Update app.py

app.py

@@ -1,4 +1,4 @@
-# app.py — PromptCraft: Refinamiento Estructural de Prompts
+# app.py — PromptCraft: Refinamiento Estructural de Prompts (con fallback a Together vía HF Router)
 import gradio as gr
 import os
 import time
@@ -18,6 +18,7 @@ class LlamaRefiner:
         hf_token = os.getenv("PS")
         if not hf_token:
             raise ValueError("Secret 'PS' (HF_TOKEN) no encontrado.")
+        # Cliente para fallback original (Hyperbolic)
         self.hf_client = InferenceClient(api_key=hf_token)
         self.agent = ImprovedSemanticAgent()
         if not self.agent.is_ready:
@@ -25,11 +26,12 @@ class LlamaRefiner:
             logger.info(f"Inicialización del agente: {init_msg}")
             if not self.agent.is_ready:
                 logger.error("❌ Agente NO está listo.")
+        
         logger.info("🚀 Cargando traductor local (es → en)...")
         self.translator = pipeline(
             "translation_es_to_en",
             model="Helsinki-NLP/opus-mt-es-en",
-            device=-1
+            device=-1  # CPU
         )
         logger.info("✅ Traductor local listo.")
 
@@ -38,31 +40,10 @@ class LlamaRefiner:
             return text
         try:
             result = self.translator(text, max_length=250, clean_up_tokenization_spaces=True)
-            raw_translation = result[0]['translation_text'].strip()
+            return result[0]['translation_text'].strip()
         except Exception as e:
             logger.warning(f"Traducción local fallida: {e}. Usando texto original.")
-            raw_translation = text
-
-        user_text_lower = text.lower()
-        output = raw_translation
-
-        if any(kw in user_text_lower for kw in ["llamas", "ardiendo", "quem", "incendi", "fuego"]):
-            output = output.replace("fiery", "on fire")
-            if not any(term in output.lower() for term in ["on fire", "burning", "in flames", "ablaze", "aflame"]):
-                output = output + " on fire"
-
-        if any(kw in user_text_lower for kw in ["oro", "dorado"]):
-            if "golden" not in output.lower() and "gold" not in output.lower():
-                if any(w in output.lower() for w in ["statue", "sculpture", "figure"]):
-                    output = output + " made of gold"
-                else:
-                    output = output + " golden"
-
-        if any(kw in user_text_lower for kw in ["congelado", "hielo", "helado", "ice"]):
-            if not any(term in output.lower() for term in ["frozen", "ice", "icy"]):
-                output = output + " frozen"
-
-        return output.strip()
+            return text
 
     def retrieve_similar_examples(self, user_prompt_en: str, category: str = "auto", k: int = 6) -> list:
         if not self.agent.is_ready:
@@ -71,6 +52,7 @@ class LlamaRefiner:
             query_embedding = self.agent.embedding_model.encode([user_prompt_en], convert_to_numpy=True, normalize_embeddings=True)[0]
             query_embedding = query_embedding.astype('float32').reshape(1, -1)
             distances, indices = self.agent.index.search(query_embedding, 50)
+            
             candidates = []
             for idx in indices[0]:
                 if 0 <= idx < len(self.agent.indexed_examples):
@@ -80,19 +62,25 @@ class LlamaRefiner:
                     if isinstance(caption, str) and len(caption) > 10:
                         if category == "auto" or ex_category == category:
                             candidates.append((idx, caption, ex_category))
+            
             if not candidates:
                 return []
+            
             if len(candidates) <= k:
                 return [cap for _, cap, _ in candidates]
+            
             candidate_texts = [cap for _, cap, _ in candidates]
             pairs = [[user_prompt_en, cand] for cand in candidate_texts]
             scores = self.agent.reranker.predict(pairs)
+            
             scored = [(candidates[i][1], scores[i]) for i in range(len(candidates))]
             scored.sort(key=lambda x: x[1], reverse=True)
+            
             top_examples = [ex for ex, _ in scored[:k]]
             return top_examples
+            
         except Exception as e:
-            logger.error(f"Error en recuperación: {e}")
+            logger.error(f"Error en recuperación con re-ranking y categoría: {e}")
             try:
                 return [
                     self.agent.indexed_examples[idx]['caption'] 
@@ -102,13 +90,103 @@ class LlamaRefiner:
             except:
                 return []
 
