[FEATURE REQ] Expand character limit on UserAgentApplicationId

[posaninagendra]

Describe the feature or improvement you are requesting

Today UserAgentApplicationId imposes a strict character limit of 24 characters, this limits the ability for our service to send client sent user agents appended to the service's own user agent and this in turn limits the ability to properly categorize token consumption at Azure OpenAI inferencing logs.

Having no character limit would be great or a big buffer for Azure services will unblock.

Character Limit Validation: https://github.com/openai/openai-dotnet/blob/main/src/Custom/Internal/TelemetryDetails.cs#L41

Additional context

No response

[KrzysztofCwalina]

@JeffreyRichter, do you remember why we have this limit?

[trrwilson]

Just as a reference point -- although the logic is copied to be project-local here, it's mirrored from the established behavior in Azure.Core:
https://github.com/Azure/azure-sdk-for-net/blob/81a4260b877f9739ab1d5b86757a0b38c11ae4ef/sdk/core/Azure.Core/src/TelemetryDetails.cs#L18

24 characters does seem to be excessively constrained to me.

[KrzysztofCwalina]

Yeah, it seems short, but it was picked like that for a reason (which I forgot). I vaguely remember that before we introduced this limit some servers/proxies were truncating the header. Jeff would remember.

[JeffreyRichter]

I don't remember exactly. I think some routers had limits our something. Brian Terlson might know more about why we set this limit. I'm not opposed to raising it.

[KrzysztofCwalina]

@bterlson, would you remember?

[posaninagendra]

@KrzysztofCwalina any update on this? Should we raise the limit?

[jsquire]

@KrzysztofCwalina, @christothes: Would appreciate your thoughts on a final outcome for the decision.

[KrzysztofCwalina]

OK, so I did some digging into the history, current state of various services, etc.
The conclusion is that it's ok to give more flexibility to the client library authors (not users) as to the maximum length of the application ID that the client is willing to accept, as it pretty much depends on the implementation of the service how much the service is able to consume.

In the OpenAI library, we do the length check in an internal TelemetryDetails class. We could relax this check, but we would have to find some reasonable length that would work for both OpenAI and Azure OpenAI. Maybe @achandmsft could follow up with the OpenAI and Azure OpenAI service teams to agree on some common length?