app.py

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


@st.experimental_memo
def load_model():
    return SentenceTransformer("TamedWicked/MathBERT_hr")


@st.experimental_memo
def load_knowledge_base_df():
    return pd.read_parquet("data/knowledge_base.parquet")


@st.experimental_memo
def load_knowledge_base_index():
    embeddings = joblib.load("data/knowledge_base_embeddings.pkl")
    index = faiss.IndexFlatL2(embeddings.shape[1])
    index.add(embeddings)
    return index


def vector_search(query: list, model: SentenceTransformer, index, num_results=10):
    vector = model.encode(list(query), show_progress_bar=False, convert_to_numpy=True)
    D, I = index.search(np.array(vector).astype("float32"), k=num_results)
    return D, I


def show_df_as_html(df: pd.DataFrame):
    return df.to_html()


def show_df_as_markdown(df: pd.DataFrame):
    return df.to_markdown()


model: SentenceTransformer = load_model()
df: pd.DataFrame = load_knowledge_base_df()
knowledge_index: np.array = load_knowledge_base_index()

query = st.text_input("Your math query:", value="Jesu li strukture koje su elementarno ekvivalentne izomorfne?")
if query:
    D, I = vector_search([query], model, knowledge_index, num_results=5)
    result = df[["Speech", "start_link"]].iloc[I[0]]
    result.index = list(range(1, len(result)+1))
    speeches = result["Speech"].tolist()
    links = result["start_link"].tolist()
    st.write(show_df_as_markdown(result))