Applies to SUSE AI Factory with NVIDIA

3 What pre-built AI blueprints are included? #

SUSE AI Factory with NVIDIA includes the following pre-built AI blueprints:

3.1 NVIDIA AI-Q with RAG #

This blueprint deploys the NVIDIA retrieval-augmented generation (RAG) blueprint with the nvidia/llama-3.3-nemotron-super-49b-v1.5 model. It also configures the NVIDIA AI-Q profile with a deep-research agent for Foundational RAG (FRAG).

We support deploying AI-Q integrated with RAG using local NIMs. Although the blueprint page includes complete and validated integration steps, we summarize them for a quick reference here:

Create a custom ConfigMap. Create a ConfigMap in the ns-aiq namespace holding the updated configuration, pointing AI-Q at the locally hosted NIMs:
```
> kubectl create configmap aiq-web-frag-config \
  -n ns-aiq \
  --from-file=config_web_frag.yml=config_web_frag_2_1_0_updated.yaml
```
Tip
You can download the config_web_frag_2_1_0_updated.yaml file from the SUSE AI Factory with NVIDIA examples repository.

Override the AI-Q Helm values. Override the AI-Q back-end to load the custom configuration and talk to the RAG services. Replace <RAG_NAMESPACE> with the namespace where RAG is deployed. It is ns-aiq if RAG and AI-Q are deployed together by this blueprint. In such cases, short service names (for example, http://rag-server:8081/v1) also work. With the following configuration, the AI-Q back-end uses the custom file config_web_frag.yml defined in the ConfigMap.

aiq:
  apps:
    backend:
      env:
        CONFIG_FILE: configs/config_web_frag.yml
        RAG_SERVER_URL: http://rag-server.<RAG_NAMESPACE>.svc.cluster.local:8081/v1
        RAG_INGEST_URL: http://ingestor-server.<RAG_NAMESPACE>.svc.cluster.local:8082/v1
        EMBED_BASE_URL: http://nemotron-vlm-embedding-ms.<RAG_NAMESPACE>.svc.cluster.local:8000/v1
        EMBED_MODEL: nvidia/llama-nemotron-embed-vl-1b-v2
      volumes:
        # Stock volume — must be kept, the chart replaces the whole list.
        - name: postgres-init
          configMap:
            name: aiq-postgres-init
        # Override the default config with our custom ConfigMap.
        - name: aiq-web-frag-config
          configMap:
            name: aiq-web-frag-config
      volumeMounts:
        # Replace only the bundled config file while leaving the rest of
        # /app/configs unchanged (subPath mounts just the single file).
        - name: aiq-web-frag-config
          mountPath: /app/configs/config_web_frag.yml
          subPath: config_web_frag.yml

3.2 NVIDIA RAG (minimal, low-GPU) #

This is a minimal NVIDIA RAG blueprint for low-GPU footprints.

It replaces the default 49B Nemotron model with the Llama 3.2 3B model. It also disables structural optical character recognition (OCR) and table and chart extraction to reclaim approximately 18 GB of VRAM.
It forces CPU vector search and indexing.
It drops chart resource bounds to fit a single node and pins Milvus standalone (text-only RAG).

This blueprint is not recommended for production use. The tracing and observability stack is disabled, and the token budget is tuned for small large language models (LLMs) instead of high throughput.