app.py

import streamlit as st
import pandas as pd
import faiss
import numpy as np
from sentence_transformers import SentenceTransformer
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, AutoModelForSeq2SeqLM
from PIL import Image
from io import BytesIO
import requests
import torch
import os

# --- 1. Load Recipes Dataset ---
@st.cache_data
def load_recipes():
    try:
        recipes_df = pd.read_csv("recipes.csv")
        recipes_df = recipes_df.rename(columns={"recipe_name": "title", "directions": "instructions"})
        recipes_df = recipes_df[['title', 'ingredients', 'instructions', 'img_src']]
        recipes_df.fillna("", inplace=True)
        recipes_df["ingredients"] = recipes_df["ingredients"].str.lower()
        recipes_df["combined_text"] = recipes_df["title"] + " " + recipes_df["ingredients"]
        return recipes_df
    except Exception as e:
        st.error(f"⚠ Error loading recipes: {e}")
        return pd.DataFrame()

recipes_df = load_recipes()

# --- 2. Load SentenceTransformer Model ---
@st.cache_resource
def load_embedding_model():
    return SentenceTransformer("all-MiniLM-L6-v2")

embedding_model = load_embedding_model()

# --- 3. Initialize FAISS Index ---
FAISS_INDEX_PATH = "faiss_index.idx"

@st.cache_resource
def build_faiss_index():
    dim = 384
    index = faiss.IndexFlatL2(dim)
    
    if os.path.exists(FAISS_INDEX_PATH):
        index = faiss.read_index(FAISS_INDEX_PATH)
    else:
        combined_texts = recipes_df["combined_text"].tolist()
        embeddings = embedding_model.encode(combined_texts, normalize_embeddings=True, convert_to_numpy=True).astype(np.float32)
        index.add(embeddings)
        faiss.write_index(index, FAISS_INDEX_PATH)
        
    return index

faiss_index = build_faiss_index()

# --- 4. Load Classification Model (for query classification) ---
@st.cache_resource
def load_classifier():
    return pipeline("zero-shot-classification", model="facebook/bart-large-mnli")

classifier = load_classifier()

# --- Load T5 Model (for sequence-to-sequence tasks and recipe generation) ---
@st.cache_resource
def load_llm_model():
    model_name = "google/flan-t5-small"  # T5-based model for sequence-to-sequence tasks
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_name, torch_dtype=torch.float32, device_map="cpu")  # Use AutoModelForSeq2SeqLM
    return model, tokenizer

llm_model, tokenizer = load_llm_model()

# --- 6. Query Classification ---
def classify_query(query):
    labels = ["recipe", "general", "ingredients"]
    result = classifier(query, candidate_labels=labels)
    return result['labels'][0]

# --- 7. Recipe Retrieval ---
def retrieve_recipes(query, top_k=3):
    query_embedding = embedding_model.encode(query, convert_to_numpy=True).reshape(1, -1)
    distances, indices = faiss_index.search(query_embedding, top_k)
    return recipes_df.iloc[indices[0]] if len(indices[0]) > 0 else None

# --- 8. Generate a Recipe (Recipe Generation with T5-based model) ---
def generate_recipe(ingredients):
    system_prompt = """You are a creative AI chef. Create a recipe based on the following ingredients: {ingredients}. Be sure to include instructions and specify the cuisine type."""
    full_prompt = system_prompt.format(ingredients=ingredients)
    
    inputs = tokenizer(full_prompt, return_tensors="pt", truncation=True).to("cpu")
    
    with torch.no_grad():
        output = llm_model.generate(**inputs, max_length=300, num_beams=2, early_stopping=True)
    
    generated_recipe = tokenizer.decode(output[0], skip_special_tokens=True)
    return generated_recipe.strip()

# --- 9. Display Image ---
def display_image(image_url, recipe_name):
    try:
        response = requests.get(image_url, timeout=5)
        response.raise_for_status()
        image = Image.open(BytesIO(response.content))
        st.image(image, caption=recipe_name, use_container_width=True)
    except requests.exceptions.RequestException:
        st.image("https://via.placeholder.com/300?text=No+Image", caption=recipe_name, use_container_width=True)

# --- 10. Handle Recipe Queries (Retrieve or Generate) ---
def handle_recipe_query(query):
    retrieved_recipes = retrieve_recipes(query, top_k=3)
    if retrieved_recipes is not None and not retrieved_recipes.empty:
        st.subheader("🍴 Found Recipes:")
        for _, recipe in retrieved_recipes.iterrows():
            st.markdown(f"### {recipe['title']}")
            st.write(f"**Ingredients:** {recipe['ingredients']}")
            st.write(f"**Instructions:** {recipe['instructions']}")
            display_image(recipe.get('img_src', ''), recipe['title'])
    else:
        st.write("⚠ No recipes found. Generating a new recipe based on your ingredients...")
        generated_recipe = generate_recipe(query)
        st.subheader("🍳 Generated Recipe:")
        st.write(generated_recipe)

# --- 11. Answer General Queries ---
def answer_question(query):
    system_prompt = """You are a helpful AI assistant. Provide clear and concise responses to user questions."""
    full_prompt = f"{system_prompt}\nUser Query: {query}"
    
    inputs = tokenizer(full_prompt, return_tensors="pt", truncation=True).to("cpu")
    
    with torch.no_grad():
        output = llm_model.generate(**inputs, max_length=200, num_beams=2, early_stopping=True)
    
    response = tokenizer.decode(output[0], skip_special_tokens=True)
    return response.strip()

# --- 12. Streamlit UI ---
st.title("🍽️ AI Recipe Assistant")

user_query = st.text_input('Enter your Ingredient or recipe search query (eg: "what can i make with strawberries? or Give me the recipe of pizza"):', "", key="main_query_input")

if st.button("Ask AI"):
    if user_query:
        with st.spinner("🔍 Searching..."):
            query_type = classify_query(user_query)
            if query_type == "recipe" or query_type == "ingredients":
                handle_recipe_query(user_query)
            else:
                st.subheader("🤖 AI Answer:")
                ai_answer = answer_question(user_query)
                st.write(ai_answer)
    else:
        st.warning("⚠ Please enter a query before clicking 'Ask AI'.")