Explicitly raise ContextWindowExceptions

[krrishdholakia]

Your feature request in one line
Explicitly raise ContextWindowExceptions

Describe the solution you'd like
Raise an exception to specifically call out when a context window exceeded exception occurs. OpenAI can raise invalidrequest errors for several reasons - but the fallback strategy specifically for context window exceptions is different

[krrishdholakia]

cc: @yujonglee feel free to add any additional details

[yujonglee]

In the OpenAI Docs, it said An InvalidRequestError indicates that your request was malformed or missing some required parameters, such as a token or an input. This could be due to a typo, a formatting error, or a logic error in your code.

And it can be context_length_exceeded too.

Like @krrishdholakia mentioned, one might want to retry with longer context window model only if InvalidRequestError is raised due to context_length_exceeded, and just raise exception for others(Since it is really invalid request).

It can be implemented as subclass of openai.error.InvalidRequestError.

[krrishdholakia]

Investigating this more - not all providers return a detailed enough error message (e.g. baseten returns a 500 error). We can implement this on a case-by-case basis (i.e. for those that do return a specific max tokens / context window exceeded message) we can return this exception (this should be the 90% case tbh).

[krrishdholakia]

added - https://docs.litellm.ai/docs/exception_mapping.

[krrishdholakia]

Should be available across all models from v0.1.504 onwards