app.py

import streamlit as st
from transformers import pipeline
import requests
from bs4 import BeautifulSoup
import nltk
import string
from streamlit.components.v1 import html
from sentence_transformers.cross_encoder import CrossEncoder as CE
import re
from typing import List, Tuple
import torch

SCITE_API_KEY = st.secrets["SCITE_API_KEY"]

class CrossEncoder:
    def __init__(self, model_path: str, **kwargs):
        self.model = CE(model_path, **kwargs)

    def predict(self, sentences: List[Tuple[str,str]], batch_size: int = 32, show_progress_bar: bool = True) -> List[float]:
        return self.model.predict(
            sentences=sentences,
            batch_size=batch_size,
            show_progress_bar=show_progress_bar)


def remove_html(x):
    soup = BeautifulSoup(x, 'html.parser')
    text = soup.get_text()
    return text.strip()


# 4 searches: strict y/n, supported y/n
# deduplicate
# search per query

# options are abstract search
# all search


def search(term, limit=10, clean=True, strict=True, all_mode=True, abstracts=True, abstract_only=False):
    term = clean_query(term, clean=clean, strict=strict)
    # heuristic, 2 searches strict and not? and then merge?
    # https://api.scite.ai/search?mode=all&term=unit%20testing%20software&limit=10&date_from=2000&date_to=2022&offset=0&supporting_from=1&contrasting_from=0&contrasting_to=0&user_slug=domenic-rosati-keW5&compute_aggregations=true
    contexts, docs = [], []
    if not abstract_only:
        mode = 'all'
        if not all_mode:
            mode = 'citations'
        search = f"https://api.scite.ai/search?mode={mode}&term={term}&limit={limit}&offset=0&user_slug=domenic-rosati-keW5&compute_aggregations=false"
        req = requests.get(
            search,
            headers={
                'Authorization': f'Bearer {SCITE_API_KEY}'
            }
        )
        try:
            req.json()
        except:
            pass

        contexts += [remove_html('\n'.join([cite['snippet'] for cite in doc['citations'] if cite['lang'] == 'en'])) for doc in req.json()['hits']]
        docs += [(doc['doi'], doc['citations'], doc['title'], doc['abstract'] or '')
            for doc in req.json()['hits']]

    if abstracts or abstract_only:
        search = f"https://api.scite.ai/search?mode=papers&abstract={term}&limit={limit}&offset=0&user_slug=domenic-rosati-keW5&compute_aggregations=false"
        req = requests.get(
            search,
            headers={
                'Authorization': f'Bearer {SCITE_API_KEY}'
            }
        )
        try:
            req.json()
            contexts += [remove_html(doc['abstract'] or '') for doc in req.json()['hits']]
            docs += [(doc['doi'], doc['citations'], doc['title'], doc['abstract'] or '')
                        for doc in req.json()['hits']]
        except:
            pass


    return (
        contexts,
        docs
    )


def find_source(text, docs, matched):
    for doc in docs:
        for snippet in doc[1]:
            if text in remove_html(snippet.get('snippet', '')):
                if matched and remove_html(snippet.get('snippet', '')).strip() != matched.strip():
                    continue
                new_text = text
                for sent in nltk.sent_tokenize(remove_html(snippet.get('snippet', ''))):
                    if text in sent:
                        new_text = sent
                return {
                    'citation_statement': snippet['snippet'].replace('<strong class="highlight">', '').replace('</strong>', ''),
                    'text': new_text,
                    'from': snippet['source'],
                    'supporting': snippet['target'],
                    'source_title': remove_html(doc[2] or ''),
                    'source_link': f"https://scite.ai/reports/{doc[0]}"
                }
        if text in remove_html(doc[3]):
            if matched and remove_html(doc[3]).strip() != matched.strip():
                    continue
            new_text = text
            sent_loc = None
            sents = nltk.sent_tokenize(remove_html(doc[3]))
            for i, sent in enumerate(sents):
                if text in sent:
                    new_text = sent
                    sent_loc = i

            context = remove_html(doc[3]).replace('<strong class="highlight">', '').replace('</strong>', '')
            if sent_loc:
                context_len = 2
                sent_beg = sent_loc - context_len
                if sent_beg <= 0: sent_beg = 0
                sent_end = sent_loc + context_len
                if sent_end >= len(sents):
                    sent_end = len(sents)
                context = ''.join(sents[sent_beg:sent_end])

            return {
                    'citation_statement': context,
                    'text': new_text,
                    'from': doc[0],
                    'supporting': doc[0],
                    'source_title': remove_html(doc[2] or ''),
                    'source_link': f"https://scite.ai/reports/{doc[0]}"
                }
    return None


