The platform, module by module
Enclave is built as discrete modules on top of the NVIDIA RAG Blueprint. This page lists every one that ships today, grouped the way a buyer reads them, with the tier each belongs to. What is specified but not built is at the bottom, under its own heading.
How to read this page
Every card below describes something that exists in the product and has been exercised in our own test and browser-QA suites. Where a module ships in part, the card says what ships and names the boundary rather than describing the finished version. Nothing on a roadmap appears as a feature.
The honest limit on all of it: proof to date is against our own development stack and document corpora. No customer identity tenant, SharePoint or Confluence estate, air-gapped site or multi-GPU installation has been used yet, and retrieval quality is a measured range, not a guarantee. Ask for the measurements — they exist, and we will show them.
Tiers. Each card carries the offer tier its module belongs to: Pilot is the 30-day proving deployment, Production the full governed rollout, and Production+ the deeper modules a regulated estate adds on top. Tier unassigned means the module's specification carries no tier line — not that it is free. Tiers and pricing.
Identity & access
Staff sign in with the company account they already have, and every request to the assistant is checked against that identity.
Each person only ever gets answers from the document sets their group is allowed to read, enforced at the retrieval layer rather than hidden in the screen.
Site and library permissions from the source system are mirrored into the assistant on every change-driven sync pass, so retrieval stays scoped to what the person asking could open themselves. The depth is deliberately bounded: library and site level, no folder-scoped unique permissions, no sharing links, no external users, and no nightly full re-sweep — mirroring follows changes.
Every question, every document retrieved and every admin change is recorded in a tamper-evident log an auditor can read and export.
Admins see who is using the assistant and what it costs, broken down by person, team and model.
Documents carry a sensitivity marking and people carry a clearance, and that clearance is enforced in retrieval — a restricted passage never reaches someone who is not cleared for it, and every clearance change is audited. Markings are declared by you, not discovered: nothing scans a document to classify it.
When you want a person to sign off before something leaves — a generated document or canvas export, an emailed digest, a tool that writes somewhere, or in a narrow classified case an answer — an admin writes that as a policy, the output is parked and the requester told at once, a reviewer in the named group sees it under their own permissions and approves or rejects with a reason, a time limit escalates it, and every step is in the ledger. Off by default: a deployment ships with no policies and nothing gated.
Connectors & ingestion
Document libraries sync in on a schedule, so answers come from the living corpus rather than a one-time upload: new and changed documents are picked up each run, with per-run counts on the admin surface. Removal is the open edge — a document deleted upstream is not yet withdrawn automatically. Every run to date has been against a Graph fixture, never a customer tenant.
Upload a document and watch it go from uploaded to searchable, with honest progress, refusals you can understand, and no silent orphans left behind.
The assistant can reach the business tools you already run — ticketing, wikis, internal services — through a standard connector instead of a fork of the product. Tool discovery and invocation are proven against fixtures; each external service you connect gets its own acceptance pass.
Extraction and chunking ship as five built-in presets, and the fit to real documents is measured rather than asserted: 91 files across 29 formats index and pass their content checks with full-page OCR enabled. The per-engagement tuning report has no generator yet, and behaviour on a corpus larger than a single workstation is unmeasured.
Recurring document pipelines run on a schedule — pull, extract to typed records, index — instead of somebody remembering to upload a folder.
For any answer, generated document, canvas version or pipeline record, the full chain behind it on one page: which files and versions, which ingestion run, which permissions were in force, which model — assembled from the records the platform already keeps, with anything you may not see named as withheld rather than hidden, and an auditor-readable report whose own digest is in the ledger.
Chat & workspace
A branded chat application your staff can actually use: streaming answers with citations you can open, uploads, cancellation, dark mode and a working phone layout.
Open a panel and see what the assistant searched, which documents it consulted and how the answer came together.
Team workspaces that hold their own documents, conversations and members, so a project's context stays with the project.
One console for grants, users and groups, connector status, pipelines, usage and the audit log.
Brand the assistant from a settings page, and drop the same governed chat into an application you already run with a few lines of pasted script.
Write a document beside the chat: the assistant proposes edits you accept or reject one at a time, versions snapshot and restore, and the result exports to Word. Source traceability inside the canvas is still being proven — the markers that tie a claim to its source have open defects, so a canvas claim is a draft to check, not a cited fact.
Agents & generation
The application brain: it decides what to do with a question, keeps conversation state, streams progress and can be cancelled mid-answer.
The retrieval specialist: it searches the document sets you are allowed to see, under a budget, and answers with citations or says it does not know.
Customer-specific actions — lookups, record queries, ticket creation — become typed tools the assistant can use under guard, instead of forks of the product.
Ask for a calculation or a chart over your own data files and get one back, computed in a sealed sandbox on your own box with no network.
One deployment can mix models — NVIDIA NIMs, open weights, any OpenAI-compatible endpoint — with health checks and failover, so a single backend going down is not an outage.
Pick a template and the assistant drafts the document section by section from your own corpus, citing every claim and marking anything it could not source, then exports to Word or Excel.
Standing watchers that check your corpus on a schedule, or whenever a pipeline or connector run finishes, run as you with your permissions re-checked every time, and tell you in the app or by email what they found — the answer is yours alone; admins see facts and counts only.
Ask a question that needs more than one search, and the platform plans the report, researches each sub-question over the collections you can read, under your own permissions, writes each section with citations into one registry, and ends every report with a “Gaps / no source found” section naming what it could not answer — delivered as a private canvas of your own or a Markdown file only you can download. Off by default; a deployment turns it on and sets the budgets.
Evaluation & safety
Acceptance is a scored run against a gold set you own, with a report and a verdict — not a demo and a handshake.
Instructions hidden inside a document cannot make the assistant break its rules, leak its prompt or exfiltrate data.
Your gold-set questions re-run on a schedule as a real user, every answer scored (grounding, citations, and four judged metrics), stored as an append-only trend with one honest classifier, drawn with gaps where nothing was measured, and a notification with a link when quality regresses — plus a curated queue for turning real questions into new gold-set entries.
Anyone can rate an answer and flag a wrong source; feedback is tied to the exact answer and its retrieval, rolled up nightly into candidates a reviewer promotes — never an automatic retune.
Deployment & operations
The whole platform stands up on your own hardware from one command set, with health checks, secrets, certificates and telemetry wired in.
The runbooks and tooling behind a managed service — model refresh, index maintenance, backup and restore, monitoring, incident response — with backup and restore exercised in the QA suite. The service commitments around them are not yet validated: no deployment has run under a managed contract, so the cadence is untested.
When a pipeline, a generated document or an ingestion needs attention, the right people hear about it — in the app and, when a relay is configured, by email — with a link back to the exact item and never any document or answer text.
Readiness
A written, code-cited account of which security controls are in place and which are not, including an authorisation matrix over every API operation. It is an honest readiness assessment and nothing more: no SOC 2, no ISO 27001, no CMMC, no independent audit and no penetration test.
Not yet available
These modules have written specifications and an agreed shape, and nothing more. They are listed so the gaps are visible, not as things you can buy today. None of them is deployable, and none should be counted on in a plan. There are 4 of them:
A specified module can be brought forward inside a Production engagement when it is what the engagement needs. That is a build, priced as one — not a switch we flip.
We will show you the modules above running against your own corpus, on hardware you control, and hand you the measurements.
Schedule a demo