Private Deployment Options
A private deployment of Vectara is designed to work with your existing infrastructure choices, offering flexible deployment models to match your organization's specific requirements.
Deployment models
Traditional on-premises
Run Vectara in your data center. Air-gapped supported.
- Physical servers or VMs
- No internet required after initial setup
- Your hardware, your rules
Private cloud (VPC)
Deploy in your existing cloud account:
- AWS: EKS or self-managed Kubernetes in your VPC
- GCP: Private GKE clusters with your security policies
- Azure: AKS with your network controls
- Other: Various Kubernetes distributions supported
What you need
Kubernetes
Container-based deployment on your cluster:
- Enterprise distributions supported
- Managed services: EKS, GKE, AKS
- Standard Kubernetes environments
Storage
- Block storage: Your existing CSI driver
- Object storage: Any S3-compatible API (MinIO, Ceph, actual S3)
Observability
Standard metrics and logging:
- Prometheus-compatible metrics
- Log export via Fluentbit
- OpenTelemetry trace support
GPU acceleration (optional)
NVIDIA GPU support available:
- Accelerated embedding generation
- Improved reranking performance
- MIG partitioning compatible
How we deploy
Infrastructure as Code
Everything deploys via Terraform or Helm:
- Version-controlled configurations
- Your Git, your process
- Test in staging, promote to prod
- Rollback is just
git revert
Updates on your schedule
- Regular version releases provided
- Customer-controlled upgrade timing
- Canary deployment support
- No forced updates
Model flexibility
Choose your LLMs based on your requirements:
- Use ours: Vectara-optimized models included
- Bring yours: Connect GPT-4, Claude, Gemini, or Llama
- Mix and match: Different models for different use cases
- Air-gapped: Local models only, no external calls
Integration ready
Compatible with existing infrastructure:
- Auth: OIDC/SAML via your IdP
- APIs: Standard REST API
- CI/CD: ArgoCD, Flux, Jenkins compatible
- Backups: Velero-compatible