Generative AI is moving from experimentation to production as organizations build AI assistants, knowledge discovery tools, agentic workflows, and retrieval-augmented generation (RAG) applications that connect large language models to enterprise data. As these systems become business-critical, protecting the data, application state, and dependencies that make them useful has become a fundamental consideration.
Red Hat is helping customers explore scenarios like this through Red Hat AI QuickStarts. These are practical, focused examples designed to help organizations deploy and evaluate real-world AI use cases on the Red Hat AI platform. The catalog covers use cases ranging from RAG and AI-powered support to agents, model evaluation, observability, secure inference, and other enterprise AI scenarios.
One of the newest QuickStarts, Protect Enterprise RAG Knowledge with Veeam Kasten on OpenShift, brings Veeam data resilience directly into that conversation by demonstrating how teams can protect and recover a RAG application running on Red Hat OpenShift, including the enterprise knowledge and application state the system depends on. For teams evaluating AI on OpenShift, the guide offers a practical way to build in resilience from the outset rather than addressing it after an application reaches production.
From AI concepts to deployment methodology
Red Hat AI QuickStarts are designed to help customers move beyond high-level AI concepts and explore deployable examples on the Red Hat AI platform.
The Veeam Kasten QuickStart is part of Red Hat’s Adopt and Scale AI collection and focuses on the increasingly important topic of understanding how to protect the knowledge that makes a RAG application valuable. The theme is relevant to the broader topic of AI since production RAG applications are more than just a model connected to a database: They’re distributed applications made up of source documents, vector embeddings, databases, object storage, ingestion pipelines, model-serving configuration, Kubernetes resources, secrets, policies, and other dependencies.
When these components fall out of sync, the application may continue running while retrieving stale or incomplete information, potentially reducing the accuracy and reliability of its responses.
Protecting the complete RAG application
The Red Hat AI QuickStart, Protect Enterprise RAG Knowledge with Veeam Kasten on OpenShift, provides a deployable RAG application running on Red Hat OpenShift AI. It demonstrates how Veeam Kasten protects the application’s Kubernetes resources and the state included in the protection policy. The example includes Llama Stack, PostgreSQL with the pgvector extension, object storage, data pipelines, model-serving services, and the Kubernetes resources that connect these components.
Rather than treating each component as an isolated infrastructure resource, Kasten protects the application state required to recover the RAG environment coherently. That includes:
- Vector data and metadata stored in PostgreSQL and PGVector
- Source documents used by the ingestion pipeline
- AI application configuration
- Pipeline definitions, history, and artifacts
- Model-serving configuration
- Kubernetes resources, including Deployments, StatefulSets, Services, Routes, ConfigMaps, Secrets, RBAC, and custom resources
Why does all this matter? Because backing up only the vector database could leave embeddings disconnected from the very source documents that created them. Protecting only object storage could require teams to rerun ingestion and embedding jobs. And protecting Kubernetes manifests alone would leave the application configuration intact while not recovering the enterprise knowledge behind it.
The QuickStart demonstrates an application-level approach, delivering a path to recover the RAG deployment as a coordinated system rather than a collection of disparate components.
Extending protection beyond storage snapshots
Stateful AI services may require more than a volume snapshot. Depending on the workload, creating a usable recovery point can require application-consistent operations, logical database backups, sequencing, and coordination across dependencies.
The guide demonstrates this approach using Veeam Kasten’s application-aware data protection capabilities. Through Kanister, the protection workflow can perform a logical backup of the PostgreSQL database containing the pgvector data. This can make recovery less dependent on a particular storage implementation and support additional use cases, including migration, cloning, and development and testing workflows.
More broadly, the example illustrates an important principle for enterprise AI: An effective resilience strategy should account for application state and dependencies, not only the underlying infrastructure.
Building resilience into the AI lifecycle
One of the most useful patterns in the QuickStart happens before a recovery is ever required.
For RAG applications, the ingestion pipeline is where the enterprise knowledge base changes. New documents may be cleaned, broken apart, embedded, classified, and written into the vector database.