@st.experimental_singleton
def init_models():
    nltk.download('stopwords')
    nltk.download('punkt')
    from nltk.corpus import stopwords
    stop = set(stopwords.words('english') + list(string.punctuation))
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    question_answerer = pipeline(
        "question-answering", model='sultan/BioM-ELECTRA-Large-SQuAD2-BioASQ8B',
        device=device
    )
    reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2', device=device)
    # queryexp_tokenizer = AutoTokenizer.from_pretrained("doc2query/all-with_prefix-t5-base-v1")
    # queryexp_model = AutoModelWithLMHead.from_pretrained("doc2query/all-with_prefix-t5-base-v1")
    return question_answerer, reranker, stop, device # uqeryexp_model, queryexp_tokenizer

qa_model, reranker, stop, device = init_models() # queryexp_model, queryexp_tokenizer


def clean_query(query, strict=True, clean=True):
    operator = ' '
    if strict:
        operator = ' AND '
    query = operator.join(
        [i for i in query.lower().split(' ') if clean and i not in stop])
    if clean:
        query = query.translate(str.maketrans('', '', string.punctuation))
    return query


def card(title, context, score, link, supporting):
    st.markdown(f"""
    <div class="container-fluid">
        <div class="row align-items-start">
             <div  class="col-md-12 col-sm-12">
                 <br>
                 <span>
                     {context}
                     [<b>Confidence: </b>{score}%]
                 </span>
                 <br>
                 <b>From <a href="{link}">{title}</a></b>
             </div>
        </div>
     </div>
        """, unsafe_allow_html=True)
    html(f"""
    <div
                    class="scite-badge"
                    data-doi="{supporting}"
                    data-layout="horizontal"
                    data-show-zero="false"
                    data-show-labels="false"
                    data-tally-show="true"
                />
    <script
    async
    type="application/javascript"
    src="https://cdn.scite.ai/badge/scite-badge-latest.min.js">
  </script>
  """, width=None, height=42, scrolling=False)


st.title("Scientific Question Answering with Citations")

st.write("""
Ask a scientific question and get an answer drawn from [scite.ai](https://scite.ai) corpus of over 1.1bn citation statements.
Answers are linked to source documents containing citations where users can explore further evidence from scientific literature for the answer.

For example try: Do tanning beds cause cancer?
""")

st.markdown("""
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bootstrap@4.0.0/dist/css/bootstrap.min.css" integrity="sha384-Gn5384xqQ1aoWXA+058RXPxPg6fy4IWvTNh0E263XmFcJlSAwiGgFAW/dAiS6JXm" crossorigin="anonymous">
""", unsafe_allow_html=True)

with st.expander("Settings (strictness, context limit, top hits)"):
    support_all = st.radio(
        "Use abstracts and titles as a ranking signal (if the words are matched in the abstract then the document is more relevant)?",
        ('yes', 'no'))
    support_abstracts = st.radio(
        "Use abstracts as a source document?",
        ('yes', 'no', 'abstract only'))
    strict_lenient_mix = st.radio(
        "Type of strict+lenient combination: Fallback or Mix? If fallback, strict is run first then if the results are less than context_lim we also search lenient. Mix will search them both and let reranking sort em out",
        ('fallback', 'mix'))
    confidence_threshold = st.slider('Confidence threshold for answering questions? This number represents how confident the model should be in the answers it gives. The number is out of 100%', 0, 100, 1)
    use_reranking = st.radio(
        "Use Reranking? Reranking will rerank the top hits using semantic similarity of document and query.",
        ('yes', 'no'))
    top_hits_limit = st.slider('Top hits? How many documents to use for reranking. Larger is slower but higher quality', 10, 300, 50)
    context_lim = st.slider('Context limit? How many documents to use for answering from. Larger is slower but higher quality', 10, 300, 10)

# def paraphrase(text, max_length=128):
#     input_ids = queryexp_tokenizer.encode(text, return_tensors="pt", add_special_tokens=True)
#     generated_ids = queryexp_model.generate(input_ids=input_ids, num_return_sequences=suggested_queries or 5, num_beams=suggested_queries or 5, max_length=max_length)
#     queries = set([queryexp_tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=True) for g in generated_ids])
#     preds = '\n * '.join(queries)
#     return preds


def group_results_by_context(results):
    result_groups = {}
    for result in results:
        if result['context'] not in result_groups:
            result_groups[result['context']] = result
            result_groups[result['context']]['texts'] = []

