Update app.py

app.py

@@ -20,13 +20,10 @@ if os.environ.get("GEMINI_API_KEY"):
     genai.configure(api_key=os.environ.get("GEMINI_API_KEY"))
 
 # --- HELPER ---
-def encode_image(image_source):
+def encode_image(image_path):
     try:
-        # Se for string (caminho), abre. Se for objeto imagem, ignora.
-        if isinstance(image_source, str):
-            with open(image_source, "rb") as image_file:
-                return base64.b64encode(image_file.read()).decode('utf-8')
-        return None
+        with open(image_path, "rb") as image_file:
+            return base64.b64encode(image_file.read()).decode('utf-8')
     except Exception:
         return None
 
@@ -36,11 +33,8 @@ def encode_image(image_source):
 def run_local_h200(messages):
     # Qwen não suporta imagens
     for m in messages:
-        content = m['content']
-        if isinstance(content, list): # Verifica se tem imagem na lista
-            for item in content:
-                if isinstance(item, dict) and (item.get('type') == 'image' or item.get('type') == 'image_url'):
-                    return "⚠️ Qwen H200 não suporta imagens. Use Gemini ou Pixtral."
+        if isinstance(m['content'], list):
+             return "⚠️ Qwen H200 não suporta imagens. Use Gemini ou Pixtral."
 
     global local_model, local_tokenizer
     if local_model is None:
@@ -50,42 +44,21 @@ def run_local_h200(messages):
             LOCAL_MODEL_ID, torch_dtype=torch.bfloat16, device_map="cuda"
         )
     
-    # Simplifica input para texto
-    text_only_msgs = []
-    for m in messages:
-        content = m['content']
-        text = ""
-        if isinstance(content, str): text = content
-        elif isinstance(content, list):
-            for item in content:
-                if item.get('type') == 'text': text += item['text'] + " "
-        text_only_msgs.append({"role": m['role'], "content": text.strip()})
-
-    text = local_tokenizer.apply_chat_template(text_only_msgs, tokenize=False, add_generation_prompt=True)
+    text = local_tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
     inputs = local_tokenizer([text], return_tensors="pt").to(local_model.device)
     outputs = local_model.generate(**inputs, max_new_tokens=2048, temperature=0.6, do_sample=True)
     return local_tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
 
 def run_groq(messages, model_id):
-    # Groq não suporta imagens
     for m in messages:
-        content = m['content']
-        if isinstance(content, list):
-             for item in content:
-                if isinstance(item, dict) and (item.get('type') == 'image' or item.get('type') == 'image_url'):
-                    return "⚠️ Groq não suporta imagens. Use Gemini ou Pixtral."
+        if isinstance(m['content'], list):
+             return "⚠️ Groq não suporta imagens. Use Gemini ou Pixtral."
 
     if not groq_client: return "❌ Erro: GROQ_API_KEY ausente."
     
     clean_msgs = []
     for m in messages:
-        content = m['content']
-        text = ""
-        if isinstance(content, str): text = content
-        elif isinstance(content, list):
-            for item in content:
-                if item.get('type') == 'text': text += item['text']
-        clean_msgs.append({"role": m['role'], "content": text})
+        clean_msgs.append({"role": m['role'], "content": m['content']})
 
     try:
         completion = groq_client.chat.completions.create(
@@ -97,7 +70,6 @@ def run_groq(messages, model_id):
 def run_mistral(messages, model_id):
     if not mistral_client: return "❌ Erro: MISTRAL_API_KEY ausente."
 
-    # Mistral espera formato específico para imagens
     formatted_msgs = []
     for m in messages:
         new_content = []
@@ -109,14 +81,14 @@ def run_mistral(messages, model_id):
             for item in content:
                 if item.get('type') == 'text':
                     new_content.append({"type": "text", "text": item['text']})
-                elif item.get('type') == 'image' or item.get('type') == 'image_url':
-                    # Gradio 5 manda 'image' com caminho, ou 'image_url' com url
-                    path = item.get('image') or item.get('image_url')
-                    if isinstance(path, dict): path = path.get('url') # Caso venha aninhado
-                    
-                    if path and os.path.exists(path):
-                        b64 = encode_image(path)
+                elif item.get('type') == 'image_url':
+                    url = item['image_url']['url']
+                    # Gradio 4 local path handling
+                    if not url.startswith("data:") and os.path.exists(url):
+                        b64 = encode_image(url)
                         new_content.append({"type": "image_url", "image_url": f"data:image/jpeg;base64,{b64}"})
+                    else:
+                        new_content.append({"type": "image_url", "image_url": url})
         
         formatted_msgs.append({"role": m['role'], "content": new_content})
 
@@ -131,7 +103,6 @@ def run_gemini(messages, model_id):
         model = genai.GenerativeModel(model_id)
         chat_history = []
         
-        # Converte histórico
         for m in messages[:-1]:
             role = "user" if m['role'] == "user" else "model"
             parts = []
@@ -140,12 +111,11 @@ def run_gemini(messages, model_id):
             elif isinstance(content, list):
                 for item in content:
                     if item.get('type') == 'text': parts.append(item['text'])
-                    elif item.get('type') == 'image': 
-                        path = item.get('image')
-                        if path: parts.append(Image.open(path))
+                    elif item.get('type') == 'image_url':
+                        path = item['image_url']['url']
+                        if os.path.exists(path): parts.append(Image.open(path))
             chat_history.append({"role": role, "parts": parts})
 
