Update app.py

app.py

@@ -20,7 +20,7 @@ def extract_title_and_ingredients(sample):
 
 def extract_each_feature(sample):
     """
-    Extract each feature of a recipe from a sample, and clean up potential duplications.
+    FIXED: Extract each feature of a recipe from a sample and clean up potential duplications.
     """
     full_text = sample['input']
     
@@ -60,12 +60,10 @@ print("Loading embedding model...")
 model_name = "all-MiniLM-L6-v2"
 embedding_model = SentenceTransformer(f"sentence-transformers/{model_name}")
 
-# Compute embeddings for the dataset
 print("Generating embeddings...")
 embeddings = embedding_model.encode(dataset['text_for_embedding'], show_progress_bar=True)
 embeddings = np.array(embeddings, dtype=np.float32)
 
-# Build FAISS index for similarity search
 print("Building FAISS index...")
 dimension = embeddings.shape[1]
 index = faiss.IndexFlatL2(dimension)
@@ -74,7 +72,7 @@ print(f"Index is ready. Total vectors in index: {index.ntotal}")
 
 # --- 3. SYNTHETIC GENERATION ---
 print("Loading generative model...")
-generator = pipeline('text-generation', model='gpt2')
+generator = pipeline('text-generation', model='distilgpt2')
 
 def get_recommendations_and_generate(query_ingredients, k=3):
     # 1. Get Recommendations
@@ -92,20 +90,46 @@ def get_recommendations_and_generate(query_ingredients, k=3):
         }
         results.append(recipe)
         
-    # 2. Generate a new recipe idea
-    prompt = f"Create a simple recipe title and a list of ingredients using: {query_ingredients}."
-    generated_text = generator(prompt, max_length=100, num_return_sequences=1)[0]['generated_text']
-    
-    # Clean up generated text to be more readable
-    generated_recipe_parts = generated_text.split("Ingredients:")
-    generated_title = generated_recipe_parts[0].replace(prompt, "").strip()
-    generated_ingredients = generated_recipe_parts[1].strip() if len(generated_recipe_parts) > 1 else "Could not determine ingredients."
-    
-    generated_recipe = {
-        "title": generated_title,
-        "ingredients": generated_ingredients,
-        "directions": "This is an AI-generated idea. Directions are not provided."
-    }
+    # 2. Generate a new recipe with a structured "few-shot" prompt
+    prompt = f"""Create a full recipe including a title, ingredients, and directions based on the following items: {query_ingredients}.
+
+### Title:
+[Recipe Title]
+
+### Ingredients:
+- [Ingredient 1]
+- [Ingredient 2]
+- [Ingredient 3]
+
+### Directions:
+1. [Step 1]
+2. [Step 2]
+3. [Step 3]
+---
+Recipe:
+### Title:
+"""
+    # Generate the recipe text
+    generated_outputs = generator(prompt, max_new_tokens=200, num_return_sequences=1)
+    generated_text = generated_outputs[0]['generated_text'].replace(prompt, "").strip()
+
+    # 3. Parse the generated text into a structured format
+    try:
+        title_part, rest = generated_text.split("### Ingredients:", 1)
+        ingredients_part, directions_part = rest.split("### Directions:", 1)
+        
+        generated_recipe = {
+            "title": title_part.strip(),
+            "ingredients": ingredients_part.strip(),
+            "directions": directions_part.strip()
+        }
+    except ValueError:
+        # Fallback if the model doesn't follow the format perfectly
+        generated_recipe = {
+            "title": "AI Generated Recipe",
+            "ingredients": "Could not determine ingredients.",
+            "directions": generated_text
+        }
     
     return results[0], results[1], results[2], generated_recipe