Feature Wiki

Information about planned and released features

Tabs

AI Base Infrastructure

1 Summary

Over the past few years, many plugins have appeared that bring AI-supported features to ILIAS. They are valuable groundwork, and several of them already cover interesting, concrete use cases. However, all of these plugins share a common base infrastructure that each one has had to build on its own: managing providers, models, agents, general guidelines and more. This leads to duplicated effort, inconsistent behaviour across installations and higher development and maintenance costs.

ILIAS should therefore offer its own AI base infrastructure. This has two goals:

  • AI features in ILIAS: Components should be able to add AI-supported features with as little effort as possible, based on a shared and consistent foundation.
  • ILIAS for AI: ILIAS should be ready to be connected to AI systems. Through an MCP (Model Context Protocol) interface, external AI clients should be able to work with ILIAS.

This request only proposes the base infrastructure. It does not include any concrete AI use cases. Those can be suggested as separate feature requests built on top of this infrastructure.

In addition to the infrastructure, this request includes a guideline for developing AI features in ILIAS. No feature may depend on whether an installation uses an AI connection. Every feature must be provider-agnostic and must keep working if no suitable agents or models are available.

2 Problem Statement

See above

3 Concept

3.1 Guiding Principles

AI in ILIAS is meant to support users, never to replace existing workflows without an alternative.

  • Supportive, not exclusive: Every AI-supported process must have a non-AI alternative. For example, if tile images for repository objects can be generated by an image generator, it must still be possible to upload an image manually.
  • Provider-agnostic: Features must not depend on a specific provider or model.
  • Graceful absence: Features must work in installations without an AI connection or without suitable agents and models.
  • Transparency: It must always be clear to users what is generated by AI and which actions use AI. For example, form fields that are pre-filled with a suggestion must be marked as AI-generated suggestions. Actions such as generating a tile image must be marked, for example with an icon in the button.

These principles will be written down as a development guideline that is part of this request.

3.2 Provider Management

Administrators can manage AI providers in a provider-agnostic way. Adding a new provider should be as easy as possible, ideally with templates for common providers. Self-hosted models running on an installation's own servers can be connected as well.

3.3 Model Management

For each provider, administrators can configure the models it offers:

  • Activate or deactivate models.
  • Define basic settings such as the context window.
  • Where possible, enter the price per token, so that limits can be enforced.

3.4 Agents and Purposes

An Agent fulfils a specific task in ILIAS, for example writing a summary or generating an image.

A Purpose describes a need in ILIAS, for example "generate tile image". Several Agents can serve the same Purpose. This allows a component to provide its own Agent for an existing Purpose, for example one that is more specialised or that adds data from that component.

In the administration, administrators decide which Agent is used for which Purpose:

  • If only one Agent is available for a Purpose, it is used automatically.
  • If several Agents are available, none is used until an administrator has chosen one.

Agents can be configured with additional instructions and with uploaded files that serve as extra context.

3.5 Tools

ILIAS Activities that are already available for other interfaces are offered as **Tools**. The same Tools are:

  • Available through the MCP server for external AI clients.
  • Assignable to Agents if possible by configuration, so Agents can fetch additional information from ILIAS when they need it.

Permission checks remain the responsibility of each Tool/Activity.

3.6 Context

Everything that is sent to an LLM is wrapped in a Context. The Context is the single point where pseudonymisation takes place. It can contain text as well as media and files, and it is built up in several layers:

  • Agent context: Instructions and files that administrators configure for an Agent.
  • Tool context: Information that the Agent fetches through its assigned Tools.
  • Caller context: Additional data from the component that calls the Agent, for example a file that should be summarised.

This keeps the infrastructure flexible enough to cover as many use cases as possible.

3.7 Pseudonymisation

The goal is that data sent to an LLM contains no directly identifying personal data. To achieve this, the infrastructure includes a pseudonymisation layer:

  • Before a request is sent, specific data such as e-mail addresses or usernames is replaced by placeholders.
  • These placeholders can appear in the LLM's response. The pseudonymisation layer then replaces them with the real values again, inside ILIAS.

The infrastructure comes with a set of standard pseudonymisers. Components can contribute their own pseudonymisers, and administrators can define additional rules per installation, for example based on regular expressions. All of them follow the same mechanism. This ensures that as little personal or otherwise sensitive data as possible is sent to LLMs.

3.8 MCP Server

ILIAS provides an MCP server that makes the existing Tools available to external AI clients.

  • Administrators define per role whether users may create MCP tokens.
  • Users with this permission can create tokens, for example in their personal profile, and administrators can manage them in the user administration.
  • All data passed through the MCP server uses the same pseudonymisation layer.

3.9 Permissions and Limits

The infrastructure does not handle permissions for AI features. Each component that uses an Agent decides who may use the feature, in line with its existing permission model.

