app.py

import gradio as gr
import torch
from safetensors.torch import load_file
from pathlib import Path
import json
from functools import lru_cache

# Import utilities
from models.model_loader import load_embed_model, load_rerank_model
from utils.search import search_embeddings, rerank_results
from utils.visualization import plot_results

# ============================================================================
# Data Loading Functions
# ============================================================================

@lru_cache(maxsize=1)
def load_metadata(metadata_path: str = "data/recipes/metadata.json") -> dict:
    """Load metadata.json (cached)."""
    with open(metadata_path, "r", encoding="utf-8") as f:
        return json.load(f)


def get_recipes_and_paths(metadata_path: str = "data/recipes/metadata.json"):
    """Load recipes, image paths, and full data."""
    print("📋 Loading metadata...")
    metadata = load_metadata(metadata_path)
    recipes = metadata["recipes"]
    
    all_recipes_markdown = [r["markdown"] for r in recipes]
    all_image_paths = [r["image_path"] for r in recipes]
    
    print(f"✅ Loaded {len(recipes)} recipes")
    return all_recipes_markdown, all_image_paths, recipes


# ============================================================================
# Load Models & Data (once at startup)
# ============================================================================

print("🔄 Loading models...")
embed_model, embed_device = load_embed_model()
rerank_model, rerank_processor, rerank_device = load_rerank_model()

print("📦 Loading embeddings...")

# Load Image embeddings
print("   - Loading image embeddings...")
image_embeddings = load_file("data/embeddings/image_embedding_500.safetensors")["image_embeddings_500"]
print(f"     ✅ Image embeddings shape: {image_embeddings.shape}")

# Load Text embeddings
print("   - Loading text embeddings...")
text_embeddings = load_file("data/embeddings/text_embeddings_500.safetensors")["text_embeddings_500"]
print(f"     ✅ Text embeddings shape: {text_embeddings.shape}")

# Load Image+Text embeddings
print("   - Loading image+text embeddings...")
image_text_embeddings = load_file("data/embeddings/image_text_embeddings_500.safetensors")["image_text_embeddings"]
print(f"     ✅ Image+Text embeddings shape: {image_text_embeddings.shape}")

print("📋 Loading recipe data...")
all_recipes, all_image_paths, full_recipes = get_recipes_and_paths()

print("\n✅ System ready!")
print(f"   📚 Recipes loaded: {len(all_recipes)}")
print(f"   🖼️  Images loaded: {len(all_image_paths)}")
print(f"   📊 Embeddings:")
print(f"      - Image: {image_embeddings.shape}")
print(f"      - Text: {text_embeddings.shape}")
print(f"      - Image+Text: {image_text_embeddings.shape}")

# ============================================================================
# Gradio Interface
# ============================================================================

def search_and_display(
    query: str,
    modality: str,
    use_rerank: bool,
    num_results: int = 3
):
    """Main search function."""
    
    if not query.strip():
        return None, {"error": "Please enter a search query"}
    
    # Select embeddings and documents based on modality
    if modality == "Image":
        embeddings = image_embeddings
        documents = [""] * len(all_image_paths)  # Empty for image-only
        search_modality = "image"
    elif modality == "Text":
        embeddings = text_embeddings
        documents = all_recipes
        search_modality = "text"
    else:  # Image+Text
        embeddings = image_text_embeddings
        documents = all_recipes
        search_modality = "image_text"
    
    print(f"\n🔍 Searching with modality: {modality}")
    print(f"   Embeddings shape: {embeddings.shape}")
    print(f"   Query: {query[:50]}...")
    
    # Initial search
    results = search_embeddings(
        query=query,
        query_image=None,
        model=embed_model,
        embeddings=embeddings,
        documents=documents,
        image_paths=all_image_paths,
        top_k=20,
        modality=search_modality
    )
    
    print(f"   ✅ Found {len(results)} results")
    
    # Optional reranking (only for image/image_text modalities)
    if use_rerank and modality != "Text":
        print(f"   🎯 Reranking top {min(10, len(results))} results...")
        results = rerank_results(
            query=query,
            results=results,
            rerank_model=rerank_model,
            rerank_processor=rerank_processor,
            device=rerank_device,
            top_k=min(10, len(results))
        )
        print(f"   ✅ Reranking complete")
    
    # Add full recipe details to results
    for i, result in enumerate(results[:num_results]):
        # Find matching recipe index
        img_path = result.get("image_path")
        if img_path:
            try:
                idx = all_image_paths.index(img_path)
                result["recipe_details"] = {
                    "name": full_recipes[idx]["name"],
                    "description": full_recipes[idx]["description"][:200] + "...",
                    "tags": full_recipes[idx]["tags"][:5],
                    "ingredients_count": len(full_recipes[idx]["ingredients"]),
                    "steps_count": len(full_recipes[idx]["steps"]),
                }
            except (ValueError, IndexError):
                pass
    
