search.py

from transformers import BertTokenizer, BertModel
import torch
import pandas as pd
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np

# Load embeddings DataFrame
df = pd.read_parquet('embeddings.parquet')
df = df.head(5)

# Initialize tokenizer and model
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = BertModel.from_pretrained('bert-base-uncased')

def search_embeddings(query):
    # Prepare the query text
    inputs = tokenizer(query_text, return_tensors="pt", padding=True, truncation=True, max_length=512)

    # Generate embeddings for the query text
    with torch.no_grad():
        outputs = model(**inputs)
        query_vector = outputs.last_hidden_state.mean(dim=1).squeeze().numpy()

    # Assuming 'embedding' column in df contains embeddings as lists or arrays
    # Convert list of embeddings to numpy array for cosine similarity calculation

    embedding_matrix = np.stack(df['embedding'].values)
    # Compute cosine similarities
    similarities = cosine_similarity([query_vector], embedding_matrix)

    # Get the top 5 most similar entries
    top_indices = np.argsort(similarities[0])[::-1][:5]
    top_scores = similarities[0][top_indices]

    results = ""

    # Print top matches with their scores
    for index, score in zip(top_indices, top_scores):
        # print(f"Index: {index}, Score: {score}, Data: {df.iloc[index]}")
        data = df.iloc[index]
        results += (f"Question: {data['text']} Answer: {data['answer']} ")
    return results


# query_text = "Paul's First Epistle to the Corinthians"
# print(search_embeddings(query_text))