Administrators can set usage limits per model or per provider for a configurable period, for example a monthly budget based on the price per token. Once a limit is reached, all features that depend on that model or provider are unavailable until the period ends. These actions are not hidden but shown as disabled, ideally with a short explanation, so users understand why they can't use them at the moment. The non-AI alternative stays available, as the guideline requires.

3.10 Extensibility

The AI infrastructure does not provide any plugin slots and has no special handling for plugins. Contributions are made solely through the ILIAS bootstrapping mechanism of the Component Revision. Any component that is integrated this way can provide its own Agents, Purposes, Tools and pseudonymisers, exactly like core components.

Existing AI plugins can use the infrastructure once they have been migrated to components. In the medium term, this should make developing such features much cheaper, because the shared infrastructure no longer has to be built separately for each plugin.

4 User Interface Modifications

4.1 List of Affected Views

  • Administration > Extending ILIAS: new menu entry for the AI infrastructure, with one tab per area:
    - Providers
    - Models
    - Agents and Purposes
    - Pseudonymisation
    - Limits
    - Log
    - Personal profile: managing personal MCP tokens
    - User administration: managing MCP tokens of users
    - Role settings: permission to create MCP tokens

4.2 User Interface Details

Mockups will be added.

4.3 New User Interface Concepts

None. The feature is built with existing UI components.

4.4 Technical Aspects

The infrastructure will be provided as a pull request with the relevant interfaces, so that it can be discussed directly in the code. The pull request will also include the development guideline and README files, so developers can review the concept and give feedback.

Feedback from providers that already run AI plugins is especially welcome. Their experience from recent years should be taken into account in the design.

The infrastructure fully relies on the ILIAS bootstrapping mechanism of the Component Revision (see 2.10).

5 Non-Functional Implications

5.1 Security

  • Encrypted storage: API keys, provider credentials and MCP tokens are always stored encrypted, never in plain text. MCP tokens are shown only once when they are created.
  • No privilege escalation: The MCP server always acts on behalf of the authenticated user and does not grant any additional permissions. Checking permissions remains the responsibility of the individual Tools, as it already is for other interfaces. The infrastructure only makes sure that each call runs in the context of the right user.
  • Token lifecycle: MCP tokens have an expiry date and can be revoked at any time, both by the user and by administrators. Who may create tokens is controlled per role.
  • Untrusted LLM output: Responses from LLMs are treated as untrusted input. They are never executed without checks and are always sanitised before they are displayed, to reduce the risk of prompt injection.
  • Logging and rate limiting: Requests to LLMs and calls to the MCP server are logged in pseudonymised form. Rate limits and the cost limits per model or provider protect against misuse and unexpected costs.
  • Transport security: Connections to external providers use TLS only.

5.2 Privacy

  • Pseudonymisation: Before any data is sent to an LLM, personal data such as names, logins or e-mail addresses is replaced by placeholders. These placeholders are resolved again in the response, inside ILIAS. Components can add their own pseudonymisers, and administrators can define additional rules per installation.
  • Logging: Every request to an LLM is logged. The log only stores the pseudonymised content of requests and responses, so the component itself never keeps unprotected personal data.
    Personal data stored: The log stores which user made which request at what time, together with the Agent and model that were used. This is needed for traceability, cost limits and investigating misuse.
  • Access: Only administrators with the appropriate permission can see the log.
  • Retention: Log entries are deleted after a period that administrators can configure. When a user account is deleted, the related log entries are deleted or anonymised as well.
  • Data processing by providers: Installations are responsible for choosing suitable providers and making data processing agreements where needed. Self-hosted models on the installation's own servers are fully supported as an alternative.

5.3 Accessibility Implications

The feature only uses existing UI components, so it introduces no new accessibility implications.

6 Process Information

6.1 Involved Authorities

If this request is related to multiple components, please list both authorities for all related components.

6.2 Contact

Person to be contacted in case of questions about the feature or for funding offers:  Schmid, Fabian [fschmid]

6.3 Funding

Funding status and funding parties are listed in the block 'Status of Feature' in the right column of this page.

If you are interested to give funding for this feature, please get into contact with the person mentioned above as 'Contact'.

7 Discussion

8 Implementation

Feature has been implemented by {Please add related profile link of this person}

8.1 Description and Screenshots

{ Description of the final implementation and the changed behaviour in the related components. Please add screenshots to visualise the changes if possible. }

8.2 Test Cases

Test cases completed at {date} by {user}

  • {Test case number linked to Testrail} : {test case title}

8.3 Privacy

Information in privacy.md of component: updated at {date} by {user} | no change required

8.4 Approval

Approved at {date} by {user}.

Last edited: 12. Nov 2026, 17:37, Schmid, Fabian [fschmid]