Classify the task
What needs doing, which data class, what quality is required?
Assign models by task and test them before use.
Models differ in size, operating location, cost and response time. Model Routing means assigning tasks to a model according to defined criteria.
In brief: Model Routing is the rule by which Swiss AI Factory distributes tasks across models. The application submits a task, and Factory chooses the model – by capability, data protection, speed, cost and availability.
Principle
Four steps per task.
Identify models
Which models meet the functional and data protection requirements?
Select
By speed, cost and availability – with a defined fallback model.
Assess
Result, cost and response time feed back into quality checks.
Examples
Matching task and model.
We name specific models and versions in the project.
| Task | Requirement | Typical assignment |
|---|---|---|
| Classify an email | fast, low-cost, high volume | small local model |
| Review a contract | precise, sensitive | larger local model |
| Transcribe a meeting | Swiss German, internal | specialised model, private cloud |
| Market research | public data, broad | external model |
| Write a report | quality, non-critical data | external or large local model |
Model customisation
When knowledge access and clear instructions are not enough.
RAG means answers from your own documents, with sources. If that is not enough, a model can be specifically adapted. Fine-tuning trains a model on examples for a particular task. With distillation, a smaller model takes on selected capabilities of a larger one. We assess effort and benefit separately.
Further reading
Related topics.
Frequently asked questions
Frequently asked questions about Model Routing
Does model assignment tie us to Swiss AI?
Dependencies on platform, interfaces and contracts must be checked separately. Separating the application can make switching models easier.
What does switching models involve?
The switch is configured in Studio. Before the changeover, quality checks, testing and possible adjustments are needed.
Do you show internal details?
Publicly, we show the principle, layers and data flow at a high level. Internal topology, routing rules, security rules and version states stay internal.
Is it already in production?
We state the maturity of each building block openly. Phrases such as “is designed to” describe the architecture, not a client deployment. Studio, Agents and parts of Factory run in our own development and operating environment and are continuously extended. Further operating models, multiple nodes and private and local installations will follow step by step.
Next step
Clarify the model needs of your tasks.
We propose a model strategy with data classes and fallback rules.