:
  - Replicas: 3-10 depending on load
  - Each pod: 1-2 GPUs, 24-32GB VRAM
  - Shared model cache (volume)
  - Load balancing: Round-robin or least-connection

Scaling Trigger:
  - CPU > 80% → Add replica
  - GPU memory > 90% → Add replica
  - Latency p95 > 1s → Add replica
  - Response queue > 50 → Add replica

Pattern 2: Stateful Vector Database

Use Case: Persistent vector storage with replication

Configuration:
  - StatefulSet (not Deployment)
  - Persistent volumes per pod
  - Multi-replica for HA
  - Backup strategy: Nightly snapshots

Topology:
  - Pod 0: Primary (read/write)
  - Pod 1: Replica 1 (read-only)
  - Pod 2: Replica 2 (read-only)

  Write operations → Pod 0
  Read operations → Round-robin (Pod 0, 1, 2)

Pattern 3: Data Connector Service

Use Case: Real-time data synchronization

Configuration:
  - CronJob for scheduled imports
  - Service for real-time connectors
  - Init containers for schema setup

Workflow:
  1. Fetch from source (DB, API, file)
  2. Process/tokenize
  3. Generate embeddings
  4. Index in vector DB
  5. Update metadata

---

Versioning & Model Management

Model Versioning Strategy

File Structure:

/models
├── llm
│   ├── phi-3-4b
│   │   ├── v1.0 (current production)
│   │   │   ├── model.gguf (quantized 4-bit)
│   │   │   ├── tokenizer.model
│   │   │   └── config.json
│   │   ├── v1.1 (canary testing)
│   │   └── v0.9 (previous stable)
│   └── mistral-7b
│       ├── v1.0
│       └── v1.1
│
├── embeddings
│   ├── bge-base-en
│   │   ├── v1.0 (current)
│   │   └── v1.1 (canary)
│   └── all-minilm-l6-v2
│       └── v1.0
│
└── indices
    ├── vector-index-v1
    ├── vector-index-v2
    └── vector-index-v3

Blue-Green Deployment

Minimize downtime during model updates:

Current State (Blue):
  ┌──────────────────┐
  │ LLM v1.0 (Prod)  │
  │ Vector DB v1     │
  │ Embeddings v1    │
  └──────────────────┘
         ↑
      Traffic

New State (Green - Staging):
  ┌──────────────────┐
  │ LLM v1.1 (Test)  │
  │ Vector DB v2     │
  │ Embeddings v2    │
  └──────────────────┘
        (offline)

Switchover:
  1. Test Green fully
  2. Redirect traffic to Green
  3. Keep Blue as instant rollback
  4. Blue becomes staging for next version

Canary Deployment

Reduce risk with gradual rollout:

Timeline:
  Hour 0: Deploy v1.1 (10% traffic)
  Hour 1: If metrics good, 25% traffic
  Hour 2: If metrics good, 50% traffic
  Hour 4: If metrics good, 100% traffic

Metrics to Monitor:
  - Error rate < 1%
  - Latency p95 < 1s
  - Model confidence > 0.7
  - User satisfaction > 4.5/5

Rollback Trigger:
  - Error rate > 2%
  - Latency p95 > 2s
  - Hallucination rate > 10%

---

CI/CD for RAG Systems

Build Pipeline

Source Code (Git)
    │
    ▼
1. Lint & Test
   - Code quality checks
   - Unit tests for prompt templates
   - Embedding validation
    │
    ▼
2. Build Containers
   - LLM service image
   - Vector DB image
   - Data connector image
    │
    ▼
3. Security Scanning
   - Vulnerability scan (Trivy)
   - Secret detection
   - Image signing
    │
    ▼
4. Push to Registry
   - Container registry (ACR)
   - Tag: :v1.2.3, :latest
    │
    ▼
5. Deploy to Dev
   - Automated deployment
   - Smoke tests
   - E2E tests
    │
    ▼
6. Deploy to Staging
   - Full test suite
   - Performance testing
   - User acceptance testing
    │
    ▼
7. Deploy to Production
   - Canary rollout (10% → 50% → 100%)
   - Continuous monitoring
   - Instant rollback capability

GitOps Workflow

Infrastructure as Code approach:

1. Developer commits model update
   git commit -m "Update phi-3 to v1.1"

2. PR triggers CI pipeline
   - Build new container
   - Run tests
   - Generate deployment manifests

3. Merge PR to main branch
   - ArgoCD detects change
   - Applies manifests to cluster
   - Monitors deployment health

4. Continuous Compliance
   - Policy checks (no unapproved images)
   - Quota validation (GPU, memory)
   - Audit logging

---

Edge-Specific Deployment Considerations

Offline-First Deployment

For disconnected/air-gapped environments:

1. Prepare on-premises:
   • Download model artifacts
   • Pre-build container images
   • Create deployment manifests
   • Generate offline documentation
2. Deploy to edge:
   • Load container images locally
   • Mount model files from NAS/storage
   • Configure with local DNS
   • Validate in air-gapped environment
3. Updates:
   • Prepare update package offline
   • Validate on staging cluster first
   • Deploy with manual approval

Resource Constraints

Optimize for edge hardware (vs. cloud):

Cloud Deployment:     Edge Deployment:
─────────────────     ────────────────
- 8+ GPUs available   - 1-2 GPUs per node
- 128GB+ VRAM         - 24-32GB VRAM
- Unlimited scaling   - Fixed hardware
- High availability   - Local resilience

Strategies:
  1. Model quantization (4-bit → 75% size reduction)
  2. Smaller base models (7B vs. 70B)
  3. Local caching (reduce network calls)
  4. Batch processing (amortize overhead)
  5. GPU sharing (time-slicing for multiple models)

Low-Bandwidth Considerations

When network bandwidth is limited:

1. Model Delivery:
   • Pre-download models to edge
   • Use compression (20-50% reduction)
   • Incremental updates (delta sync)
2. Data Ingestion:
   • Batch imports (daily/weekly vs. real-time)
   • Lossy compression for non-critical data
   • Local processing before cloud sync
3. Monitoring:
   • Local metrics collection
   • Batch telemetry upload (hourly)
   • Low-bandwidth observability (sampling)

---

Troubleshooting Deployment Issues

Common Issues & Solutions

Issue	Cause	Solution
High latency (>2s)	Model too large	Reduce quantization, use smaller model
OOM errors	Insufficient VRAM	Enable disk offloading, batch smaller
Vector search slow	Poor index	Rebuild with HNSW, reduce vectors
Models not loaded	Network timeout	Pre-download, check storage
High error rate	Hallucination	Increase retrieval context, improve prompt

Health Checks

Kubernetes liveness & readiness probes:

livenessProbe:
  httpGet:
    path: /health
    port: 8000
  initialDelaySeconds: 60
  periodSeconds: 10
  timeoutSeconds: 5

readinessProbe:
  httpGet:
    path: /ready
    port: 8000
  initialDelaySeconds: 30
  periodSeconds: 5
  successThreshold: 1

Health Endpoints:

GET /health
  Response: {"status": "healthy", "uptime": 3600}

GET /ready
  Response: {"ready": true, "models": 3, "vectors": 1000000}

GET /metrics
  Response: Prometheus metrics (latency, throughput, errors)

---

Related Topics

• Main Page: Edge RAG Implementation
• Vector Databases: Vector Databases & Indexing
• LLM Optimization: LLM Inference Optimization
• Operations: RAG Operations & Monitoring
• Assessment: RAG Implementation Knowledge Check

---

Last Updated: October 21, 2025