update prompt and output format

src/agents/mise_en_place.py

@@ -5,40 +5,40 @@ from typing import Optional
 import os
 import torch
 from transformers import AutoProcessor, MiniCPMV4_6ForConditionalGeneration
+from src import config
 
 MODEL_ID = "openbmb/MiniCPM-V-4.6"
 
 # Detectar el entorno: Si estás en Hugging Face ZeroGPU, forzamos "auto" (CPU inicial).
 # Si estás en tu PC local con una GPU Nvidia, usamos "cuda". De lo contrario, "cpu".
 if os.environ.get("SPACE_ID") is not None:
-    # Estamos dentro de un Hugging Face Space
+    # Hugging Face Space
     ENV_DEVICE = "auto"
     print("Entorno detectado: Hugging Face Spaces (ZeroGPU).")
 else:
-    # Estamos en tu máquina local
+    # máquina local
     ENV_DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
     print(f"Entorno detectado: Local. Usando dispositivo: {ENV_DEVICE.upper()}")
 
-print(f"Loading processor: {MODEL_ID}")
 processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
 
-print(f"Loading model: {MODEL_ID}")
 model = MiniCPMV4_6ForConditionalGeneration.from_pretrained(
     MODEL_ID,
     torch_dtype=torch.bfloat16,
     attn_implementation="sdpa",
     trust_remote_code=True,
-    device_map=ENV_DEVICE # <--- Configuración dinámica según el entorno
+    device_map=ENV_DEVICE
 ).eval()
 
-_prompt = "Describe the given image in detail. Be honest, do not say liers."
-
+_PROMPT_PATH = config.PROMPTS_DIR / "vision_prompt.txt"
+def _prompt() -> str:
+    return _PROMPT_PATH.read_text(encoding="utf-8")
 
 @spaces.GPU(duration=30)
 def identify_ingredients(image: Optional[Image.Image]) -> list[str]:
     try:
         img = image.convert("RGB")
-        messages = [{"role": "user", "content": [{"type": "image", "image": img}, {"type": "text", "text": _prompt}]}]
+        messages = [{"role": "user", "content": [{"type": "image", "image": img}, {"type": "text", "text": _prompt()}]}]
 
         # Procesamiento de tokens
         inputs = processor.apply_chat_template(
@@ -46,8 +46,6 @@ def identify_ingredients(image: Optional[Image.Image]) -> list[str]:
             enable_thinking=False, processor_kwargs={"downsample_mode": "16x", "max_slice_nums": 9, "use_image_id": True}
         )
 
-        # LOCAL: model.device será 'cuda' (Tu tarjeta gráfica)
-        # ZERO_GPU: model.device cambiará a 'cuda:0' dinámicamente gracias al decorador
         inputs = inputs.to(model.device)
 
         for k, v in inputs.items():
@@ -60,10 +58,11 @@ def identify_ingredients(image: Optional[Image.Image]) -> list[str]:
         generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
         raw_output = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
 
-        # (Tu lógica de limpieza de JSON e ingredientes aquí...)
         print(raw_output)
-        return ['testing']
+        data = json.loads(raw_output)
+        ingredients = [str(x).lower().strip() for x in data.get("ingredients", [])]
+        return [x for x in ingredients if x] or list("There are no ingredients detected")
         
     except Exception as e:
         print(f"Error: {e}")
-        return ["there are not ingredients detected"]
+        return ["There are not ingredients detected"]

src/prompts/vision_prompt.txt

@@ -1 +1,20 @@
-Describe the given image in detail. Be honest, do not say liers.
+# Context: 
+You are going to be provided by an image of a fridge, or image where you will see different ingredientes or food supplies.
+
+# Task:
+You have to detect every kind of food, such as vegetables, meat, condiments, oil, milk, eggs, etc. After detect specific ingredients, you have to return that list as your response.
+
+# Output format:
+Your output should be in json format, follow the example: 
+    {
+        'ingredients': [
+            'ingredient 1',
+            'ingredient 2',
+            ...
+        ]
+    }
+
+# Contraints
+If you do not see any kind of ingredient you must return a empty json.
+You have to be honest, only add existent ingredients.
+Follow the output format always.