app.py

import gradio as gr
import numpy as np
import h5py
import faiss
import json
from sentence_transformers import SentenceTransformer

def load_data():
    with h5py.File('patent_embeddings.h5', 'r') as f:
        embeddings = f['embeddings'][:]
        patent_numbers = f['patent_numbers'][:]
    
    metadata = {}
    with open('patent_metadata.jsonl', 'r') as f:
        for line in f:
            data = json.loads(line)
            metadata[data['patent_number']] = data
    
    print(f"Embedding shape: {embeddings.shape}")
    return embeddings, patent_numbers, metadata

embeddings, patent_numbers, metadata = load_data()

# Create FAISS index
index = faiss.IndexFlatL2(embeddings.shape[1])
index.add(embeddings)

# Load BERT model for encoding search queries
embedding_dim = embeddings.shape[1]
print(f"Embedding dimension: {embedding_dim}")

if embedding_dim == 384:
    model = SentenceTransformer('all-MiniLM-L6-v2')
elif embedding_dim == 768:
    model = SentenceTransformer('all-mpnet-base-v2')
else:
    print(f"Unexpected embedding dimension: {embedding_dim}")
    model = SentenceTransformer('all-MiniLM-L6-v2')  # Default to this model

def search(query, top_k=5):
    # Encode the query
    query_embedding = model.encode([query])[0]
    
    # Ensure the query embedding has the same dimension as the index
    if query_embedding.shape[0] != index.d:
        print(f"Query embedding dimension ({query_embedding.shape[0]}) does not match index dimension ({index.d})")
        # Option 1: Pad or truncate the query embedding
        if query_embedding.shape[0] < index.d:
            query_embedding = np.pad(query_embedding, (0, index.d - query_embedding.shape[0]))
        else:
            query_embedding = query_embedding[:index.d]
    
    # Perform similarity search
    distances, indices = index.search(np.array([query_embedding]), top_k)
    
    results = []
    for i, idx in enumerate(indices[0]):
        patent_number = patent_numbers[idx]
        patent_data = metadata[patent_number]
        result = f"Patent Number: {patent_number}\n"
        result += f"Abstract: {patent_data['abstract'][:200]}...\n"
        result += f"Similarity Score: {1 - distances[0][i]:.4f}\n\n"
        results.append(result)
    
    return "\n".join(results)

# Create Gradio interface
iface = gr.Interface(
    fn=search,
    inputs=gr.Textbox(lines=2, placeholder="Enter your search query here..."),
    outputs=gr.Textbox(lines=10, label="Search Results"),
    title="Patent Similarity Search",
    description="Enter a query to find similar patents based on their embeddings."
)

if __name__ == "__main__":
    iface.launch(share=True)