Update app.py

app.py

@@ -1,11 +1,53 @@
 import gradio as gr
 from huggingface_hub import InferenceClient
 
-"""
-For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-"""
-client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
+# Defina os nomes dos modelos na Hugging Face que quer usar
+MODEL_1 = "HuggingFaceH4/zephyr-7b-beta"       # Modelo gerador 1
+MODEL_2 = "facebook/blenderbot-400M-distill"   # Modelo gerador 2 (exemplo)
+JUDGE_MODEL = "google/flan-t5-large"            # Modelo julgador (exemplo)
 
+client_1 = InferenceClient(MODEL_1)
+client_2 = InferenceClient(MODEL_2)
+client_judge = InferenceClient(JUDGE_MODEL)
+
+def call_model(client, messages, max_tokens, temperature, top_p):
+    # Chama o modelo de chat_completion e monta resposta completa (não stream)
+    response = ""
+    for message in client.chat_completion(
+        messages,
+        max_tokens=max_tokens,
+        stream=True,
+        temperature=temperature,
+        top_p=top_p,
+    ):
+        token = message.choices[0].delta.content
+        if token:
+            response += token
+    return response
+
+def judge_response(client, prompt, resp1, resp2):
+    # Monta prompt para o julgador comparar respostas
+    judge_prompt = (
+        f"Usuário perguntou: {prompt}\n"
+        f"Resposta 1: {resp1}\n"
+        f"Resposta 2: {resp2}\n"
+        f"Qual das respostas é melhor? Responda apenas com 1 ou 2."
+    )
+    inputs = [{"role": "user", "content": judge_prompt}]
+
+    response = ""
+    for message in client.chat_completion(
+        inputs,
+        max_tokens=20,
+        stream=True,
+        temperature=0.0,  # julgador mais determinístico
+        top_p=1.0,
+    ):
+        token = message.choices[0].delta.content
+        if token:
+            response += token
+    # Retorna o número da resposta escolhida, ou "1" como default
+    return response.strip()
 
 def respond(
     message,
@@ -15,34 +57,31 @@ def respond(
     temperature,
     top_p,
 ):
+    # Monta o contexto de conversa para os geradores
     messages = [{"role": "system", "content": system_message}]
-
     for val in history:
         if val[0]:
             messages.append({"role": "user", "content": val[0]})
         if val[1]:
             messages.append({"role": "assistant", "content": val[1]})
-
     messages.append({"role": "user", "content": message})
 
-    response = ""
+    # Chama os dois modelos geradores
+    response1 = call_model(client_1, messages, max_tokens, temperature, top_p)
+    response2 = call_model(client_2, messages, max_tokens, temperature, top_p)
 
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        token = message.choices[0].delta.content
+    # Chama o julgador para decidir qual resposta é melhor
+    choice = judge_response(client_judge, message, response1, response2)
+
+    # Escolhe a resposta com base no julgamento
+    if choice == "2":
+        final_response = response2
+    else:
+        final_response = response1
 
-        response += token
-        yield response
+    yield final_response
 
 
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
 demo = gr.ChatInterface(
     respond,
     additional_inputs=[
@@ -59,6 +98,5 @@ demo = gr.ChatInterface(
     ],
 )
 
-
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()