Refactor Proposal: Migrate to Official OpenAI SDK with Extensible Model Wrapper

[pmoharana-cmd]

Summary

This issue proposes a refactor of our existing VLLMModel implementation in model_logic.py, which currently wraps an OpenAI-compatible chat completion endpoint using raw HTTPX requests. The goal is to:

• Leverage the official OpenAI Python SDK for native support of OpenAI-hosted models.
• Retain BaseModel as an abstract interface and use backend-specific subclasses (OpenAIModel, VLLMModel, etc.) to support multiple providers.
• Improve maintainability, compatibility with newer OpenAI features, and allow smoother integration of other providers or local models in the future.

---

Motivation

• Reduced Maintenance Burden: By using the official SDK, we avoid having to manually replicate request formatting, error handling, and streaming logic.
• Better Compatibility: New features (e.g., assistants, tool use, parallel function calls) are only available through the official SDK.
• Modular Architecture: Separating provider logic into individual subclasses of BaseModel enables easier extension and testing of alternatives (e.g., local VLLM, Ollama, LM Studio).
• Standardized Behavior: Less likelihood of subtle bugs from inconsistent behavior between our wrapper and the official API.

---

Proposed Structure

# model_logic.py
class BaseModel(ABC):
    @abstractmethod
    async def call_stream(self, messages: List[Dict[str, str]], **kwargs) -> AsyncGenerator[str, None]:
        ...
    @abstractmethod
    async def close_client(self): ...

class OpenAIModel(BaseModel):
    def __init__(self, model_name: str, api_key: str, ...): ...
    async def call_stream(...):  # use openai.AsyncOpenAI or openai.ChatCompletion.acreate(...)
    async def close_client(): ...

class VLLMModel(BaseModel):
    # existing behavior preserved for local or remote deployments
    # refactored to be more modular if needed