import gradio as gr
import pandas as pd
import pickle
import numpy as np
import os
import random
import base64
from huggingface_hub import InferenceClient
from sklearn.metrics.pairwise import cosine_similarity
from datasets import load_dataset
from IO_pipeline import RecipeDigitalizerPipeline
# ==========================================
# 1. SETUP & DATA LOADING (HYBRID)
# ==========================================
hf_token = os.getenv("HF_TOKEN")
API_MODEL = "BAAI/bge-small-en-v1.5"
client = InferenceClient(token=hf_token) if hf_token else None
print("⏳ Initializing Data Loading...")
# --- A. Load Text Data from Hugging Face Dataset ---
try:
print(" ...Downloading recipes from HF Dataset (Liori25/10k_recipes)")
dataset = load_dataset("Liori25/10k_recipes", split="train")
df_recipes = dataset.to_pandas()
print(f"✅ Recipes Loaded! Count: {len(df_recipes)}")
except Exception as e:
print(f"❌ Error loading HF Dataset: {e}")
df_recipes = pd.DataFrame({'Title': [], 'Raw_Output': []})
# --- B. Load Embeddings from Local File (Space Repo) ---
try:
print(" ...Loading embeddings from local 'recipe_embeddings.pkl'")
if os.path.exists('recipe_embeddings.pkl'):
with open('recipe_embeddings.pkl', 'rb') as f:
data = pickle.load(f)
# Logic to handle different pickle formats
if isinstance(data, dict):
stored_embeddings = np.array(data['embeddings'])
elif isinstance(data, pd.DataFrame):
target_col = next((c for c in ['embedding', 'embeddings', 'vectors'] if c in data.columns), None)
stored_embeddings = np.vstack(data[target_col].values) if target_col else data
else:
stored_embeddings = data
print(f"✅ Embeddings Loaded! Shape: {stored_embeddings.shape}")
else:
print("❌ 'recipe_embeddings.pkl' not found locally.")
stored_embeddings = None
except Exception as e:
print(f"❌ Error loading pickle file: {e}")
stored_embeddings = None
# --- C. Safety Check ---
if stored_embeddings is not None and not df_recipes.empty:
if len(stored_embeddings) != len(df_recipes):
print(f"⚠️ WARNING: Row mismatch! Recipes: {len(df_recipes)}, Embeddings: {len(stored_embeddings)}")
# ==========================================
# 2. HELPER: IMAGE TO BASE64
# ==========================================
def image_to_base64(image_path):
if not os.path.exists(image_path):
return "R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7"
with open(image_path, "rb") as img_file:
return base64.b64encode(img_file.read()).decode('utf-8')
logo_b64 = image_to_base64("cookbook logo.png")
profile_b64 = image_to_base64("chef.avif")
process_b64 = image_to_base64("preview of process.jpg")
# ==========================================
# 3. BACKEND LOGIC
# ==========================================
def get_embedding_via_api(text):
if not client: raise ValueError("HF_TOKEN missing")
response = client.feature_extraction(text, model=API_MODEL)
return np.array(response)
def find_similar_recipes_list(query_text):
if stored_embeddings is None: return ["Database error: Embeddings missing."] * 3
if df_recipes.empty: return ["Database error: Recipes missing."] * 3
query_vec = get_embedding_via_api("Represent this recipe for retrieving similar dishes: " + query_text)
if len(query_vec.shape) == 1: query_vec = query_vec.reshape(1, -1)
# Calculate Similarity
scores = cosine_similarity(query_vec, stored_embeddings)[0]
top_indices = scores.argsort()[-3:][::-1]
# Identify column names
cols = df_recipes.columns
ing_col = next((c for c in cols if 'ingredient' in c.lower()), None)
inst_col = next((c for c in cols if 'instruction' in c.lower()), None)
results_list = []
# --- HELPER TO CHECK FOR ERRORS & HIDE BLOCKS ---
def clean_and_validate(raw_text):
val = str(raw_text).strip()
# 1. Clean list syntax (['...'])
if val.startswith("[") and val.endswith("]"):
val = val[1:-1].replace("'", "").replace('"', "").strip()
val_lower = val.lower()
# 2. Basic Empty Checks
if val_lower in ['nan', 'none', 'null', '[]', '']:
return None
# 3. STRICT ERROR CHECK: If "parse error" or "error" is in the text, return None to hide the block
if "parse error" in val_lower or "error" in val_lower:
return None
return val
for idx in top_indices:
score = scores[idx]
row = df_recipes.iloc[idx]
title = row.get('Title', 'Unknown Recipe')
score_display = f"{score:.3%}"
content_parts = []
# 1. Check Ingredients
if ing_col:
cleaned_ing = clean_and_validate(row[ing_col])
# Only add to display if valid AND no error found
if cleaned_ing:
content_parts.append(f"🛒 INGREDIENTS: {cleaned_ing}")
# 2. Check Instructions
if inst_col:
cleaned_inst = clean_and_validate(row[inst_col])
# Only add to display if valid AND no error found
if cleaned_inst:
content_parts.append(f"🍳 INSTRUCTIONS: {cleaned_inst}")
# 3. Fallback logic
if not content_parts:
# If both were hidden (due to errors) or empty, check raw output
raw_out = str(row.get('Raw_Output', 'No details available.'))