+    def _clean_output(self, text: str) -> str:
+        text = text.strip()
+        if text.startswith(("Here is", "Final:", "Output:", '"', "'")):
+            text = text.split(":", 1)[-1].strip().strip("\"'")
+        return text
+
     def refine_with_llm(self, user_prompt: str, category: str = "auto") -> Tuple[str, str, list]:
         user_prompt_en = self.translate_to_english(user_prompt)
         examples = self.retrieve_similar_examples(user_prompt_en, category=category, k=6)
 
-        # ✅ Usar SOLO enriquecimiento local (como en la versión que funcionaba)
+        if not examples:
+            fallbacks = {
+                "entity": [
+                    "an elderly maya man weaving a hammock under a ceiba tree, golden hour light filtering through leaves, Antigua Guatemala setting, hyperrealistic style",
+                    "a young indigenous woman in traditional Kekchi attire by Lake Atitlán, morning mist, volcano backdrop, soft natural light, documentary photography"
+                ],
+                "style": [
+                    "oil painting of a forest in autumn, warm amber and crimson tones, impasto brushstrokes, style of Vincent van Gogh",
+                    "cyberpunk cityscape at night, neon reflections on wet streets, cinematic lighting, style of Blade Runner 2049"
+                ],
+                "composition": [
+                    "a lone wolf on a snowy mountain peak, northern lights in the sky, wind blowing snow, rule of thirds composition, photorealistic wildlife"
+                ],
+                "imaginative": [
+                    "a floating island with ancient ruins, waterfalls cascading into clouds, golden hour, fantasy concept art, highly detailed"
+                ],
+                "text": [
+                    "minimalist typography design, the word 'LIBERTAD' in bold sans-serif, high contrast black on white, professional layout"
+                ],
+                "auto": [
+                    "an elderly maya man weaving a hammock under a ceiba tree, golden hour light filtering through leaves, Antigua Guatemala setting, hyperrealistic style, intricate textures of rope and bark",
+                    "a cyberpunk street at night in Tokyo, neon signs reflecting on wet pavement, rain mist in air, distant flying cars, cinematic wide shot, Blade Runner atmosphere",
+                    "a library interior with tall oak bookshelves, sunbeams through stained glass windows, dust particles floating, oil painting style, warm amber tones, masterpiece",
+                    "a lone wolf howling on a snowy mountain peak, northern lights in the sky, wind blowing snow, photorealistic wildlife photography, 8k detailed fur",
+                    "a steampunk airship floating above Victorian London, copper pipes and brass gears, cloudy sky, detailed machinery, concept art by Jakub Rozalski",
+                    "a young woman in traditional Kekchi attire standing by Lake Atitlán, morning mist, volcano backdrop, soft natural light, documentary photography style"
+                ]
+            }
+            examples = fallbacks.get(category, fallbacks["auto"])
+            logger.warning(f"No se encontraron ejemplos para categoría '{category}'. Usando fallback.")
+
+        system_message = (
+            "You are a prompt engineering analyst for diffusion models (Midjourney, FLUX, SDXL). "
+            "Analyze the DESCRIPTIVE GRAMMAR (word order, phrasing, element sequence) used in the reference prompts below. "
+            "Reconstruct the user's concept using that exact same descriptive logic. "
+            "Do NOT follow a predefined template (e.g. subject→lighting→style). "
+            "Do NOT invent elements not implied by the user. "
+            "Preserve the user's core intent exactly. "
+            "Output ONLY the final prompt in English. No explanations, no markdown."
+        )
+
+        user_message = "User concept:\n" + user_prompt_en + "\n\nReference prompts (observe their descriptive grammar):\n" + "\n".join(examples)
+
+        # ✅ PRIMERO: Intentar con Together vía HF Router (más estable)
+        try:
+            client = OpenAI(
+                base_url="https://router.huggingface.co/v1",
+                api_key=os.getenv("PS")
+            )
+            completion = client.chat.completions.create(
+                model="meta-llama/Llama-3.2-3B-Instruct:together",
+                messages=[
+                    {"role": "system", "content": system_message},
+                    {"role": "user", "content": user_message}
+                ],
+                max_tokens=250,
+                temperature=0.2,
+                timeout=30.0
+            )
+            refined = self._clean_output(completion.choices[0].message.content)
+            info = f"🧠 Refinado con Llama-3.2-3B vía Together (HF Router, {len(examples)} ejemplos, categoría: {category})."
+            return refined, info, examples
+
+        except (APIError, Timeout, Exception) as e1:
+            logger.error(f"Error con Together (HF Router): {e1}")
+
+            # 🔄 SEGUNDO: Intentar con HF predeterminado (Hyperbolic)
+            try:
+                completion = self.hf_client.chat.completions.create(
+                    model="meta-llama/Llama-3.2-3B-Instruct",
+                    messages=[
+                        {"role": "system", "content": system_message},
+                        {"role": "user", "content": user_message}
+                    ],
+                    max_tokens=250,
+                    temperature=0.2
+                )
+                refined = self._clean_output(completion.choices[0].message.content)
+                info = f"🧠 Fallback: HF (Hyperbolic, {len(examples)} ejemplos)."
+                return refined, info, examples
+
+            except Exception as e2:
+                logger.error(f"También falló Hyperbolic: {e2}")
+
+        # ⚠️ ÚLTIMO RECURSO: enriquecimiento semántico local
         enhanced_prompt, _ = self.agent.enhance_prompt(user_prompt_en, category=category)
-        return enhanced_prompt.strip(), "✨ Enriquecimiento semántico local", examples
+        return enhanced_prompt.strip(), f"⚠️ LLMs no disponibles. Usando enriquecimiento semántico (categoría: {category}).", examples
 
