Practical guide to AI-powered catalog enrichment: blueprints from NVIDIA & Nebius

The NVIDIA Retail Catalog Enrichment Blueprint combines product images, existing catalog fields and optional documents to generate richer product content. Outputs include titles, descriptions, categories, FAQs, localized images and 3D assets.

In the live demo, Manuel Gomez uploads an air fryer image and its PDF manual. The workflow generates text, produces images for different markets and creates a 3D asset. Brand instructions can guide the wording, while existing catalog information provides additional context.

A vision-language model interprets the product image, while a language model generates text and applies brand or localization instructions. An embedding model supports retrieval from documents. Separate models edit product images and generate 3D assets.

The demonstrated architecture uses NVIDIA Nemotron models, FLUX Kontext and Microsoft TRELLIS. A vision-language model also evaluates generated images. The blueprint is modular, so teams can adapt individual components to their product categories, quality requirements and deployment environment.

The workflow extracts text from uploaded PDFs, divides it into chunks and indexes those chunks using embeddings. Relevant passages can then be retrieved to support answers about the product.

In the demonstration, the air fryer manual supplies information for FAQ generation. The same approach can incorporate policy documents. Retrieval makes selected information available without placing an entire document into the model’s context, while the generated answers still need checks against the source material.

The session demonstrates a vision-language model acting as a judge to compare generated images with the original product and score their quality. The speakers also discuss applying evaluation to text and checking whether outputs follow catalog and business requirements.

In the Q&A, Antonio Martinez describes using thresholds, limited retries and human review for outputs that remain below the required standard. Teams need to establish and test those criteria. A judge score is a review signal, not a guarantee that every product detail is correct.

A GPU virtual machine provides a starting point for trying the blueprint. For a larger service, separating model components allows teams to scale the parts of the workflow that receive the most demand.

Manuel discusses managed Kubernetes, managed inference through Token Factory, and serverless deployment options. For example, generating images for many markets may put more pressure on the image model than on text generation. Scaling those services independently can help match resources to the actual workload.

Batch processing suits work that can run on a schedule, such as enriching a seasonal product release or refreshing catalog entries overnight. An online service is more appropriate when individual requests need a prompt response.

The session explains how jobs can process a defined set of products and stop when the work finishes. It also discusses using interruptible capacity for suitable offline workloads. The choice should reflect completion deadlines, interruption tolerance and how frequently the catalog changes.