Update FastAPI_app.py

FastAPI_app.py

@@ -7,7 +7,6 @@ import time
 import traceback
 import threading
 import signal
-from dotenv import load_dotenv
 
 import uvicorn
 import numpy as np
@@ -25,10 +24,6 @@ import google.generativeai as genai
 from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
 
 
-# CONFIGURATION
-# Load environment variables
-load_dotenv()
-
 # Ingredient model (load once)
 MODEL_PATH = "models/ingredient_model.h5"
 if not os.path.exists(MODEL_PATH):
@@ -50,7 +45,7 @@ else:
         'sweetpotato', 'tomato', 'turnip', 'watermelon'
     ]
 
-# Phi-3.5-mini-instruct local model loading
+# Timeout handler
 def timeout_handler(signum, frame):
     raise TimeoutError("Model load timed out after 300s")
 
@@ -86,21 +81,25 @@ def load_Qwen():
 def generate_recipe_qwen(ingredient_names):
     tokenizer, model = load_Qwen() 
     
-    prompt = f"""
-            You are an AI chef. Create a short recipe using only: {', '.join(ingredient_names)}.
-            Include:
-            - Recipe name
-            - One-sentence description
-            - Ingredients list with quantities
-            - 6-10 concise steps
-            - Optional tips
-            RETURN RESULT IN MARKDOWN FORMAT ONLY.
-            """
+    messages = [
+        {"role": "system", "content": "You are a helpful chef. Always respond ONLY with clean markdown, no extra text, no greetings, no explanations."},
+        {"role": "user", "content": f"""Create a delicious recipe using only these ingredients: {', '.join(ingredient_names)}
+
+        Return ONLY clean markdown with:
+        - Recipe title (## Title)
+        - One-sentence description
+        - Ingredients list with quantities
+        - Numbered steps
+        - Optional tip"""}
+            ]
         
-    inputs = tokenizer(prompt, return_tensors="pt")
+    # Use Qwen chat template
+    input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
+    
     outputs = model.generate(
         inputs.input_ids,
-        max_new_tokens=512,
+        max_new_tokens=450,
         temperature=0.7,
         top_p=0.9,
         do_sample=True
@@ -108,7 +107,8 @@ def generate_recipe_qwen(ingredient_names):
     
     recipe_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
     # Strip the prompt part
-    return recipe_text.split("assistant")[-1].strip()
+    return recipe_text.split("<|assistant|>")[-1].strip()
+    
 
 
 # Infer uploaded image function