How the AI assistant works

This section describes the AI assistant's components, how it answers questions, and how it handles data.

Components

Component	Role
Assistant panel	Part of the Authorization Hub user interface. Sends the user's message and the open policy, and displays the answer.
API gateway	Routes assistant requests to the assistant service with the signed-in user's token.
Assistant service (hub-ai-service)	Builds the request to the model, checks the policy and runs tests, and stores conversations.
Model provider	The language model configured for the deployment. Not included with Authorization Hub.
PostgreSQL	Holds the conversations and preferences in the assistant schema of the Authorization Hub database.

How a question is answered

1. A message is sent from the assistant panel. The open policy file, the cursor position, the selection, and the editor's errors are added by the panel.

2. The request is forwarded by the API gateway, with the user's token, to the assistant service.

3. The message, the conversation so far, and the assistant's instructions are sent to the model provider.

4. The model can request policies to be read, checked, or tested. The project's data is read as the signed-in user, with that user's permissions, and checks and tests are run with the built-in ALFA compiler and decision engine.

5. The answer is streamed back to the panel as it is written.

6. If a change is proposed, the proposed text is checked where possible, and the result is shown below the change in the panel.

Data handling

• Sent to the model provider:
  • The user's message.
  • The open policy file, including unsaved edits, the cursor position, the selection, and the editor's errors.
  • The earlier messages of the conversation.
  • The user's preferences for the project.
  • The policies, staged changes in My workspace, attributes, and test results that are read to prepare the answer.
• Stored: conversations and preferences, in the Authorization Hub database. Conversations are removed after a period without activity. Read the Conversation limits section for details.
• Not logged: messages, answers, policy content, and the content of policy checks. Only the user, the check, and its outcome are logged.
• Not stored by Authorization Hub: the model provider's own processing. Your provider's data retention terms apply.

Conversations and preferences are retained when the assistant service is restarted.

Trust boundary

• The assistant is instructed to stay on authorization topics and to decline unsafe requests. These instructions guide the language model but they are not a security control.
• Messages and answers are not filtered. A user who deliberately tries to make the assistant act outside its scope may succeed.
• The assistant cannot exceed the user's permissions. It reads data as the user, and it cannot save or publish.
• Every answer must be treated as a suggestion. Proposed changes are checked where possible, but a check proves only the cases that were tested.