# Also hide Raw Output if it contains an error
if "parse error" in raw_out.lower() or "error" in raw_out.lower():
display_text = "Details unavailable for this recipe."
else:
display_text = raw_out
else:
display_text = "
CookBook AITurning Handwritten Recipes into a Digital Recipe.
v4.4
""")
with gr.Row():
# --- LEFT SIDEBAR ---
with gr.Column(scale=1, min_width=200):
gr.HTML(f"""
My Profile
""")
gr.HTML("")
nav_digital = gr.Button("✨ AI Digitizer", elem_classes=["nav-btn", "selected"])
nav_feed = gr.Button("📰 News Feed", elem_classes=["nav-btn"])
nav_about = gr.Button("ℹ️ About", elem_classes=["nav-btn"])
# --- CENTER CONTENT ---
with gr.Column(scale=3):
# === VIEW 1: AI DIGITALIZER ===
with gr.Group(visible=True) as digitalizer_view:
with gr.Row(elem_classes=["gap-fix"]):
with gr.Column(scale=1):
with gr.Group(elem_classes=["content-card"]):
input_img = gr.Image(type="filepath", label="Upload", height=300)
magic_btn = gr.Button("✨ Convert to Digital", variant="primary", size="lg")
# --- UPDATED EXAMPLES HERE ---
gr.Examples(
examples=[
["quick_tries_images/applecrisp.jpg"],
["quick_tries_images/meatballs recipe.jpg"],
["quick_tries_images/Apple Dapple (aka Fresh Apple Cake).jfif"]
],
inputs=input_img,
label="Or try these examples:",
cache_examples=False
)
with gr.Column(scale=1):
with gr.Group(elem_classes=["content-card"]):
out_text = gr.Textbox(label="Result", value="Here your digitalized recipe will be presented", lines=20, interactive=False, show_label=False)
gr.HTML("")
gr.Markdown("### 3. Similar Recipes from Database")
with gr.Row():
with gr.Column(elem_classes=["sim-card"]) as c1_box:
sim1 = gr.Markdown("Once you will upload your scanned recipe, we will share similar recipes!")
with gr.Row(visible=False) as c1_btns:
gr.Button("👍 Like", size="sm", variant="secondary")
gr.Button("↗️ Share", size="sm", variant="secondary")
with gr.Column(elem_classes=["sim-card"], visible=False) as c2_box:
sim2 = gr.Markdown("")
with gr.Row():
gr.Button("👍 Like", size="sm", variant="secondary")
gr.Button("↗️ Share", size="sm", variant="secondary")
with gr.Column(elem_classes=["sim-card"], visible=False) as c3_box:
sim3 = gr.Markdown("")
with gr.Row():
gr.Button("👍 Like", size="sm", variant="secondary")
gr.Button("↗️ Share", size="sm", variant="secondary")
magic_btn.click(ui_update_pipeline, input_img, [out_text, sim1, c1_btns, c2_box, sim2, c3_box, sim3])
# === VIEW 2: FEED ===
with gr.Column(visible=False, elem_id="feed-container") as feed_view:
if not df_recipes.empty:
feed_samples = df_recipes.sample(10)
for index, row in feed_samples.iterrows():
user_name = random.choice(["Grandma Rose", "Chef Mike", "Sarah J."])
emoji = random.choice(["🥘", "🥗", "🍰", "🌮"])
time_options = ["2h", "3h", "4h", "6h", "9h", "12h", "a day ago", "2 days ago"]
post_time = random.choice(time_options)
raw_desc = str(row.get('Raw_Output', 'Delicious recipe...'))[:250]
title_feed = row.get('Title', 'Recipe')
with gr.Group(elem_classes=["content-card"]):
gr.HTML(f"""
{emoji}
{user_name} {post_time} · 🌍 Public
""")
gr.Markdown(f"### {title_feed}")
gr.Markdown(f"{raw_desc}...")
with gr.Row():
gr.Button("👍 Like", size="sm", variant="secondary")
gr.Button("💬 Comment", size="sm", variant="secondary")
gr.Button("↗️ Share", size="sm", variant="secondary")
else:
gr.Markdown("⚠️ Database is empty.")
# === VIEW 3: ABOUT (UPDATED) ===
with gr.Group(visible=False) as about_view:
with gr.Group(elem_classes=["content-card"]):
gr.Markdown("""
# Goal Project
The goal of this project is to develop an app that takes a scanned image of a handwritten recipe as input, generates text using a VLM, and based on the extracted text, suggests 3 similar recipes from a 10K dataset of synthetic recipes. Our app will bridge the gap between analog culinary heritage and digital discovery.
### About Us
This app was developed by **Shahar Firshtman** and **Lior Feinstein**, 2nd year students for Economics and data science.
""")
# Process Image
gr.HTML(f"""
Process Overview
""")
# --- RIGHT COLUMN ---
with gr.Column(scale=1, min_width=200):
# 1. Trending
gr.Markdown("### Trending Recipes")
def trend_box(title, likes):
return f"
My Profile
