app.py

import gradio as gr

from datetime import datetime
import concurrent.futures
from io import StringIO
import requests

from urllib.parse import urlparse
import time

import demo as du

endpoint_url = "https://d34hcsxnegbpcslxzqsmesvr7m0ljtuz.lambda-url.us-west-2.on.aws/"
executor = concurrent.futures.ThreadPoolExecutor(max_workers=10)

def output_credit_dist_nollm(msg, cur_idx, _out_credit, _out_claims):
    print(f'Start output_credit_dist at {datetime.now()}.')
    start_time = time.perf_counter()
    print(f'Target is ```{msg}```')
    _out_credit.truncate(0)
    _out_credit.seek(0)
    _out_claims.truncate(0)
    _out_claims.seek(0)

    print(du.style2_str, file=_out_credit)
    print(du.style2_str, file=_out_claims)

    atoms_l, atom_topkmatches_l, credit_l = [], [], []

    if len(msg) > 10:
        params = { "msg": msg, "max_workers": 10 }
        response = requests.get(endpoint_url, json=params).json()

        atoms_l = response['atoms_l']
        atom_topkmatches_l = response['atom_topkmatches_l']
        # atomidx_w_single_url_aggmatch_l = response['atomidx_w_single_url_aggmatch_l']
        atom_support_l = response['atom_support_l']
        credit_dist = response['credit_dist']

    if atoms_l:
        (
            url_to_cid_to_ctext_map,
            url_to_cid_to_ctext_formatted_map,
            url_to_cid_to_nquotes_map,
        ) = du.create_url_to_cid_to_ctext_formatted_map(atom_support_l)

        url_to_title = {}
        for atom_topkmatches in atom_topkmatches_l:
            for match in atom_topkmatches:
                url_to_title[match['metadata']['url']] = match['metadata']['title']

        credit_l = [(url, w) for url, w in credit_dist.items()]
        credit_l = sorted(credit_l, key=lambda x: x[1], reverse=True)
        du.print_w_time_elapsed('Computed credit_l', start_time)

    if not atom_topkmatches_l:
        print(f"<div>", file=_out_credit)
        print(f"<div class=\"doc-title\">No sources were found that are relevant this target.</div>", file=_out_credit)
        print(f"</div>", file=_out_credit)
        du.print_w_time_elapsed(f'End output_credit_dist (no matches) at {datetime.now()}', start_time)
        return '', _out_credit.getvalue()
    
    if not credit_l:
        print(f"<div>", file=_out_credit)
        print(f"<div class=\"doc-title\">No sources were found that strongly support this target.</div>", file=_out_credit)
        print(f"</div>", file=_out_credit)

    url_to_body_w_credit = {}
    futures = []
    for url, _ in credit_l:
        futures.append(executor.submit(du.get_article_from_url, url))
    for f in futures:
        article = f.result()
        url_to_body_w_credit[article['url']] = article['text']
    du.print_w_time_elapsed(f'Got url bodies', start_time)

    for url, w in credit_l:
        match_text = du.format_chunk_texts_for_display3(
            url,
            url_to_cid_to_ctext_map[url],
            url_to_cid_to_ctext_formatted_map[url],
            url_to_cid_to_nquotes_map[url],
        )
        print(f"<div>", file=_out_credit)
        favicon = f"<img src=\"https://www.google.com/s2/favicons?sz=128&amp;domain={urlparse(url).netloc}\"/>"
        print(f"<div class=\"doc-title\">{favicon}&nbsp&nbsp;{url_to_title[url]}<score>{100*w:.0f}%</score></div>", file=_out_credit)
        print(f"<div class=\"doc-url\"><a href=\"{url}\"  target=\"_blank\">{url}</a></div>", file=_out_credit)
        print(f"<div class=\"doc-text-wrapper\">", file=_out_credit)
        print(f"<div class=\"doc-text-left\">{url_to_body_w_credit[url]}</div>", file=_out_credit)
        print(f"<div class=\"doc-text-right\">{match_text}</div>", file=_out_credit)
        print(f"</div>", file=_out_credit)
        print(f"</div>", file=_out_credit)

    print(f"<div>", file=_out_claims)
    print(f"<div class=\"section-title\">Breakdown of article support for each extracted claim</div>", file=_out_claims)
    for j, atom_support in enumerate(atom_support_l):
        n_urls = len(atom_support.keys())
        n_support = sum([1 if determination['true'] else 0 for determination in atom_support.values()])

        print(f"<div class=\"claim-header\"><strong>Claim {j+1} ({n_support}/{n_urls}):</strong> \"{atoms_l[j]}\"</div>", file=_out_claims)
        
        for url, aggmatch_determination in atom_support.items():
            title = url_to_title[url]

            print(f"<div class=\"claim-doc-title\">{title}</div>", file=_out_claims)
            print(f"<div class=\"claim-doc-url\"><a href=\"{url}\"  target=\"_blank\">{url}</a></div>", file=_out_claims)

            print(f"<div class=\"claim-determination\"><strong>Determination:</strong> {'Supported' if aggmatch_determination['true'] else 'NOT supported'}.</div>", file=_out_claims)
            print(f"<div class=\"claim-text\"><strong>Rationale:</strong> {aggmatch_determination['rationale']}</div>", file=_out_claims)

            # if aggmatch_determination.get('quote_matches_l', None):
            #     for qid, qtext in enumerate(aggmatch_determination['quote_matches_l']):
            #         print(f"<div class=\"claim-text\"><strong>Quote {qid}:</strong> {qtext}</div>", file=_out_claims)

            use_formatted = 'chunk_text_formatted_l' in aggmatch_determination and aggmatch_determination['chunk_text_formatted_l']
            chunk_text_l_key = 'chunk_text_formatted_l' if use_formatted else 'chunk_text_l'

            for cid, ctext in zip(aggmatch_determination['id_l'], aggmatch_determination[chunk_text_l_key]):
                print(f"<div class=\"claim-text\"><strong>Chunk {cid}:</strong> {ctext}</div>", file=_out_claims)

    print(f"</div>", file=_out_claims)

    cur_idx[0] = 0
    du.print_w_time_elapsed(f'End output_credit_dist at {datetime.now()}', start_time)
    return 'Show claim breakdown', _out_credit.getvalue()

def toggle_output(cur_idx, _out_credit, _out_claims):
    if cur_idx[0] < 0:
        return ''
    cur_idx[0] += 1
    if cur_idx[0] % 2 == 0:
        return 'Show claim breakdown', _out_credit.getvalue()
    return 'Back to attribution', _out_claims.getvalue()

with gr.Blocks(theme=gr.themes.Default(text_size="lg")) as demo:
    _out_credit_var = gr.State(StringIO)
    _out_claims_var = gr.State(StringIO)
    cur_idx_var = gr.State([0])

    msg = gr.Textbox(label='Target')
    results_box = gr.HTML(label='Matches')
    toggle = gr.Button("")

    msg.submit(output_credit_dist_nollm, [msg, cur_idx_var, _out_credit_var, _out_claims_var], [toggle, results_box], queue=False)
    toggle.click(toggle_output, [cur_idx_var, _out_credit_var, _out_claims_var], [toggle, results_box], queue=False)

    results_box.change(None, scroll_to_output=True)
    
if __name__ == "__main__":
    demo.queue()
    demo.launch()