-        # Mensagem atual
         last_msg = messages[-1]
         current_parts = []
         content = last_msg['content']
@@ -153,39 +123,49 @@ def run_gemini(messages, model_id):
         elif isinstance(content, list):
             for item in content:
                 if item.get('type') == 'text': current_parts.append(item['text'])
-                elif item.get('type') == 'image': 
-                    path = item.get('image')
-                    if path: current_parts.append(Image.open(path))
+                elif item.get('type') == 'image_url':
+                    path = item['image_url']['url']
+                    if os.path.exists(path): current_parts.append(Image.open(path))
 
         chat = model.start_chat(history=chat_history)
         response = chat.send_message(current_parts)
         return response.text
     except Exception as e: return f"❌ Gemini Error ({model_id}): {e}"
 
-# --- ROTEADOR (FIXED FOR GRADIO 5) ---
+# --- ROTEADOR (COMPATÍVEL GRADIO 4) ---
 def router(message, history, model_selector):
-    # No Gradio 5 com type="messages", history já vem formatado como lista de dicts!
-    # message é um dict {'text': '...', 'files': [...]}
+    # Em Gradio 4, history é [[user, bot], [user, bot]]
+    # message é {'text': '...', 'files': ['path']}
     
-    # 1. Prepara a mensagem atual no formato OpenAI-like
-    current_msg = {"role": "user", "content": []}
+    formatted_history = []
     
-    # Texto
-    if message.get("text"):
-        current_msg["content"].append({"type": "text", "text": message["text"]})
-    
-    # Imagens (Arquivos)
-    for file_path in message.get("files", []):
-        # Gradio 5 passa o caminho local do arquivo
-        current_msg["content"].append({"type": "image", "image": file_path})
+    # 1. Processa o histórico antigo
+    for user_turn, bot_turn in history:
+        # User turn (pode ser string ou tupla/lista com imagem no Gradio 4 antigo, mas vamos tratar como string safe)
+        u_text = str(user_turn) 
+        # Tenta limpar se vier com sujeira de arquivo
+        if isinstance(user_turn, tuple): u_text = user_turn[0] 
         
-    # Se não tiver imagem, simplifica (opcional, mas bom pra debug)
-    if not message.get("files") and message.get("text"):
-        # Se for só texto, mantém estrutura de lista pra compatibilidade com funções novas
-        pass 
+        formatted_history.append({"role": "user", "content": u_text})
+        if bot_turn:
+            formatted_history.append({"role": "assistant", "content": str(bot_turn)})
+
+    # 2. Processa a mensagem ATUAL
+    current_content = []
+    text_input = message.get("text", "")
+    files = message.get("files", [])
+
+    if text_input:
+        current_content.append({"type": "text", "text": text_input})
+    
+    for file_path in files:
+        current_content.append({"type": "image_url", "image_url": {"url": file_path}})
 
-    # 2. Junta tudo
-    full_history = history + [current_msg]
+    # Adiciona ao histórico final
+    if not files:
+        formatted_history.append({"role": "user", "content": text_input})
+    else:
+        formatted_history.append({"role": "user", "content": current_content})
 
     # 3. Roteamento
     if "Gemini" in model_selector:
@@ -194,7 +174,7 @@ def router(message, history, model_selector):
         elif "2.5 Pro" in model_selector: tid = "gemini-2.5-pro"
         elif "2.5 Flash" in model_selector: tid = "gemini-2.5-flash"
         elif "2.0 Flash" in model_selector: tid = "gemini-2.0-flash-exp"
-        return run_gemini(full_history, tid)
+        return run_gemini(formatted_history, tid)
     
     elif "Mistral" in model_selector:
         tid = "mistral-large-latest"
@@ -202,19 +182,19 @@ def router(message, history, model_selector):
         elif "2512" in model_selector: tid = "mistral-large-2512"
         elif "Magistral" in model_selector: tid = "magistral-medium-latest"
         elif "Codestral" in model_selector: tid = "codestral-2508"
-        return run_mistral(full_history, tid)
+        return run_mistral(formatted_history, tid)
 
     elif "Groq" in model_selector:
-        return run_groq(full_history, "llama-3.3-70b-versatile")
+        return run_groq(formatted_history, "llama-3.3-70b-versatile")
         
     elif "H200" in model_selector:
-        return run_local_h200(full_history)
+        return run_local_h200(formatted_history)
         
     return "Modelo desconhecido."
 
 # --- UI ---
-with gr.Blocks(fill_height=True) as demo:
-    gr.Markdown("# 🔀 APIDOST V7 (Gradio 5 Native)")
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# 🔀 APIDOST V7 (Gradio 4 Fix)")
     
     with gr.Row():
         model_dropdown = gr.Dropdown(
@@ -231,9 +211,9 @@ with gr.Blocks(fill_height=True) as demo:
             interactive=True
         )
 
+    # REMOVIDO type="messages"
     chat = gr.ChatInterface(
         fn=router,
-        type="messages", # O SEGREDO! Isso diz pro Gradio mandar a history como dict
         additional_inputs=[model_dropdown],
         multimodal=True,
     )