Fix stop reason for a few providers

[narengogi]

Title:

• Transform Stop Reason from Provider specific enums to OpenAI Response Schema compliant ones

Description:

• I'm using Enums instead of string types to be explicit, it adds a few lines of code which would have been simply overlapping types if we used string types instead, I find this more readable
• Did not make any of the existing response types non nullable

TO-DO

• At some of the places, stop reason is hardcoded to be length/stop, I did not change those as part of this PR.
• Types and Utils can be extracted from inside the providers complete.ts and chatComplete.ts files to make the code more readable

Motivation:

• Portkey Gateway currently does not transform the stop_reason at a lot of places and it fails with strictly typed languages and fails validations.

Related Issues:

• https://github.com/Portkey-AI/gateway/issues/459

Also thanks to @unsync for his initial PR on this. I've extended it to all the providers

[narengogi]

@VisargD for your review

[VisargD]

Hey! This PR is not currently handling stop reason transformation for streaming calls.

[narengogi]

provider	chat	chat streaming	completion	completion streaming
Ai21	:white_check_mark:	N/A	:white_check_mark:	N/A
Anthropic	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
Anyscale	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
Azure OpenAI	:white_check_mark:	N/A	:white_check_mark:	N/A
Cohere	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
DeepInfra	:white_check_mark:	:white_check_mark:	N/A	N/A
Fireworks	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
Google	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
Groq	:white_check_mark:	:white_check_mark:	N/A	N/A
Lingyi	:white_check_mark:	:white_check_mark:	N/A	N/A
Mistral	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
MonsterAPI	:white_check_mark:	N/A	N/A	N/A
Moonshot	:white_check_mark:	N/A	N/A	N/A
Novita	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
Ollama	[ ]	[ ]	[ ]	[ ]
Openrouter	Not possible, they return different stop reason depending on deployed model/provider
Palm	:white_check_mark:	N/A	:white_check_mark:	N/A
Perplexity	:white_check_mark:	:white_check_mark:	N/A	N/A
Predibase	Not possible, they return different stop reason depending on deployed model/provider
Reka	:white_check_mark:	N/A	N/A	N/A
Together	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
Workers	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
Zhipu	:white_check_mark:	:white_check_mark:	N/A	N/A

[narengogi]

@VisargD please review this, I'll have to make changes again if new providers are added

[narengogi]

closing this for now as discussed with @VisargD and @ayush-portkey Stems from decision not to transform any of the json values, and only stick to transforming keys.

Transforming stop reason	Not Transforming
✅ Compatible with OpenAI SDK, even when type validation is applied (zod, pydantic, etc)	Will break with OpenAI SDK if type validation applied
Context loss	✅ No Context loss
Added maintenance	✅ No maintenance

Illustrated example on why doing this would mean context loss: VERTEX_AI provider returns following types with key finishReason

export enum finishReason {
  FINISH_REASON_UNSPECIFIED = 'FINISH_REASON_UNSPECIFIED',
  STOP = 'STOP',
  MAX_TOKENS = 'MAX_TOKENS',
  SAFETY = 'SAFETY',
  RECITATION = 'RECITATION',
  OTHER = 'OTHER',
  BLOCKLIST = 'BLOCKLIST',
  PROHIBITED_CONTENT = 'PROHIBITED_CONTENT',
  SPII = 'SPII',
}

since OPENAI only has 4 finish_reason enums, we map the above keys to one of the 4.

const transformVertexGeminiChatStopReason = (
  stopReason?: VERTEX_GEMINI_GENERATE_CONTENT_FINISH_REASON | string
): OPEN_AI_CHAT_COMPLETION_FINISH_REASON => {
  switch (stopReason) {
    case VERTEX_GEMINI_GENERATE_CONTENT_FINISH_REASON.STOP:
      return OPEN_AI_CHAT_COMPLETION_FINISH_REASON.stop;
    case VERTEX_GEMINI_GENERATE_CONTENT_FINISH_REASON.SAFETY:
    case VERTEX_GEMINI_GENERATE_CONTENT_FINISH_REASON.RECITATION:
    case VERTEX_GEMINI_GENERATE_CONTENT_FINISH_REASON.BLOCKLIST:
    case VERTEX_GEMINI_GENERATE_CONTENT_FINISH_REASON.PROHIBITED_CONTENT:
    case VERTEX_GEMINI_GENERATE_CONTENT_FINISH_REASON.SPII:
      return OPEN_AI_CHAT_COMPLETION_FINISH_REASON.content_filter;
    case VERTEX_GEMINI_GENERATE_CONTENT_FINISH_REASON.MAX_TOKENS:
      return OPEN_AI_CHAT_COMPLETION_FINISH_REASON.length;
    default:
      return OPEN_AI_CHAT_COMPLETION_FINISH_REASON.stop;
  }
};