Microsoft.Extensions.AI.* future plans #5498
-
|
Hi, I recently read your blog post about the exciting new libraries. At DevExpress, we are eager to transition our implementation to use the new API members you’ve introduced, replacing our own (you can find details about our AI-powered enhancements in this blog post by Daniel Roth). However, we have a few questions, and your answers will influence our decision:
Cheers, Product Manager @ DevExpress |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
Thanks!
In this regard, the Microsoft.Extensions.AI libraries are on the same plan as the rest of the core libraries produced by the .NET team. They have assets for all in-support versions of the platform. That means the 9.0 versions of the package include assets for .NET 8 and .NET 9 (but not .NET 6, which will be out of support by the time the 9.0 packages GA). I expect the Microsoft.Extensions.AI.Abstractions and Microsoft.Extensions.AI libraries will GA with the same assets they currently have, which is net9.0, net8.0, net462, and netstandard2.0.
At the moment, we're not anticipating any significant changes, but we are currently shipping the libraries as preview because we're reserving the right to react to feedback and fix anything that's considered problematic. At a minimum, I do expect there will be at least minor tweaks to the surface area here and there before it's all done. And we do very much want to hear about both things that work really well for folks and things that are problematic, sooner rather than later, so that we can appropriately react to those and ensure the APIs are great for the long-term. That said, the shape that's currently there is based on almost two years of experience with similar APIs in Semantic Kernel, work done to understand commonalities between all these systems as part of Azure AI Inference APIs, and doing our own implementations of these abstractions on a multitude of systems.
Right now we're focused on abstractions for which there are many different implementations / services that all map to the same concepts. We do not plan to ship any APIs specific to a particular provider's services. We're very much focused on being glue for the .NET ecosystem rather than being a framework meant to surplant others; we simply want everyone to be able to pick the framework / libraries they want to use while still being able to participate in an ecosystem with everyone else. As patterns around agents solidify, it's quite possible we'd look at introducing new abstractions focused on such patterns, of which one possible implementation is based on OpenAI assistants. We are also looking at whether IChatClient might be relevant as well, eg getting an IChatClient to represent a particular thread of conversation with an assistant. For now at least, we'd recommend folks looking to use OpenAI assistants use one of the existing libraries for doing so, such as the OpenAI library, Semantic Kernel's agents framework (which includes an implementation based on OpenAI assistants), etc. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for the prompt response and all the details, Stephen! Your answers have given us the confidence to proceed with using this library, and we plan to migrate our implementation to these APIs in the coming weeks. Once we complete the work, or if we encounter any blockers, I’ll be sure to share our feedback with you. Keep up the great work! Best regards, |
Beta Was this translation helpful? Give feedback.
Thanks!
In this regard, the Microsoft.Extensions.AI libraries are on the same plan as the rest of the core libraries produced by the .NET team. They have assets for all in-support versions of the platform. That means the 9.0 versions of the package include assets for .NET 8 and .NET 9 (but not .NET 6, which will be out of support by the time the 9.0 packages GA). I expect the Microsoft.Extensions.AI.Abstractions and Microsoft.Extensions.AI libraries will GA with the same assets they currently have, which is net9.0, net8.0, net462, and netstandard2.0.