        result_groups[result['context']]['texts'].append(
            result['answer']
        )
        if result['score'] > result_groups[result['context']]['score']:
            result_groups[result['context']]['score'] = result['score']
    return list(result_groups.values())


def matched_context(start_i, end_i, contexts_string, seperator='---'):
    # find seperators to identify start and end
    doc_starts = [0]
    for match in re.finditer(seperator, contexts_string):
        doc_starts.append(match.end())

    for i in range(len(doc_starts)):
        if i == len(doc_starts) - 1:
            if start_i >= doc_starts[i]:
                return contexts_string[doc_starts[i]:len(contexts_string)].replace(seperator, '')

        if start_i >= doc_starts[i] and end_i <= doc_starts[i+1]:
            return contexts_string[doc_starts[i]:doc_starts[i+1]].replace(seperator, '')
    return None


def run_query(query):
#     if use_query_exp == 'yes':
#         query_exp = paraphrase(f"question2question: {query}")
#         st.markdown(f"""
# If you are not getting good results try one of:
# * {query_exp}
# """)

    # address period in highlitht avoidability. Risk factors
    # address poor tokenization Deletions involving chromosome region 4p16.3 cause WolfHirschhorn syndrome (WHS, OMIM 194190) [Battaglia et al, 2001].
    # address highlight html

    # could also try fallback if there are no good answers by score...
    limit = top_hits_limit or 100
    context_limit = context_lim or 10
    contexts_strict, orig_docs_strict = search(query, limit=limit, strict=True, all_mode=support_all == 'yes', abstracts= support_abstracts == 'yes', abstract_only=support_abstracts == 'abstract only')
    if strict_lenient_mix == 'fallback' and len(contexts_strict) < context_limit:
        contexts_lenient, orig_docs_lenient = search(query, limit=limit, strict=False, all_mode=support_all == 'yes',  abstracts= support_abstracts == 'yes', abstract_only= support_abstracts == 'abstract only')
        contexts = list(
            set(contexts_strict + contexts_lenient)
        )
        orig_docs = orig_docs_strict + orig_docs_lenient
    elif strict_lenient_mix == 'mix':
        contexts_lenient, orig_docs_lenient = search(query, limit=limit, strict=False)
        contexts = list(
            set(contexts_strict + contexts_lenient)
        )
        orig_docs = orig_docs_strict + orig_docs_lenient
    else:
        contexts = list(
            set(contexts_strict)
        )
        orig_docs = orig_docs_strict


    if len(contexts) == 0 or not ''.join(contexts).strip():
        return st.markdown("""
        <div class="container-fluid">
        <div class="row align-items-start">
             <div  class="col-md-12 col-sm-12">
            Sorry... no results for that question! Try another...
         </div>
        </div>
     </div>
        """, unsafe_allow_html=True)

    if use_reranking == 'yes':
        sentence_pairs = [[query, context] for context in contexts]
        scores = reranker.predict(sentence_pairs, batch_size=len(sentence_pairs), show_progress_bar=False)
        hits = {contexts[idx]: scores[idx] for idx in range(len(scores))}
        sorted_contexts = [k for k,v in sorted(hits.items(), key=lambda x: x[0], reverse=True)]
        context = '\n---'.join(sorted_contexts[:context_limit])
    else:
        context = '\n---'.join(contexts[:context_limit])

    results = []
    model_results = qa_model(question=query, context=context, top_k=10)
    for result in model_results:
        matched = matched_context(result['start'], result['end'], context)
        support = find_source(result['answer'], orig_docs, matched)
        if not support:
            continue
        results.append({
            "answer": support['text'],
            "title": support['source_title'],
            "link": support['source_link'],
            "context": support['citation_statement'],
            "score": result['score'],
            "doi": support["supporting"]
        })

    grouped_results = group_results_by_context(results)
    sorted_result = sorted(grouped_results, key=lambda x: x['score'], reverse=True)

    if confidence_threshold == 0:
        threshold = 0
    else:
        threshold = (confidence_threshold or 10) / 100

    sorted_result = filter(
        lambda x: x['score'] > threshold,
        sorted_result
    )

    for r in sorted_result:
        ctx = remove_html(r["context"])
        for answer in r['texts']:
            ctx = ctx.replace(answer.strip(), f"<mark>{answer.strip()}</mark>")
        # .replace( '<cite', '<a').replace('</cite', '</a').replace('data-doi="', 'href="https://scite.ai/reports/')
        title = r.get("title", '')
        score = round(round(r["score"], 4) * 100, 2)
        card(title, ctx, score, r['link'], r['doi'])

query = st.text_input("Ask scientific literature a question", "")
if query != "":
    with st.spinner('Loading...'):
        run_query(query)