app.py

import streamlit as st

import numpy as np

import pickle
from collections import OrderedDict

from sentence_transformers import SentenceTransformer, CrossEncoder, util
import torch

import requests
import io


import gdown



device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

import pandas as pd
st.title('Unofficial ASA 2022 Program Search')
st.write(''' * Retrieves the twenty most relevant talks to your seach phrase.
* The first search can take up to 30 seconds as the files load. After that, it's quicker to respond.
* Behind the scenes, the semantic search uses [text embeddings](https://www.sbert.net) with a [retrieve & re-rank](https://colab.research.google.com/github/UKPLab/sentence-transformers/blob/master/examples/applications/retrieve_rerank/retrieve_rerank_simple_wikipedia.ipynb) process to find the best matches.
* Let [me](mailto:neal.caren@unc.edu) know what you think or if it looks broken.
''')


def sent_trans_load():
    #We use the Bi-Encoder to encode all passages, so that we can use it with sematic search
    bi_encoder = SentenceTransformer('multi-qa-MiniLM-L6-cos-v1')
    bi_encoder.max_seq_length = 256     #Truncate long passages to 256 tokens, max 512
    return bi_encoder

def sent_cross_load():
    #We use the Bi-Encoder to encode all passages, so that we can use it with sematic search
    cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
    return cross_encoder


@st.cache
def load_data():
    #df = pd.read_json('https://www.dropbox.com/s/82lwbaym3b1o6uq/passages.jsonl?raw=1', lines=True)

    output = "asa_talks.jsonl"
    gdown.download(id='1-028z9eUkceUonK9YSb-ICv5ZgA3y0-K', output=output, quiet=False)

    df = pd.read_json(output, lines=True)

    df.reset_index(inplace=True, drop=True)
    return df


with st.spinner(text="Loading data..."):
    df = load_data()
    passages = df['text'].values

@st.cache
def load_embeddings():
    #efs = [np.load(f'embeddings_{i}.pt.npy')  for i in range(0,5)]
    #corpus_embeddings = np.concatenate(efs)

    output = "embeddings.npy"
    gdown.download(id='112Z5t9bVHbbZxlx0R7MKdBy_VLR-pEsO', output=output, quiet=False)

    corpus_embeddings = np.load(output)
    #response = requests.get("https://www.dropbox.com/s/px8kjdd3p5mzw6j/corpus_embeddings.pt.npy?raw=1")
    #corpus_embeddings = np.load(io.BytesIO(response.content))


    return corpus_embeddings

with st.spinner(text="Loading embeddings..."):
    corpus_embeddings = load_embeddings()





def search(query, top_k=40):

    ##### Sematic Search #####
    # Encode the query using the bi-encoder and find potentially relevant passages
    question_embedding = bi_encoder.encode(query, convert_to_tensor=True).to(device)


    hits = util.semantic_search(question_embedding, corpus_embeddings, top_k=top_k)
    hits = hits[0]  # Get the hits for the first query
    ##### Re-Ranking #####
    # Now, score all retrieved passages with the cross_encoder
    cross_inp = [[query, passages[hit['corpus_id']]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp)

    # Sort results by the cross-encoder scores
    for idx in range(len(cross_scores)):
        hits[idx]['cross-score'] = cross_scores[idx]

    # Output of top-5 hits from re-ranker
    print("\n-------------------------\n")
    print("Search Results")
    hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)

    hd = OrderedDict()
    for hit in hits[0:20]:

        row_id = hit['corpus_id']

        title = df.loc[row_id]['title']
        panel = df.loc[row_id]['panel'].split(' - ')[-1]
        details = df.loc[row_id]['details']
        author = df.loc[row_id]['author']
        abstract = df.loc[row_id]['abstract']
        session_id = df.loc[row_id]['session_id']
        paper_id = df.loc[row_id]['paper_id']

        session_url = f'https://convention2.allacademic.com/one/asa/asa22/index.php?program_focus=view_session&selected_session_id={session_id}&cmd=online_program_direct_link&sub_action=online_program'
        paper_url =  f'https://convention2.allacademic.com/one/asa/asa22/index.php?program_focus=view_paper&selected_paper_id={paper_id}&cmd=online_program_direct_link&sub_action=online_program'

        st.markdown(f'## [{title}]({paper_url})')
        st.markdown(f'Panel: [{panel}]({session_url})')
        st.markdown(details)
        st.markdown(author)
        for graph in abstract.splitlines():
            st.markdown(f'> {graph}')
        st.write('')



search_query = st.text_input('Enter your search phrase:')
if search_query!='':
    with st.spinner(text="Searching and sorting results."):
        bi_encoder = sent_trans_load()
        cross_encoder = sent_cross_load()
        search(search_query)