import gradio as gr
from huggingface_hub import InferenceClient
from transformers import AutoTokenizer, AutoModelForQuestionAnswering, pipeline
from sentence_transformers import SentenceTransformer, util

# Carregar modelos
model_name = "deepset/roberta-base-squad2"
qa_pipeline = pipeline('question-answering', model=model_name, tokenizer="HF_TOKEN_1")
chat_client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
embed_model = SentenceTransformer('all-MiniLM-L6-v2')


class MultiModelQA:
    def __init__(self, qa_pipeline, chat_client, embed_model):
        self.qa_pipeline = qa_pipeline
        self.chat_client = chat_client
        self.embed_model = embed_model

    def answer_with_qa_model(self, question, context):
        return self.qa_pipeline({'question': question, 'context': context})['answer']

    def answer_with_chat_model(self, question, system_message, max_tokens, temperature, top_p):
        messages = [
            {"role": "system", "content": system_message},
            {"role": "user", "content": question}
        ]
        response = ""
        for msg in self.chat_client.chat_completion(
            messages,
            max_tokens=max_tokens,
            stream=True,
            temperature=temperature,
            top_p=top_p,
        ):
            token = msg.choices[0].delta.content
            response += token
        return response

    def comparar_semanticamente(self, resp1, resp2):
        emb1 = self.embed_model.encode(resp1, convert_to_tensor=True)
        emb2 = self.embed_model.encode(resp2, convert_to_tensor=True)
        similarity = util.cos_sim(emb1, emb2).item()
        return similarity


multiqa = MultiModelQA(qa_pipeline, chat_client, embed_model)


def responder_e_comparar(question, context, system_message, max_tokens, temperature, top_p):
    qa_resp = multiqa.answer_with_qa_model(question, context)
    chat_resp = multiqa.answer_with_chat_model(question, system_message, max_tokens, temperature, top_p)
    similaridade = multiqa.comparar_semanticamente(qa_resp, chat_resp)

    result = f"""### Resposta do modelo QA:
{qa_resp}

### Resposta do modelo Chat:
{chat_resp}

### Similaridade semântica (coseno): {similaridade:.2%}
"""
    return result


# Interface Gradio
demo = gr.Interface(
    fn=responder_e_comparar,
    inputs=[
        gr.Textbox(label="Pergunta"),
        gr.Textbox(label="Contexto"),
        gr.Textbox(value="Você é um assistente útil.", label="Mensagem do sistema"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Máximo de tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperatura"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p"),
    ],
    outputs=gr.Markdown(),
    title="Comparador de Respostas de Modelos",
    description="Compara as respostas de um modelo de QA e um modelo de chat (Zephyr-7B) e calcula a similaridade semântica entre elas."
)

if __name__ == "__main__":
    demo.launch()