Composition of Experts

The SambaNova Composition of Experts (CoE) model architecture combines the broad capabilities and accuracy of the world’s largest models with the performance of much smaller models, model ownership, and the ability to apply role based access controls.

A CoE model is an ensemble of many models that work in conjunction with each other. There can be any number of models in the ensemble to meet the needs of any organization. The composition can consist of any combination of base, pre-trained, and fine-tuned models.

Access Control

With a model dedicated to act as a router, the SambaNova CoE enables role based access controls, so customers can maintain the RBAC they use to ensure that only those in their organization with permission to use a model are able to do so.

Fine Tuning and RAG

The best AI uses the most accurate models. SambaNova customers can easily fine-tune their models with private data for the highest accuracy. They can also use RAG to ensure that their models always have the latest, most accurate information. By using both fine tuning and RAG, SambaNova customers can be assured that they always have the most accurate models.

Compliance

Closed, cloud-based models do not provide any insight into model training data or weights, making explainability with those models impossible. SambaNova customers get the latest open source models, which can be securely trained on their data, and they then own. This gives every customer insight into the models, so they can better meet compliance requirements.