That process introduces operational risk. A problematic document set, failed embedding process, corrupted data, or other ingestion issues can reduce the quality of the information the AI system retrieves.
With Veeam Kasten, teams can trigger a protection workflow before ingestion modifies the knowledge base. In the QuickStart, the ingestion workflow uses the Veeam Kasten API to initiate protection before modifying the knowledge base. If something goes wrong during the ingestion process, teams have a recovery point immediately before the change happened. That transforms backup from a separate infrastructure task into part of the AI application’s operational lifecycle.
And that is precisely where this example becomes particularly useful for AI platform teams. Resilience does not have to operate only on a background schedule; it can be integrated directly into the workflows that modify enterprise knowledge.
Simplifying RAG recovery
Without a coordinated recovery point, restoring a RAG application can require multiple teams and disconnected recovery steps. Teams may need to restore the database, recover source documents, recreate pipelines and embeddings, and reconcile Kubernetes configuration, routes, secrets, and services before the environment can be validated and returned to service.
Veeam Kasten demonstrates a simpler model that allows teams to recover the protected application state from a coordinated restore point. For customers, that approach enables:
- Reducing recovery time and helping organizations meet their RTOs
- Lower operational overhead compared to a manual rebuild
- Reduced need to rerun expensive ingestion and embedding jobs
- Repeatable and testable recovery workflows
- Application recovery or mobility across compatible namespaces and clusters
- Exporting backups to a separate, appropriately secured location for stronger resilience
Rather than rebuilding the knowledge layer piece by piece, teams can focus on recovering the entire working application.
Why this matters as enterprise AI moves into production
An AI proof of concept can often tolerate experimentation, manual recovery, and temporary data loss, but production applications can’t. As organizations begin relying on RAG for internal knowledge, customer support, sales enablement, IT operations, policy lookup, or other enterprise use cases, platform teams need to answer a new set of operational questions:
- Can the vector database be restored to a recovery point created before corruption occurred?
- Can the knowledge base be rolled back after a bad ingestion run?
- Can the application be restored into another namespace or cluster?
- Can source documents and embeddings be protected together?
- Can recovery be tested before an incident occurs?
- And can all of that happen without rebuilding the application’s knowledge base from scratch?
The Red Hat AI QuickStart featuring Veeam Kasten gives customers a practical environment for exploring those questions on Red Hat AI. It also reinforces the important principle that’s worth repeating:
AI resilience is most effective when it’s designed into the application lifecycle from the beginning.
RAG applications rely on more than just a model or a database. They require enterprise knowledge, application configuration, pipelines, and persistent data to stay synchronized and recoverable. Protecting that complete application state is essential to maintaining trust in the AI experience.
Get started with the AI QuickStart
These useful guides deliver a practical way to explore enterprise AI use cases on the Red Hat AI platform, and the Veeam Kasten QuickStart extends that approach by demonstrating how data resilience can be built into a real RAG application.
Inside, there’s a GitHub repository and a Helm-based deployment, so teams can explore the architecture and protection workflow directly in an OpenShift AI environment. For organizations building or evaluating RAG, it provides a useful reference for designing protection, recovery, and application mobility into AI from the beginning.
As enterprise AI moves from experimentation to production, resilience should be part of the platform conversation. Red Hat OpenShift AI and Veeam Kasten give customers a practical way to explore how the data, configuration, and application state behind RAG can remain recoverable, portable, and ready for what comes next.
Explore the resources:
Protect Enterprise RAG Knowledge with Veeam Kasten on OpenShift | Red Hat Documentation
3-minute Video Demo: Protect and recover a pgvector database with Veeam Kasten
The post Resilience by Design: Protecting Enterprise RAG on Red Hat OpenShift AI appeared first on Veeam Software Official Blog.
The post Resilience by Design: Protecting Enterprise RAG on Red Hat OpenShift AI appeared first on Veeam Software Official Blog.