Deployment
One container. Same behaviour wherever you run it.
Model, data, retrieval, connectors, enforcement and orchestration ship as one container. Evaluate it as managed software; run it with no route to the internet.
What is in the box
Every layer, in one unit.
Nothing here is a separate installation, a sidecar you operate, or a bolt-on you buy later. That is why the deployment story is short.
The model layer
A configured model, not a bundled one — commercial, open-weight, hosted by a provider or self-hosted on your own hardware. It is a layer inside the container, so swapping it disturbs nothing above it.
Your data and retrieval
The knowledge base, the retrieval configuration and the grounding checks that run over it, pointed at your own corpus and your own search infrastructure.
Connectors and actions
The nodes that reach your systems, plus REST for everything else, with credentials read from the secrets you keep on the Configure tab rather than baked in.
Guardrails and audit
Screening, redaction, grounding gates and the audit stream — in the request path, in the same unit, running identically in every topology.
Orchestration
The agent runtime and the editor that authors it, so building the next application does not mean standing up another environment.
How to get it
Three marketplaces, one appliance, or your own cluster.
Every route is the same software with the same feature set — a purchasing path through a vendor you already have a contract with, or hardware you already own. Routes, not editions; a few plans are listed on one cloud only.
Slide 1 of 5
01 / 05
AWS MarketplaceProcure through your existing AWS agreement and draw on the model catalogue you already have access to there.
The model is a layer
Add an LLM: pick a platform, then a model.
Each card is a platform in the Add an LLM dialog. Choose one, then any model it offers — the dialog shows what that model can and cannot do before you add it. Add several and NeuralSeek load-balances each function across the models that serve it.
Amazon BedrockModels served through your AWS account on Amazon Bedrock, under the agreement you already have.
In the console
Swap it live.
One routing category, its own configuration, a model added from a provider list — the walk-through.
With multi-agent routing, configuration flows through a tree of categories: each category can carry its own configuration for the levels beneath it, or inherit the default. Open a category and its modal holds the knowledge-base connection, the tuning and the LLM Details. Add an LLM lists the providers, the new model becomes a card, and its functions are ticked in place — then Save.
A category, its own configuration
Once it is running
Why the packaging is the compliance answer.
A security review is really asking whose controls apply, not what the software does. For a container inside your boundary, the answer is yours.
It inherits the boundary you put it in
Drop it into an environment that is already authorised and it operates under that environment's network controls, key management, logging and incident process. The honest answer to “is it certified?” is a question about your tenant.
The same feature set in every topology
There is no sovereign edition and no capability held back for the deployments that are hardest to reach. Managed service or fully air-gapped, it is the same isolation, audit and enforcement primitives.
One way to ship AI, not one per team
The same governed container carries any number of applications, each inheriting the same enforcement and the same audit trail. The second use case does not start over, and the tenth does not need its own security review.
Three questions procurement asks first.
All three are about the shape of the thing being bought rather than what it can do, and all three have short answers.
Do we get a lesser version if we do not use your preferred cloud?
There is no preferred cloud and no reduced version. The marketplace listings and the Fusion appliance are routes to the same container; a cluster you already operate gets the same one. Deployments differ in topology and who administers it — never in feature set.
Can it run with no internet connection at all?
Yes, as a standard topology rather than a special engagement. Fully isolated on-prem sits on the same list as managed service — and an air-gapped deployment has no vendor kill switch, because there is nothing for a vendor to reach.
What are we actually taking on operationally?
Containers on Kubernetes or OpenShift, with the model, retrieval, connectors, enforcement and orchestration inside — not five components to keep in step. Secrets live in the configuration, the audit stream exports to your collector, and configuration changes carry a visible history.
See it running inside your own boundary.
Talk to a NeuralSeek expert about how it fits your stack, where your data has to live, and the governance your auditors already expect.