What's the difference between AzureChatClient and AzureOpenAIChatCompletionService?

[grafanaKibana]

What’s the difference between AzureChatClient and AzureOpenAIChatCompletionService?

I replaced ChatCompletionService with AzureChatClient, but it didn’t seem to have any effect in my app.

To clarify, I’m not calling either of them directly, I’m using plugins created from the Prompt Templates.

[jluocsa]

TL;DR — they register different abstractions over the same Azure OpenAI deployment, so for plugin / prompt-template execution the observable behavior is usually the same.

Helper	Registers	Source
AddAzureOpenAIChatCompletion(...)	IChatCompletionService (and ITextGenerationService) — the SK-native abstraction. Concrete type: AzureOpenAIChatCompletionService.	AzureOpenAIServiceCollectionExtensions.cs
AddAzureOpenAIChatClient(...) (experimental — SKEXP0010)	Microsoft.Extensions.AI.IChatClient — the cross-cutting Microsoft.Extensions.AI abstraction, wrapped in a pipeline that already includes UseKernelFunctionInvocation, UseOpenTelemetry, and (optionally) UseLogging.	AzureOpenAIServiceCollectionExtensions.DependencyInjection.cs

Both ultimately hit the same Azure OpenAI deployment via the same Azure.AI.OpenAI client. That is why your prompt-template plugins behaved the same after the swap — KernelFunctionFromPrompt resolves an AI service through IAIServiceSelector, which accepts IChatCompletionService, ITextGenerationService, or IChatClient; whichever you registered, it ended up calling the same deployment with the same execution settings.

Where the two actually differ

• AddAzureOpenAIChatClient builds a Microsoft.Extensions.AI pipeline you can extend with .AsBuilder().Use(...) for custom middleware, and ships with OpenTelemetry instrumentation out of the box (openTelemetrySourceName / openTelemetryConfig parameters).
• AddAzureOpenAIChatCompletion is the established SK 1.x API — no [Experimental] warning, broader sample coverage, and it also registers ITextGenerationService for older prompt configurations.

What to pick

• You want Microsoft.Extensions.AI middleware, the built-in OTel pipeline, or to share an IChatClient with non-SK code (Aspire DevUI, Microsoft.Agents.AI, custom middleware) → AddAzureOpenAIChatClient and accept the SKEXP0010 warning.
• You just want SK plugins / prompt templates / things that take IChatCompletionService to work as documented in most samples → AddAzureOpenAIChatCompletion.

Avoid registering both with the same serviceId on the same kernel — the IAIServiceSelector will pick one and the other is dead weight. If you want to confirm which one is actually serving your prompts, register a distinct serviceId on each and inspect what your kernel function resolved, or watch the logger source names (Microsoft.SemanticKernel.Connectors.OpenAI vs Microsoft.Extensions.AI.OpenTelemetryChatClient).