 class SDXLGenerator:
     def __init__(self):
@@ -145,6 +223,7 @@ def create_interface():
             return "", "", "Prompt vacío."
         if refiner is None:
             return "", "", "Servicios no disponibles."
+
         progress(0.2, desc="🌍 Traduciendo...")
         category_map = {
             "Automática": "auto",
@@ -155,6 +234,7 @@ def create_interface():
             "Texto": "text"
         }
         category_en = category_map.get(category_es, "auto")
+        
         refined, info, examples = refiner.refine_with_llm(prompt, category_en)
         examples_text = "\n".join(f"{i+1}. {ex}" for i, ex in enumerate(examples)) if examples else "Ninguno"
         status = f"Prompt refinado: {refined}\n{info}"
@@ -165,6 +245,7 @@ def create_interface():
             return None, "❌ No hay prompt refinado. Primero haz clic en 'Refinar prompt'."
         if generator is None:
             return None, "❌ Generador no inicializado."
+
         aspect_ratios = {
             "1:1": (1024, 1024),
             "16:9": (1344, 768),
@@ -175,12 +256,12 @@ def create_interface():
             "9:21": (640, 1536),
         }
         width, height = aspect_ratios.get(aspect_ratio, (1024, 1024))
-        progress(0.5, desc="🎨 Generando imagen...")
+        progress(0.5, desc="🎨 Generando imagen (puede tardar 10-20s)...")
         try:
             image_path, gen_msg = generator.generate_image(refined_prompt, width, height)
             return image_path, gen_msg
         except Exception as e:
-            error_msg = f"❌ Error: {str(e)}"
+            error_msg = f"❌ Error al generar: {str(e)}"
             logger.error(error_msg)
             return None, error_msg
 
@@ -209,8 +290,10 @@ def create_interface():
                 prompt_input = gr.Textbox(label="Tu idea (en castellano)", lines=3, placeholder="Ej: un mercado en Antigua Guatemala...")
                 category_es = gr.Dropdown(label="Categoría", choices=CATEGORY_CHOICES_ES, value="Automática")
                 aspect = gr.Dropdown(label="Proporción", choices=["1:1", "16:9", "9:16", "4:3", "3:4", "21:9", "9:21"], value="1:1")
+                
                 refine_btn = gr.Button("🔄 Refinar prompt", variant="secondary")
                 generate_btn = gr.Button("🎨 Generar imagen", variant="primary")
+                
             with gr.Column():
                 refined_output = gr.Textbox(
                     label="Prompt refinado (inglés)",