Skip to Content
AdminModel Providers

Model Providers

Model providers are the AI services used by Nairo to generate analysis, summaries, comparisons, extractions and other AI-assisted outputs.

Model availability can differ by workspace, product surface and configuration.

Today, Nairo primarily uses Gemini-family models for generative AI features. In backend deployments, Gemini can be configured through Vertex AI or Gemini API infrastructure depending on the deployment setup.

How models are used in Nairo

AI is used across several Nairo surfaces.

SurfaceTypical model experience
AssistantUsers can choose and change the model during a conversation where model selection is available
ProjectsThe Project Assistant works over uploaded Project materials using backend-managed model execution
ExpertsUsers may select a model in the Expert workspace where available
ActionsUsers may select a model when running an Action where available
Insight TableColumn execution uses backend-managed model execution

The exact model options available to a user depend on the workspace configuration and the product surface being used.

Model selection

Where model selection is available, users can choose from the models exposed in that surface.

In the Assistant, users can change the selected model during a conversation. Later messages or retries can use a different selected model where the model selector is available.

Other surfaces may expose model selection differently.

For example:

  • Assistant supports model selection where enabled
  • Projects use the Project Assistant over uploaded materials
  • Experts may expose model selection in the Expert workspace
  • Actions may expose model selection for an Action run
  • Insight Table uses configured backend execution for column runs

Not every surface exposes the same model options.

Execution paths today

Different product surfaces may use different model execution paths.

Backend-managed execution

The following surfaces use backend-managed model execution today:

  • Assistant
  • Projects
  • Insight Table

In these surfaces, the frontend sends the request to the Nairo backend. The backend manages model execution, streaming responses and related processing for that surface.

Projects and Assistant both use backend-managed model execution, but they may use different assistant services behind the scenes.

Client-side execution in current Experts and Actions

Current Experts and Actions flows may use client-side Gemini calls where configured.

This means that, in the current implementation, some Expert and Action interactions can be executed from the browser against the configured model provider rather than through the same backend-managed assistant pipeline used by Assistant, Projects and Insight Table.

This is an implementation detail that may change as Experts and Actions move to backend-managed execution.

Customers evaluating data handling for Experts and Actions should confirm the current execution path for their workspace with Nairo before using sensitive materials.

Persistence and model metadata

Assistant model selection behaves at message execution level.

Where supported, users can change the model during a conversation. The selected model can affect later messages or retries.

The user’s selected model preference may be stored client-side for convenience. Executed model information may also be associated with generated messages where supported by the backend.

There is no customer-facing workspace-wide model policy object in the admin UI today.

What administrators can configure today

Model configuration is currently limited in the customer-facing admin UI.

ConfigurationCurrent status
Change Assistant model during a conversationAvailable where model selection is enabled
Choose Expert model during a sessionAvailable where supported
Choose Action model for a runAvailable where supported
Configure workspace-wide model policy in admin UINot available today
Enforce different model policies by user role in admin UINot available today
Configure self-serve BYOK or BYOM in admin UINot available today

If a customer requires a specific model setup for a pilot, this should be confirmed with Nairo before rollout.

BYOK and BYOM

Nairo does not currently provide a self-serve customer admin UI for bring-your-own-key or bring-your-own-model configuration.

Model provider configuration is handled at deployment or implementation level, not through a general customer-facing admin page today.

Customers with specific provider, region, procurement or data classification requirements should discuss those requirements with Nairo before using sensitive materials.

Governance considerations

Model choice does not remove the need for human review.

When evaluating model providers for a workspace, teams should consider:

  • Which product surfaces will use AI
  • Which documents and data types will be processed
  • Which surfaces use backend-managed execution
  • Whether any current flows use client-side model calls
  • Whether internet search or external tools are enabled
  • Which models are acceptable for the organisation’s data classification
  • What human review process applies before outputs are used
  • What contractual and subprocessors terms apply to the workspace

AI-generated outputs should be reviewed before they are used in business-critical workflows.

What Model Providers is not

Model Providers is not currently a self-serve admin policy engine.

It does not provide customer administrators with full workspace-wide model governance in the application UI today.

It does not provide self-serve BYOK or BYOM configuration today.

It does not replace legal, security, procurement or compliance review of AI providers.

It does not mean all product surfaces use the same model path or expose the same model controls.

Getting started

For pilot planning, confirm which models are available for the workspace and which surfaces expose model selection.

If the organisation has specific model requirements, data classification rules or provider restrictions, confirm those before users begin working with sensitive materials.

Also confirm the current execution path for the surfaces the pilot will use, especially if the pilot includes Experts or Actions.