Keep models, knowledge and workflows replaceable.
Separate model serving, retrieval, evaluation and the application layer. Model weights, training data and application code can carry different licences.
Check the reviewed edition, migration constraints and ongoing operating costs. The options below are starting points for a fit-gap assessment, not a universal ranking.
Design a portable AI architecture
Keep the application separate from model providers. Put retrieval behind a documented interface, retain source documents and evaluation datasets, and test a second model before committing.
Review each licence separately
Open-source application code does not establish the rights to model weights or training data. Verify the exact model licence, usage constraints and deployment requirements.
Test quality and control
Use task-specific evaluations, protect sensitive retrieval data, record model and prompt versions, and budget for inference capacity and ongoing review.
Before you choose
Define the organisation size, peak usage and workflows that the system must support. Test a realistic data sample, permission model and integration before estimating delivery.
Migration and operating considerations
Map source fields, historical records, attachments and business rules. Plan a coexistence period and reconcile outputs before switching the authoritative system.
Assign an owner for upgrades, dependency patches, monitoring, backups and recovery tests. Hosting and support remain real costs.
Security and integration
Confirm standards-based identity where available, least privilege and audit requirements. Test the exact API and edition, handle integration failures explicitly, and protect secrets outside application code.
Need a stack that fits your organisation?
Tell iGears what your team needs to do. We can assess the software, integrations and custom development required.