Code Inline Completion change to v1/chat/completions

[Q0]

What happened?

I am using a dedicated service provider for OpenAI models and use Base host. In a new function (Code Inline Completion), the v1/completions method is called. This method is not available from my service provider. This method is Legacy in the OpenAI documentation, and also it is most often not available on new models.

Is it possible to change the method to v1/chat/completions?

CodeGPT version

2.3.0

[mgls23]

As supporting evidence, the one for OpenAI API, seems correctly set to v1/chat/completions

https://github.com/carlrobertoh/CodeGPT/blob/4ccba567d98bacedf485428ba97ee196f0f12ab4/src/main/java/ee/carlrobert/codegpt/settings/service/openai/OpenAISettingsState.java#L8

This seems like high priority to me as it impairs local LLMs from being used

[mgls23]

@Q0 in the meantime, it looks like you can use OpenAI API query to get around this problem. You can find this in

Pycharm Settings > Tools > CodeGPT

[Q0]

The main functionality of the plugin works correctly due to the use of Path /v1/chat/competions. However, the Code Inline Completion function uses a different method

https://github.com/carlrobertoh/CodeGPT/blob/4ccba567d98bacedf485428ba97ee196f0f12ab4/src/main/java/ee/carlrobert/codegpt/completions/CompletionRequestService.java#L81

which in turn calls the method

https://github.com/carlrobertoh/llm-client/blob/6df58820199596dc0531c67af47fa01289e04c53/src/main/java/ee/carlrobert/llm/client/openai/OpenAIClient.java#L121

and it uses /v1/completions.

As I said earlier, this method is not available from many service providers for OpenAI models.

Another reason is that the Path setting does not affect the Code Inline Completion function

[PhilKes]

I am going to look into it, the legacy /v1/completions used the suffix parameter for the code-completion, and I could not yet get the FIM prompt to work properly with /v1/chat/completions but I will open a PR for this

[carlrobertoh]

This method is Legacy in the OpenAI documentation, and also it is most often not available on new models.

@Q0 The only way to gain access to their gpt-3.5-turbo-instruct model is through the /v1/completions API endpoint. Their docs are misleading; although it states that the entire service is deprecated, their recent deprecation docs recommend it as a replacement for older models, and no words of deprecation.

I'm happy to change it to /v1/chat/completions if someone could tell me how.

[gmontamat]

@carlrobertoh check #1103

This PR adds support for /v1/chat/completions with any chat model without removing the default one.