The foundation
NVIDIA describes its Blueprints as reference solutions — flexible starting points that developers adapt and extend into production. That downstream role is Enclave's entire product: the Blueprint is the retrieval engine, and Enclave is the governed application built on top of it. Enclave tracks upstream Blueprint releases, so customer deployments can benefit from validated upstream improvements in ingestion, retrieval, and inference.
The stack
Enclave runs as four hardened services above the NVIDIA stack. Enclave builds on upstream NVIDIA services without forking them, adding deployment configuration and operational safeguards around the stack.
The orchestration
The Blueprint provides the retrieval-and-generation pipeline. Enclave adds a governed supervisor loop that routes a turn across retrieval, analysis, and connected tools:
Enclave's LangGraph orchestrator does not replace that pipeline — it makes it one of several specialists. A supervisor routes each turn with a structured decision; specialists do the work and report back; the supervisor looks at what has been gathered and routes again — up to six hops, under wall-clock and search budgets — before composing the answer. The Blueprint's round trip is what the rag specialist calls:
And because the loop runs inside the platform, governance is woven through every hop — the retrieval filter is built server-side from the asker's permissions before the engine is ever called, and the whole turn lands in the audit log:
The division of labor
The NVIDIA RAG Blueprint is deliberately decomposable — a modular reference architecture, not a finished enterprise application. Its own documentation describes the included interface as a sample, not a production surface. That is by design: the application layer is where independent software vendors are meant to build. This table is the honest split.
| Capability | NVIDIA RAG Blueprint | Enclave adds |
|---|---|---|
| Ingestion | Multimodal extraction — text, tables, charts, images | Upload lifecycle, per-corpus collections, connector-driven sync with deletion reconciliation, chunking tuned per corpus |
| Retrieval | Hybrid vector search with reranking | Permission-aware scoping per user and group, enforced server-side on every query; multi-collection strategy under budgets |
| Generation | Accelerated inference with citations | Agent orchestration (routing, multi-step tool use, sandboxed analysis), answer verification, streaming execution trail |
| Identity | — | OIDC single sign-on, role- and group-based access control, group sync from the identity provider |
| Interface | Sample interface for evaluation | Production chat application (white-label), projects/workspaces, admin console |
| Governance | Optional guardrails | Append-only audit trail with per-turn permission scope, usage accounting, guardrail profiles, prompt-injection defense |
| Quality | — | Scored evaluation harness, customer gold sets with negative controls, written acceptance criteria, quarterly quality reports |
| Operations | Docker Compose / Helm reference deploys | Hardened deployment profiles, production mTLS for Enclave service connections, secrets management, backup/restore runbooks, managed-operations tier |
System requirements
Tested single- and dual-GPU reference configurations, with larger systems sized to your workload during the assessment. The $2,500 assessment turns your corpus, user count, and document mix into an exact bill of materials your hardware partner can quote.
| Configuration | GPUs | What runs | Fits |
|---|---|---|---|
| Pilot / demo reference | 2× NVIDIA RTX PRO 6000 Blackwell (96 GB) | gpt-oss-120b served locally on one GPU; retrieval (embedding + reranking) and the full document-extraction stack — page layout, tables, charts, OCR for scans — on the other | A pilot team through departmental rollout |
| Production | 4× RTX PRO 6000; H200/H100 alternatives assessed per workload | Adds LLM concurrency and headroom for connector sync and sustained ingestion alongside serving | Larger deployments; concurrency and ingestion capacity validated during the assessment |
| Evaluation minimum | 1× RTX PRO 6000 (96 GB) | The full platform with the 120B model serving locally and text-first extraction; structured table/chart parsing and OCR staged rather than concurrent. We have measured this shape to the gigabyte — the 120B model and the multimodal extraction pipeline cannot share one card, which is exactly why the reference configuration is two. | Technical evaluation and development — not a production shape |
About the chat model. gpt-oss-120b is OpenAI's open-weight reasoning model: 117 billion parameters in a mixture-of-experts design, Apache 2.0 licensed, built to fit a single 80 GB-class GPU. NVIDIA publishes it as an NVIDIA NIM microservice with NVIDIA AI Enterprise support, which is how the NVIDIA edition serves it — the same OpenAI-compatible API the rest of the Blueprint already speaks, with the weights resident on the customer's GPU and no external call in the path. Extranetic Inc. is a member of the OpenAI Partner Network.
Host & software: x86-64 Linux (Ubuntu 22.04/24.04), NVIDIA driver + container toolkit, Docker Compose · 24+ cores, 256 GB+ RAM, 1 TB NVMe recommended per node · an OIDC identity provider (Microsoft Entra ID, Keycloak, or equivalent) · NVIDIA AI Enterprise licensing per GPU for the NIM runtime. Model inference, retrieval, storage, and audit run locally. Connections to your identity provider and source systems are configured to match your network policy.
Why this matters to the ecosystem
Every NVIDIA-edition deployment requires an NVIDIA AI Enterprise subscription, licensed per GPU and resold with the engagement — recurring NVIDIA software revenue on hardware that would otherwise run an unsupported stack.
Enclave converts Blueprint evaluations into production deployments — the exact pilot-to-production step the Blueprint is designed to seed. We track upstream releases and validate upgrades before rolling them into customer deployments.
Enclave serves the $30–250k deployment band: below the threshold where global systems integrators engage, above what an IT team can self-assemble from a repository. Coverage where the ecosystem is thinnest.
Public-sector and regulated buyers are Enclave's focus: environments governed by DFARS 252.204-7012 and NIST SP 800-171, no-cloud-AI policies, and disconnected networks — where a supported, fully on-premises stack is the requirement, not a preference.
INDUSTRY FOCUS · PUBLIC SECTOR & REGULATED MID-MARKET
Provenance
Enclave's founder, Rian Sousa, architected the initial end-to-end design of a U.S. state treasury's private AI platform and worked with a team to develop it from scratch — multi-agent retrieval over sensitive financial data, local models, enterprise SSO and role-based access. The architecture evolved as the team grew and the scope expanded. Enclave is an independent, AI-native product built from the ground up with the Blueprint as its backbone from inception. Enclave is a member of the NVIDIA Developer Program; NVIDIA Partner Network (ISV) enrollment is underway.
Next step
A live walkthrough of Enclave on NVIDIA hardware — sign-in to cited, permission-scoped answer — takes about fifteen minutes.
Contact us