app.py

import gradio as gr
from pydantic import BaseModel, Field
from typing import List, Dict, Any
from groq import Groq
import time

# Real Groq API call
def call_llm(model: str, api_key: str, question: str) -> str:
    try:
        client = Groq(api_key=api_key)
        completion = client.chat.completions.create(
            model=model,
            messages=[
                {"role": "user", "content": question}
            ],
            temperature=0.7,
            max_tokens=1024
        )
        return completion.choices[0].message.content
    except Exception as e:
        return f"Error calling {model}: {str(e)}"

def reevaluate_answer(model: str, api_key: str, answer: str) -> str:
    try:
        client = Groq(api_key=api_key)
        completion = client.chat.completions.create(
            model=model,
            messages=[
                {"role": "user", "content": f"Please review and improve this answer if needed: {answer}"}
            ],
            temperature=0.5,
            max_tokens=1024
        )
        return completion.choices[0].message.content
    except Exception as e:
        return f"Error re-evaluating with {model}: {str(e)}"

def master_review(model: str, api_key: str, all_answers: List[Dict]) -> str:
    try:
        client = Groq(api_key=api_key)
        answers_text = "\n\n".join([f"Answer {i+1} from {ans['model']}: {ans['answer']}" for i, ans in enumerate(all_answers)])
        completion = client.chat.completions.create(
            model=model,
            messages=[
                {"role": "user", "content": f"Review these answers and select the best one. Only return the number (1, 2, 3, etc.) of the best answer:\n\n{answers_text}"}
            ],
            temperature=0.3,
            max_tokens=50
        )
        return completion.choices[0].message.content.strip()
    except Exception as e:
        return "1"  # Default to first answer if error

models_list = ["qwen/qwen3-32b", "openai/gpt-oss-120b", "openai/gpt-oss-20b", "moonshotai/kimi-k2-instruct", "llama-3.1-8b-instant", "meta-llama/llama-guard-4-12b", "groq/compound", "groq/compound-mini"]

def process_question(api_key, selected_models, question):
    if not api_key or not selected_models:
        return "Please provide API key and select models", "", "", "", "", "", "", "", "", "", "", "", ""
    
    outputs = [""] * 13  # Initialize all outputs
    answers = []
    
    # Process each selected model
    model_idx = 0
    for model in selected_models[:4]:  # Limit to 4 models max
        base_idx = model_idx * 3
        
        # Initial answer
        outputs[base_idx] = f"🤔 Getting answer from {model}..."
        yield tuple(outputs)
        
        initial_ans = call_llm(model, api_key, question)
        outputs[base_idx] = initial_ans
        yield tuple(outputs)
        
        # Re-evaluate
        outputs[base_idx + 1] = f"🔄 Re-evaluating..."
        yield tuple(outputs)
        
        reevaluated = reevaluate_answer(model, api_key, initial_ans)
        outputs[base_idx + 1] = reevaluated
        answers.append({"model": model, "answer": reevaluated})
        yield tuple(outputs)
        
        model_idx += 1
    
    # Master review - only use openai/gpt-oss-120b as master
    master_model = "openai/gpt-oss-120b"
    outputs[3] = f"🏆 Master {master_model} reviewing all answers..."
    yield tuple(outputs)
    
    choice = master_review(master_model, api_key, answers)
    try:
        choice_idx = int(choice) - 1
        if 0 <= choice_idx < len(answers):
            chosen_model = answers[choice_idx]["model"]
            chosen_answer = answers[choice_idx]["answer"]
        else:
            chosen_model = answers[0]["model"]
            chosen_answer = answers[0]["answer"]
    except:
        chosen_model = answers[0]["model"]
        chosen_answer = answers[0]["answer"]
    
    outputs[3] = f"Master chose: {chosen_model}"
    
    # Clear unused master vote boxes
    outputs[6] = ""
    outputs[9] = ""
    
    # Final answer
    outputs[12] = f"🏆 Master LLM ({master_model}) selected:\n\n{chosen_answer}"
    
    yield tuple(outputs)

with gr.Blocks() as demo:
    gr.Markdown("""
    # Multi-LLM Master Evaluator with Real-time Streaming
    Enter your Groq API key, select LLM models, and ask a question. 
    Watch real-time processing as each LLM answers, re-evaluates, and then the master LLM (openai/gpt-oss-120b) selects the best answer.
    """)
    
    with gr.Row():
        api_key = gr.Textbox(label="🔑 Groq API Key", type="password", placeholder="Enter your API key...")
        models = gr.CheckboxGroup(models_list, label="🤖 Select Models", value=["llama-3.1-8b-instant", "openai/gpt-oss-20b"])
    
    question = gr.Textbox(label="❓ Your Question", placeholder="Ask your question...")
    run_btn = gr.Button("🚀 Start Evaluation")
    
    # Model outputs (4 models max, 3 columns each)
    with gr.Row():
        with gr.Column():
            gr.Markdown("### Model 1")
            model1_initial = gr.Textbox(label="Initial Answer", lines=3, interactive=False)
            model1_reevaluated = gr.Textbox(label="Re-evaluated", lines=3, interactive=False)
            model1_vote = gr.Textbox(label="Master LLM Decision", lines=1, interactive=False)
        
        with gr.Column():
            gr.Markdown("### Model 2")
            model2_initial = gr.Textbox(label="Initial Answer", lines=3, interactive=False)
            model2_reevaluated = gr.Textbox(label="Re-evaluated", lines=3, interactive=False)
            model2_vote = gr.Textbox(label="(Master uses Model 1 slot)", lines=1, interactive=False, visible=False)
    
    with gr.Row():
        with gr.Column():
            gr.Markdown("### Model 3")
            model3_initial = gr.Textbox(label="Initial Answer", lines=3, interactive=False)
            model3_reevaluated = gr.Textbox(label="Re-evaluated", lines=3, interactive=False)
            model3_vote = gr.Textbox(label="(Master uses Model 1 slot)", lines=1, interactive=False, visible=False)
        
        with gr.Column():
            gr.Markdown("### Model 4")
            model4_initial = gr.Textbox(label="Initial Answer", lines=3, interactive=False)
            model4_reevaluated = gr.Textbox(label="Re-evaluated", lines=3, interactive=False)
            model4_vote = gr.Textbox(label="(Master uses Model 1 slot)", lines=1, interactive=False, visible=False)
    
    final_answer = gr.Textbox(label="🏆 Final Best Answer", lines=6, interactive=False)
    
    run_btn.click(
        process_question,
        inputs=[api_key, models, question],
        outputs=[model1_initial, model1_reevaluated, model1_vote,
                model2_initial, model2_reevaluated, model2_vote,
                model3_initial, model3_reevaluated, model3_vote,
                model4_initial, model4_reevaluated, model4_vote,
                final_answer]
    )

demo.launch()