NVIDIA Blueprints

NVIDIA AI Blueprints are reference architectures with pre-built, validated components for common enterprise AI use cases. Each blueprint includes NIM microservices for model serving, NeMo for customisation, and curated integration patterns that accelerate deployment from months to weeks. They are designed to run on NVIDIA AI Enterprise software on DGX, certified OEM servers, or cloud GPU instances. Think of them as production-grade starting points that we customise to your specific data, infrastructure, and compliance requirements.

The Enterprise Research Copilot enables knowledge workers to query vast document repositories using natural language. It combines dense retrieval with reranking models to find relevant passages across millions of documents, then uses a large language model to synthesise answers with citations. We deploy this for law firms searching case law, pharmaceutical companies reviewing clinical literature, and financial institutions analysing regulatory filings. The key differentiator from generic RAG is the multi-stage retrieval pipeline that handles domain-specific terminology and document structures.

The RAG Agent blueprint goes beyond simple retrieval-augmented generation by adding agentic capabilities — the model can plan multi-step research tasks, use tools like calculators or APIs, and iteratively refine its search strategy based on intermediate results. It uses NVIDIA NIM for model serving, NeMo Retriever for embedding and reranking, and a tool-use framework for extending the agent with custom capabilities. We configure guardrails to constrain agent behaviour within your security policies and monitor agent trajectories for quality assurance.

The Video Search and Summarisation (VSS) Agent processes video content at scale — extracting visual features, transcribing audio, detecting objects and events, and building a searchable index across your entire video library. Users can query video archives with natural language questions like "show me all safety incidents near loading dock 3 in the last month" and get timestamped results with summaries. We deploy this for manufacturing quality inspection, security surveillance, media asset management, and compliance monitoring.

A standard implementation follows a predictable timeline. Week one covers requirements alignment and infrastructure provisioning. Weeks two and three handle core deployment, data ingestion pipeline setup, and initial model configuration. Weeks four through six focus on customisation — fine-tuning retrieval models on your domain data, building integrations, and implementing guardrails. Weeks seven and eight are dedicated to load testing, security review, and user acceptance testing. Most blueprints reach production readiness in eight weeks, with simpler deployments completing in as few as four.

Extensively. Blueprints are starting points, not finished products. We customise at every layer: swapping foundation models for domain-specific variants, training custom embedding models on your terminology, adding data connectors for your specific systems, implementing business logic in the orchestration layer, and designing user interfaces for your workflows. The blueprint architecture is modular specifically to enable this — each NIM microservice can be replaced or extended independently without affecting the rest of the pipeline.

Requirements vary by blueprint and scale. For development and proof-of-concept, a single NVIDIA A100 or H100 GPU with 80GB VRAM is typically sufficient. Production deployments for the Enterprise Research Copilot serving hundreds of concurrent users usually require two to four H100 GPUs. The Video Search blueprint needs additional compute for video processing — typically A100 or L40S GPUs dedicated to the vision pipeline. We can deploy on DGX systems, certified OEM servers from Dell, HPE, or Lenovo, or on cloud instances from any major provider.

We implement a model lifecycle management process around each blueprint. New model versions are deployed to a staging environment first, evaluated against a held-out test suite specific to your use case, and promoted to production only after passing accuracy, latency, and safety benchmarks. Rollback is instantaneous because we maintain the previous version in a warm standby state. NIM microservices support blue-green deployment natively, so model updates happen with zero downtime and no disruption to end users.

Yes. NVIDIA Blueprints are designed with extensible data connectors. We build integrations with common enterprise systems including SharePoint, Confluence, S3, Azure Blob Storage, Snowflake, Databricks, and relational databases. For real-time use cases, we connect Kafka streams or CDC pipelines directly to the ingestion layer. Authentication integrates with your existing identity provider through OIDC or SAML. The goal is to make the blueprint a natural extension of your existing data ecosystem rather than a separate silo.

This blueprint creates a self-improving loop. A large foundation model handles initial production traffic while logging inputs, outputs, and quality signals. High-quality interactions are automatically curated into training datasets. Smaller, specialised models are periodically distilled from these datasets using NeMo. Once a distilled model meets quality thresholds, traffic is routed to it at dramatically lower inference cost. The flywheel continues as the distilled model generates new training signal. In practice, this drives inference cost reductions of 60 to 98 percent while maintaining or improving output quality over time.