RAG support in Semantic Kernel

[marcominerva]

Hi!

I have just see this: https://github.com/microsoft/semantic-kernel/releases/tag/vectordata-dotnet-9.0.0-preview.1.24515.1. Will this affect future development of Kernel Memory?

[dluc]

Yes, we're always looking for opportunities to port to SK and other products features we explored in KM, such as flexible schemas, filters, async ingestion pipelines, custom RAG, etc. Once those features are ported over, we might refactor KM to leverage them.

[marcominerva]

Do you think that, eventually, Semantic Kernel and Kernel Memory will become a single product?

[dluc]

Semantic Kernel (SK) and Kernel Memory (KM) will not become a single product. SK is an SDK, while KM is a service. SK serves as an SDK to support AI applications, facilitating the use of similar features across different products, such as AI models, vector stores, and enabling automated processes like Agentic AI. KM, on the other hand, is focused on organizing information in an AI-friendly format, supporting scenarios like Retrieval Augmented Generation (RAG).

Within Microsoft, SK is used across most AI projects as a common foundation for architecture, language, and approach. It will continue to evolve as an SDK, supporting more AI application patterns. Some SK patterns have even been incorporated into Microsoft.Extensions.AI, which itself is a library to extend those capabilities. Although SK could evolve into a service, its current roadmap aims to leverage SK within existing services, like Copilot and Copilot Agents.

KM will continue as an exploration project, focusing on porting its features into other products, such as SK's new vector store interfaces, Microsoft.Extensions.AI, and future products.