app.py

import gradio as gr
import numpy as np
import h5py
import faiss
import json
import re
from collections import Counter
import torch
from nltk.corpus import stopwords
from nltk.tokenize import word_tokenize
import nltk
from sentence_transformers import SentenceTransformer
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity

# Download necessary NLTK data
nltk.download('stopwords', quiet=True)
nltk.download('punkt', quiet=True)

# Load SentenceTransformer model
model = SentenceTransformer('anferico/bert-for-patents')

def preprocess_query(text):
    # Remove "[EN]" label and claim numbers
    text = re.sub(r'\[EN\]\s*', '', text)
    text = re.sub(r'^\d+\.\s*', '', text, flags=re.MULTILINE)
    
    # Convert to lowercase while preserving acronyms and units
    words = text.split()
    text = ' '.join(word if word.isupper() or re.match(r'^\d+(\.\d+)?[a-zA-Z]+$', word) else word.lower() for word in words)
    
    # Remove special characters except hyphens and periods in numbers
    text = re.sub(r'[^\w\s\-.]', ' ', text)
    text = re.sub(r'(?<!\d)\.(?!\d)', ' ', text)  # Remove periods not in numbers
    
    # Normalize spaces
    text = re.sub(r'\s+', ' ', text).strip()
    
    # Tokenize
    tokens = word_tokenize(text)
    
    # Remove stopwords
    stop_words = set(stopwords.words('english'))
    tokens = [word for word in tokens if word.lower() not in stop_words]
    
    # Join tokens back into text
    text = ' '.join(tokens)
    
    # Preserve numerical values with units
    text = re.sub(r'(\d+(\.\d+)?)([a-zA-Z]+)', r'\1_\3', text)
    
    # Handle ranges and measurements
    text = re.sub(r'(\d+(\.\d+)?)(\s*to\s*)(\d+(\.\d+)?)(\s*[a-zA-Z]+)', r'\1_to_\4_\6', text)
    text = re.sub(r'between\s*(\d+(\.\d+)?)(\s*and\s*)(\d+(\.\d+)?)\s*([a-zA-Z]+)', r'between_\1_and_\4_\5', text)
    
    # Preserve chemical formulas
    text = re.sub(r'\b([A-Z][a-z]?\d*)+\b', lambda m: m.group().replace(' ', ''), text)
    
    return text

def extract_key_features(text):
    # For queries, we'll just preprocess and return all non-stopword terms
    processed_text = preprocess_query(text)
    # Split the processed text into individual terms
    features = processed_text.split()
    # Remove duplicates while preserving order
    features = list(dict.fromkeys(features))
    return features

def encode_texts(texts):
    embeddings = model.encode(texts, show_progress_bar=True)
    return embeddings

def load_data():
    try:
        with h5py.File('patent_embeddings.h5', 'r') as f:
            embeddings = f['embeddings'][:]
            patent_numbers = f['patent_numbers'][:]

        metadata = {}
        texts = []
        with open('patent_metadata.jsonl', 'r') as f:
            for line in f:
                data = json.loads(line)
                metadata[data['patent_number']] = data
                texts.append(data['text'])

        print(f"Embedding shape: {embeddings.shape}")
        print(f"Number of patent numbers: {len(patent_numbers)}")
        print(f"Number of metadata entries: {len(metadata)}")
        return embeddings, patent_numbers, metadata, texts
    except Exception as e:
        print(f"An error occurred while loading data: {e}")
        raise

def compare_features(query_features, patent_features):
    common_features = set(query_features) & set(patent_features)
    similarity_score = len(common_features) / max(len(query_features), len(patent_features))
    return common_features, similarity_score

def hybrid_search(query, top_k=5):
    print(f"Original query: {query}")
    processed_query = preprocess_query(query)
    query_features = extract_key_features(processed_query)

    # Encode the processed query using the SentenceTransformer model
    query_embedding = encode_texts([processed_query])[0]
    query_embedding = query_embedding / np.linalg.norm(query_embedding)

    # Perform semantic similarity search
    semantic_distances, semantic_indices = index.search(np.array([query_embedding]).astype('float32'), top_k * 2)

    # Perform TF-IDF based search
    query_tfidf = tfidf_vectorizer.transform([processed_query])
    tfidf_similarities = cosine_similarity(query_tfidf, tfidf_matrix).flatten()
    tfidf_indices = tfidf_similarities.argsort()[-top_k * 2:][::-1]

    # Combine and rank results
    combined_results = {}
    for i, idx in enumerate(semantic_indices[0]):
        patent_number = patent_numbers[idx].decode('utf-8')
        text = metadata[patent_number]['text']
        patent_features = extract_key_features(text)
        common_features, feature_similarity = compare_features(query_features, patent_features)
        combined_results[patent_number] = {
            'score': semantic_distances[0][i] * 1.0 + tfidf_similarities[idx] * 0.5 + feature_similarity,
            'common_features': common_features,
            'text': text
        }

    for idx in tfidf_indices:
        patent_number = patent_numbers[idx].decode('utf-8')
        if patent_number not in combined_results:
            text = metadata[patent_number]['text']
            patent_features = extract_key_features(text)
            common_features, feature_similarity = compare_features(query_features, patent_features)
            combined_results[patent_number] = {
                'score': tfidf_similarities[idx] * 1.0 + feature_similarity,
                'common_features': common_features,
                'text': text
            }

    # Sort and get top results
    top_results = sorted(combined_results.items(), key=lambda x: x[1]['score'], reverse=True)[:top_k]
    results = []
    for patent_number, data in top_results:
        result = f"Patent Number: {patent_number}\n"
        result += f"Text: {data['text'][:200]}...\n"
        result += f"Combined Score: {data['score']:.4f}\n"
        result += f"Common Key Features: {', '.join(data['common_features'])}\n\n"
        results.append(result)

    return "\n".join(results)

# Load data and prepare the FAISS index
embeddings, patent_numbers, metadata, texts = load_data()

# Normalize embeddings for cosine similarity
embeddings = embeddings / np.linalg.norm(embeddings, axis=1, keepdims=True)

# Create FAISS index for cosine similarity
index = faiss.IndexFlatIP(embeddings.shape[1])
index.add(embeddings)

# Create TF-IDF vectorizer
tfidf_vectorizer = TfidfVectorizer(stop_words='english')
tfidf_matrix = tfidf_vectorizer.fit_transform(texts)

# Create Gradio interface
iface = gr.Interface(
    fn=hybrid_search,
    inputs=[
        gr.Textbox(lines=2, placeholder="Enter your patent query here..."),
        gr.Slider(minimum=1, maximum=20, step=1, value=5, label="Top K Results"),
    ],
    outputs=gr.Textbox(lines=10, label="Search Results"),
    title="Patent Similarity Search",
    description="Enter a patent description to find similar patents based on key features."
)

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