Breaking changes with Ollama and Aspire 9

[aaronpowell]

Thanks to the excellent work from @Mrxx99 we have a major overhaul coming to the Ollama integration in the Aspire 9 release, and this is being coupled with a major update to OllamaSharp.

Models as resources

Originally models were just appended to the Ollama resource and you would pass the Ollama resource as a reference, but this resulted in some workarounds on how you'd set (and discover) the default model to use. But in the v9 release we have an OllamaModelResource that is returned and can be passed as a reference, which will then provide connection information to clients on which model to use.

New connection string format

Originally the "connection string" from an Ollama resource was just the HTTP endpoint, but to support the Model as resource feature better the resources will create a "real" connection string of Endpoint=<...>;Model=<...>. The Model part will only be included if you're passing the OllamaModelResource though. This is a breaking change that you may need update if you're not using the OllamaSharp integration.

OllamaSharp 4 + Microsoft.Extensions.AI

OllamaSharp has been updated to a new major version and now supports the interfaces from Microsoft.Extensions.AI. We've bumped our integration to support that and you can register the IOllamaApiClient (native OllamaSharp client), or from M.E.AI IChatClient and IEmbeddingGenerator (depending on your model type).

Using these new interfaces makes code more portable across LLM/SLM options.

API deprecations and removals

With all the refactoring that's gone on some APIs are being deprecated or straight up removed.

Pre-release builds

Pre-release builds are available on nuget.org and via the Azure Artifacts (if you want the really bleeding edge stuff), so if you're using the Ollama integrations, make sure you test in the lead up to the v9 release.

Relevant issues

• Allow specifying selected model when calling WithReference with a Ollama resource #119
• Update to OllamaSharp 4 #164
• Ollama hosting component should expose resource commands for model management #127