Feature Overview

Creating RAG artifacts, like a vector database or a fine-tuned embedding model, requires a base embedding model.

This Feature card is for the work to select a default text-embedding model for RHEL AI. The selected model will be used to generate embeddings for text data, fine-tuning the embedding model, enabling various natural language processing tasks such as search, clustering, and similarity analysis.

Goals

• Provide a default text-embedding model for use in RHEL AI
• Ensure the selected model is not from a Chinese organization due to customer restrictions.
• Obtain legal clearance for the selected model from Red Hat Legal.
• Measure improvement when fine-tuning the embedding model (if applicable).

Requirements

• The selected model must be an Apache 2.0 licensed embedding model.
• The selected model must not be from a Chinese organization.
• The selected model must be cleared by Red Hat's legal for use in RHEL AI.
• If fine-tuning is required, the improvement in performance must be measurable.

Background

IBM has not released the Slate embedding models as open-source models. Therefore, we need to identify an Apache 2.0 licensed embedding model that meets our customer's restrictions and legal requirements.

Done

• [ ] The selected model is an Apache 2.0 licensed embedding model.
• [ ] The selected model is not from a Chinese organization.
• [ ] The selected model has been cleared by Red Hat's legal for use in RHEL AI.
• [ ] The improvement in performance when fine-tuning the embedding model has been measured (if applicable).

Questions to Answer

• What are the specific Apache 2.0 licensed embedding models that we can consider?
• Red Hat Legal has cleared all-minilm-l6-v2 and all-minilm-l12-v2 for OpenShift Lightspeed. Could those be an option for RHEL AI? It requires research to validate there are measurable improvements when fine-tuning the embedding model.
• Have all customer restrictions and legal requirements been met for the selected model?
• When fine-tuning the selected model, what is the expected improvement in performance?

Out of Scope

• The development and integration of the selected text-embedding model beyond RHEL AI
• The optimization of the selected model for specific use cases.

Customer Considerations

• Ensure that the selected model meets all customer restrictions and legal requirements.