app.py

import gradio as gr

import pandas as pd
import numpy as np

import torch.nn.functional as F

from torch import Tensor
from transformers import AutoTokenizer, AutoModel
from sklearn.metrics.pairwise import cosine_similarity

import re


def average_pool(last_hidden_states: Tensor,
                 attention_mask: Tensor) -> Tensor:
    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]


df = pd.read_csv('wiki.csv')
data_embeddings = np.load("wiki-embeddings.npy")

print("loading the model...")
tokenizer = AutoTokenizer.from_pretrained('intfloat/multilingual-e5-large')
model = AutoModel.from_pretrained('intfloat/multilingual-e5-large')

with gr.Blocks() as demo:
    chatbot = gr.Chatbot(label="semantic search for 230k+ wikipedia articles")
    msg = gr.Textbox(label="simple wikipedia semantic search query", placeholder="for example, \"medieval battles\"")
    clear = gr.ClearButton([msg, chatbot])

    def _search(message, chat_history):
        batch_dict = tokenizer(["query: " + message], max_length=512, padding=True, truncation=True, return_tensors='pt')
    
        outputs = model(**batch_dict)
        input_embedding = average_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
    
        # normalize embeddings
        input_embedding = F.normalize(input_embedding, p=2, dim=1)
        input_embedding = input_embedding[0].tolist()
    
        # Compute cosine similarities
        input_embedding = np.array(input_embedding).reshape(1, -1)
        cos_similarities = cosine_similarity(data_embeddings, input_embedding).flatten()
    
        # Get top k similar points' indices
        k = 10  # replace with your value of k
        top_k_idx = cos_similarities.argsort()[-k:][::-1]
    
        # Get corresponding 'text' for top k similar points
        top_k_text = df['title'].iloc[top_k_idx].tolist()
        
        bot_message = "\n".join(f"{i+1}. {top_k_text[i]} // {top_k_idx[i]}" for i in range(len(top_k_text)))
    
        chat_history.append((message, f"results (you can enter article number 1-{k} to see its contents):\n" + bot_message))
        return "", chat_history

    def _retrieve(message, chat_history):
        idx = int(message)
        for _, m in chat_history[::-1]:
            if m.startswith("results"):
                for n in m.split("\n")[1:]:
                    print(n)
                    if str(idx) == n.split(".")[0]:
                        df_idx = int(n.split(" // ")[-1])
                        print(df_idx)
                        article = df.iloc[df_idx]['text']
                        article = re.sub(r'(===?=?[A-Z ].+?===?=?)', r'\n\n\1\n', article)
                        chat_history.append((message, f"contents of {n}:\n{article}"))
                        return "", chat_history
        print("nothing found")
        chat_history.append((message, "🤔 article not found"))
        return "", chat_history
    
    def respond(message, chat_history):
        print(f"received input '{message}'")
        try:
            int(message)
            print(f"retrieving #{message}")
            return _retrieve(message, chat_history)
        except ValueError:
            print(f"searching for {message}")
            return _search(message, chat_history)

    msg.submit(respond, [msg, chatbot], [msg, chatbot])

demo.launch()