app.py

import os
import gradio as gr
from openai import OpenAI

title = None  # "ServiceNow-AI Chat" # modelConfig.get('MODE_DISPLAY_NAME')
description = None

model_config = {
    "MODEL_NAME": os.environ.get("MODEL_NAME"),
    "MODE_DISPLAY_NAME": os.environ.get("MODE_DISPLAY_NAME"),
    "MODEL_HF_URL": os.environ.get("MODEL_HF_URL"),
    "VLLM_API_URL": os.environ.get("VLLM_API_URL"),
    "AUTH_TOKEN": os.environ.get("AUTH_TOKEN")
}

# Initialize the OpenAI client with the vLLM API URL and token
client = OpenAI(
    api_key=model_config.get('AUTH_TOKEN'),
    base_url=model_config.get('VLLM_API_URL')
)


def chat_fn(message, history):
    # Remove any assistant messages with metadata from history
    print(f"Original History: {history}")
    history = [item for item in history if
               not (isinstance(item, dict) and
                    item.get("role") == "assistant" and
                    isinstance(item.get("metadata"), dict) and
                    item.get("metadata", {}).get("title") is not None)]
    print(f"Updated History: {history}")

    messages = history + [{"role": "user", "content": message}]
    print(f"Messages: {messages}")

    # Create the streaming response
    stream = client.chat.completions.create(
        model=model_config.get('MODEL_NAME'),
        messages=messages,
        temperature=0.8,
        stream=True
    )

    history.append(gr.ChatMessage(
        role="assistant",
        content="Thinking...",
        metadata={"title": "🧠 Thought"}
    ))

    output = ""
    completion_started = False
    for chunk in stream:
        # Extract the new content from the delta field
        content = getattr(chunk.choices[0].delta, "content", "")
        output += content

        parts = output.split("[BEGIN FINAL RESPONSE]")

        if len(parts) > 1:
            if parts[1].endswith("[END FINAL RESPONSE]"):
                parts[1] = parts[1].replace("[END FINAL RESPONSE]", "")
            if parts[1].endswith("[END FINAL RESPONSE]\n<|end|>"):
                parts[1] = parts[1].replace("[END FINAL RESPONSE]\n<|end|>", "")

        history[-1 if not completion_started else -2] = gr.ChatMessage(
            role="assistant",
            content=parts[0],
            metadata={"title": "🧠 Thought"}
        )
        if completion_started:
            history[-1] = gr.ChatMessage(
                role="assistant",
                content=parts[1]
            )
        elif len(parts) > 1 and not completion_started:
            completion_started = True
            history.append(gr.ChatMessage(
                role="assistant",
                content=parts[1]
            ))

        # only yield the most recent assistant messages
        messages_to_yield = history[-1:] if not completion_started else history[-2:]
        yield messages_to_yield


# Add the model display name and Hugging Face URL to the description
# description = f"### Model: [{MODE_DISPLAY_NAME}]({MODEL_HF_URL})"

print(f"Running model {model_config.get('MODE_DISPLAY_NAME')} ({model_config.get('MODEL_NAME')})")

gr.ChatInterface(
    chat_fn,
    title=title,
    description=description,
    theme=gr.themes.Default(primary_hue="green"),
    type="messages",
).launch()