ADR 001: AI Model Ownership Structure

Status

Accepted

Context

FeedForward allows the use of AI models for providing feedback on student assignments. There are two potential sources of API keys for these models:

1. Institution-wide API keys (managed by administrators)
2. Personal API keys (provided by individual instructors)

We need a database schema that supports both cases while ensuring proper separation of concerns.

Decision

We have decided to enhance the ai_models table with two additional fields:

1. owner_type: Indicates whether the model is system-wide ('system') or instructor-specific ('instructor')
2. owner_id: For instructor-specific models, stores the instructor's email

This allows us to: - Clearly separate system models from instructor models - Enable instructors to use their own API keys - Allow administrators to provide institution-wide models - Support filtering models based on access rights - Maintain clean separation of concerns between admin and instructor roles

Consequences

Positive

• Instructors can use their own API keys before institutional funding is secured
• System can mix and match various models from different sources
• Clear ownership model prevents unauthorized access to API keys
• Flexible structure that works for both small and large deployments
• Simple query pattern to find available models for a specific instructor

Negative

• Slightly more complex queries when retrieving available models
• Additional migration step for existing databases
• Need to implement ownership checks in UI and API endpoints

Implementation

The implementation includes: 1. Adding the new fields to the ai_models table schema 2. Creating a migration script to update existing databases 3. Updating model retrieval logic to filter based on ownership 4. Adding UI elements for instructors to manage their models