Update app.py

app.py

@@ -1,78 +1,51 @@
-# app.py — AURA no HF Spaces (CPU), flan-t5-small, resposta limpa
+# app.py — AURA no HF Spaces (CPU), mt0-small (multilíngue), saída limpa
 
 import gradio as gr
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
-import torch
+from transformers import pipeline
 
-# Modelo leve (cabe no limite gratuito)
-MODEL_ID = "google/mt5-small"
+# Modelo pequeno, instruído, multilíngue e leve (cabe no plano grátis)
+MODEL_ID = "bigscience/mt0-small"
 
-# Carrega modelo/tokenizer (CPU)
-tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
-model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID)
+# Carrega via pipeline (mais estável para chat curto)
+generator = pipeline("text2text-generation", model=MODEL_ID)
 
-# Prompt de sistema curto (o completo está em aura_prompt.json, mas Flan-T5
-# funciona melhor com instruções compactas)
 SYSTEM_PROMPT = (
     "You are AURA, an Integrative Therapeutic Intelligence. "
-    "Your role is to offer emotional support, reflective questions, and short micro-practices "
-    "(e.g., breathing, grounding, brief journaling, or a simple 3-step CBT reframing). "
-    "Never diagnose, label, prescribe, or claim to replace a licensed clinician. "
-    "Adjust your tone to the user's emotional state; be warm, calm, non-judgmental, and concise. "
-    "Prefer 3–6 sentences unless the user asks for more. Offer at most one micro-practice per turn. "
-    "Always reply in the user's language; if the user writes in Portuguese, answer in Portuguese. "
-    "Crisis protocol: if the user mentions suicide, self-harm, violence, or immediate danger, "
-    "stop normal guidance and respond with a safety message encouraging immediate human help. "
-    "Use this safety text in Portuguese if relevant: "
-    "'Sinto muito que você esteja passando por isso. Eu me preocupo com a sua segurança agora. "
-    "Eu não substituo ajuda humana. Se houver risco imediato, ligue para 190/192 (ou o número de emergência da sua região) "
-    "ou procure alguém de confiança imediatamente. Posso te ajudar a encontrar um serviço de apoio perto de você?'."
+    "Speak with empathy, calm and clarity. Ask short reflective questions. "
+    "Offer at most one micro-practice (e.g., breathing for 1 minute) when appropriate. "
+    "Never diagnose or replace a clinician. "
+    "Always reply in the user's language."
 )
 
-
-def generate_reply(message: str) -> str:
-    # Prompt format that MT5 understands
+def chat_fn(message, history):
+    # Prompt simples e compatível com mt0
     prompt = (
-        f"Instruction: Reply as AURA, an integrative therapeutic AI who speaks with empathy, calm and wisdom. "
-        f"You respond kindly to the user's message, offering short, compassionate reflections and gentle guidance. "
-        f"Always reply in the same language used by the user.\n\n"
-        f"User message: {message}\n\n"
+        f"{SYSTEM_PROMPT}\n\n"
+        f"User: {message}\n"
         f"AURA:"
     )
+    out = generator(
+        prompt,
+        max_new_tokens=120,
+        do_sample=True,
+        top_p=0.9,
+        temperature=0.7
+    )[0]["generated_text"].strip()
+
+    # Limpeza leve: mantém só o trecho final após "AURA:"
+    if "AURA:" in out:
+        out = out.split("AURA:")[-1].strip()
+    # Fallback seguro
+    if not out:
+        out = "I’m here with you. Would you like to tell me a little more about how you’re feeling right now?"
+
+    return out
 
-    inputs = tokenizer(prompt, return_tensors="pt", truncation=True)
-    with torch.no_grad():
-        outputs = model.generate(
-            **inputs,
-            max_new_tokens=120,
-            temperature=0.7,
-            do_sample=True,
-            top_p=0.9
-        )
-
-    text = tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
-    # Remove prompt echoes
-    if "AURA:" in text:
-        text = text.split("AURA:")[-1].strip()
-    if "User message:" in text:
-        text = text.split("User message:")[0].strip()
-
-    if not text:
-        text = "I'm here with you. Would you like to tell me more about how you're feeling right now?"
-    return text
-
-
-def chat_fn(message, history):
-    reply = generate_reply(message)
-    return reply
-
-# Interface simples e estável do Gradio
 demo = gr.ChatInterface(
     fn=chat_fn,
     title="AURA — Inteligência Terapêutica Integrativa",
-    description="Prototipo experimental (não substitui terapia)."
+    description="Protótipo experimental (não substitui terapia)."
 )
 
 if __name__ == "__main__":
     demo.launch()
-