app.py

import faiss
import numpy as np
import pandas as pd
from sentence_transformers import SentenceTransformer
import streamlit as st

import os

os.environ['KMP_DUPLICATE_LIB_OK']='True'
        
        
@st.cache_resource
def load_model():
    model = SentenceTransformer("sbintuitions/sarashina-embedding-v1-1b")
    
    return model


@st.cache_resource
def load_title_data():
    title_df = pd.read_csv('anlp2025.tsv', names=["pid", "title"], sep="\t")
    
    return title_df


@st.cache_resource
def load_title_embeddings():
    npz_comp = np.load("anlp2025.npz")
    title_embeddings = npz_comp["arr_0"]

    return title_embeddings


def get_retrieval_results(index, input_text, top_k, model, title_df):
    query_embeddings = model.encode([input_text])
    _, ids = index.search(x=query_embeddings, k=top_k)
    retrieved_titles = []
    retrieved_pids = []

    for id in ids[0]:
        retrieved_titles.append(title_df.loc[id, "title"])
        retrieved_pids.append(title_df.loc[id, "pid"])

    df = pd.DataFrame({"pids": retrieved_pids, "paper": retrieved_titles})
    
    return df
    

if __name__ == "__main__":
    model = load_model()
    title_df = load_title_data()
    title_embeddings = load_title_embeddings()

    index = faiss.IndexFlatL2(1792)
    index.add(title_embeddings)
    
    st.markdown("## NLP2025 論文検索")
    input_text = st.text_input('query', '', placeholder='')
    top_k = st.number_input('top_k', min_value=1, value=10, step=1)
    
    if st.button('検索'):
        stripped_input_text = input_text.strip()
        df = get_retrieval_results(index, stripped_input_text, top_k, model, title_df)
        st.table(df)