LLM exposure levels

An LLM exposure level controls how much detail RebelCore Agent may return from a dataset. Choose the level while building a module, based on the sensitivity of the data and how the resulting answers will be used.

The available levels are Full, Limited, and Advisory only.

Compare the levels

Level	Suitable output	Does not permit
Full	Exact values, records, aggregates, summaries, and advice.	Fabricated or unsupported values.
Limited	Trends, categories, aggregates, summaries, data quality observations, and qualitative profiles.	Raw rows or enough record-level detail to reconstruct a source record.
Advisory only	Qualitative guidance, classifications, risks, recommendations, and next steps.	Raw records, exact dataset values, counts, percentages, averages, ranges, or other dataset-derived statistics.

Full

Use Full when authorised users may receive the available record-level detail. For example, the Agent may report exact salary fields, dates, or identifiers when the question and dataset support them.

Full exposure does not bypass role permissions, project access, personal-information masking, or audit controls.

Limited

Use Limited when users need analysis but should not receive source records. The Agent can discuss patterns and safe summaries while avoiding row-level disclosure.

If a user asks for an individual record, the answer should provide an allowed summary or explain that the current exposure level does not permit the requested detail.

Advisory only

Use Advisory only for the most sensitive datasets. The Agent uses the selected data to provide business guidance without returning the underlying values or statistics.

For example, it may describe a case as requiring attention and recommend a review, without quoting the score or amount behind that recommendation.

Choose an exposure level

During module creation:

Open the module builder from Datasets.
Find the LLM exposure setting.
Select Full, Limited, or Advisory only.
Review the choice before building the module.

Limited is the default. To change the exposure level later, rebuild the module with the intended setting.

Exposure badges appear on:

Dataset modules.
Project cards and project forms.
Dataset choices in Agent.
Agent responses.
Administrative audit views.

Projects with multiple modules

When a project contains modules with different exposure levels, the most restrictive level wins:

Selected modules	Effective level
Full only	Full
Full and Limited	Limited
Any selection containing Advisory only	Advisory only

The project form shows the effective level before creation. Agent also shows the effective level for the selected datasets.

Understand response badges

Each Agent response displays the exposure level applied to that turn. Use the badge to interpret the answer:

A Full answer can contain exact authorised values.
A Limited answer should avoid source records.
An Advisory answer should focus on qualitative guidance.

Selecting another dataset can make the effective level more restrictive. It cannot make an already selected, stricter dataset less restrictive.

Exposure and personal-information masking

Exposure and masking solve different problems:

Exposure controls the level of dataset detail the Agent may return.
Mask to AI controls whether personal identifiers are replaced before AI processing.
Mask on interface controls whether authorised personal information is shown in the chat.

A Full-exposure dataset can still use Mask to AI. An Advisory-only dataset remains Advisory-only even when the interface is unmasked.

Choosing the right level

Before selecting a level, consider:

Whether users are authorised to see individual records.
Whether exact values are needed to complete the task.
Whether summaries or recommendations provide enough value.
Your organisation’s contractual, regulatory, and information-handling requirements.

When uncertain, start with the more restrictive option and confirm the intended use with your data owner or compliance team.