    # Visualize
    if modality != "Text":
        output_img = plot_results(results, query, num_images=num_results)
        return output_img, results[:num_results]
    else:
        # For text-only, return formatted text
        text_output = "\n\n".join([
            f"**Rank {r['rank']}** (Score: {r['score']:.4f})\n{r.get('text', '')[:500]}..."
            for r in results[:num_results]
        ])
        return None, results[:num_results]


# ============================================================================
# Gradio UI
# ============================================================================

with gr.Blocks(title="🍳 Multimodal Recipe Search", theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        """
        # 🍳 Multimodal Recipe RAG System
        Search **500 recipes** using text queries across images and documents.
        
        **Three Search Modes:**
        - 🖼️ **Image**: Visual similarity search (find similar-looking dishes)
        - 📝 **Text**: Semantic text search (find by ingredients, instructions, reviews)
        - 🎨 **Image+Text**: Combined multimodal search (best of both worlds)
        
        Powered by **NVIDIA Nemotron Embed-VL** and **Rerank-VL** models.
        """
    )
    
    with gr.Row():
        with gr.Column(scale=1):
            query_input = gr.Textbox(
                label="🔍 Search Query",
                placeholder="e.g., 'chocolate cake recipe', 'healthy breakfast', 'pasta with tomatoes'",
                lines=2
            )
            
            modality_radio = gr.Radio(
                choices=["Image", "Text", "Image+Text"],
                value="Image+Text",  # ✅ Default to Image+Text
                label="📊 Search Modality",
                info="Choose how to search: visual, text, or combined"
            )
            
            rerank_check = gr.Checkbox(
                label="🎯 Use Reranking (Cross-Encoder)",
                value=True,
                info="Rerank top results for better accuracy (adds ~1-2s)"
            )
            
            num_results_slider = gr.Slider(
                minimum=1,
                maximum=5,
                value=3,
                step=1,
                label="📈 Number of Results to Display"
            )
            
            search_btn = gr.Button("🚀 Search", variant="primary", size="lg")
            
            # Add info box
            gr.Markdown(
                """
                **💡 Tips:**
                - Use **Image** for visual similarity (e.g., "desserts with chocolate")
                - Use **Text** for ingredient/instruction search (e.g., "vegetarian pasta")
                - Use **Image+Text** for best overall results
                """
            )
        
        with gr.Column(scale=2):
            output_image = gr.Image(label="🖼️ Top Recipe Results", type="pil")
            output_json = gr.JSON(label="📋 Detailed Results")
    
    # Examples for each modality
    with gr.Accordion("💡 Example Queries", open=False):
        gr.Examples(
            examples=[
                # Image searches
                ["recipes with steak", "Image", True, 3],
                ["chocolate desserts", "Image", True, 3],
                ["colorful salads", "Image", False, 4],
                
                # Text searches
                ["healthy breakfast ideas", "Text", False, 5],
                ["vegetarian meals with pasta", "Text", False, 3],
                ["quick dinner under 30 minutes", "Text", False, 4],
                
                # Image+Text searches (best results)
                ["creamy pasta dishes", "Image+Text", True, 3],
                ["spicy chicken recipes", "Image+Text", True, 3],
                ["fresh summer salads", "Image+Text", True, 4],
            ],
            inputs=[query_input, modality_radio, rerank_check, num_results_slider],
            label="Try these examples"
        )
    
    # Footer
    gr.Markdown(
        """
        ---
        ### 🔧 Technical Details
        
        **Models:**
        - 🔍 Embedding: [nvidia/llama-nemotron-embed-vl-1b-v2](https://huggingface.co/nvidia/llama-nemotron-embed-vl-1b-v2)
        - 🎯 Reranking: [nvidia/llama-nemotron-rerank-vl-1b-v2](https://huggingface.co/nvidia/llama-nemotron-rerank-vl-1b-v2)
        
        **Dataset:** [TurkishCodeMan/recipe-synthetic-images-10k](https://huggingface.co/datasets/TurkishCodeMan/recipe-synthetic-images-10k) (500 recipes)
        
        **Embeddings:**
        - Image: 2048-dim visual embeddings
        - Text: 2048-dim semantic embeddings  
        - Image+Text: 2048-dim multimodal embeddings
        """
    )
    
    # Connect button
    search_btn.click(
        fn=search_and_display,
        inputs=[query_input, modality_radio, rerank_check, num_results_slider],
        outputs=[output_image, output_json]
    )

# ============================================================================
# Launch
# ============================================================================

if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False,
        show_error=True
    )