app.py

from functools import lru_cache
import time, aiohttp, asyncio, json, os, multiprocessing, torch, \
    requests, xmltodict, fitz, io
from minivectordb.embedding_model import EmbeddingModel
from minivectordb.vector_database import VectorDatabase
from text_util_en_pt.cleaner import structurize_text, detect_language, Language
import gradio as gr

torch.set_num_threads(2)

openrouter_key = os.environ.get("OPENROUTER_KEY")
model = EmbeddingModel(use_quantized_onnx_model=True)

def convert_xml_to_json(xml):
    return xmltodict.parse(xml)

def clean_title(title):
    title = title.replace('\n', ' ')
    while '  ' in title:
        title = title.replace('  ', ' ')
    return title

@lru_cache(maxsize=500)
def fetch_arxiv_links(query, max_results=5):
    url = f'http://export.arxiv.org/api/query?search_query=all:{query}&start=0&max_results={max_results}'
    response = requests.get(url)
    json_response = convert_xml_to_json(response.text)

    # Return a list of titles and links, and pdf links
    entries = []
    for entry in json_response['feed']['entry']:

        title = entry['title']
        id = entry['id'].split('/abs/')[-1]

        link = f'http://arxiv.org/abs/{id}'
        pdf_link = f'http://arxiv.org/pdf/{id}.pdf'

        entries.append({
            'title': clean_title(title),
            'link': link,
            'pdf_link': pdf_link
        })
    return entries

def download_pdf_from_link(link):
    # Download the file and hold it in memory
    response = requests.get(link)
    return io.BytesIO(response.content)

@lru_cache(maxsize=100)
def read_remote_pdf(pdf_metadata):
    pdf_metadata = json.loads(pdf_metadata)

    link = pdf_metadata['pdf_link']
    title = pdf_metadata['title']

    pdf_content = download_pdf_from_link(link)
    pdf_file = fitz.open("pdf", pdf_content.read())
    text_content = [page.get_text() for page in pdf_file]
    pdf_file.close()
    del pdf_file
    return {'title': title, 'text': '\n'.join(text_content)}

def fetch_data_from_pdfs(links):
    links = [ json.dumps(link) for link in links ]
    with multiprocessing.Pool(10) as pool:
        pdf_metadata = pool.map(read_remote_pdf, links)
    return pdf_metadata

def index_and_search(query, pdf_metadata):
    start = time.time()
    query_embedding = model.extract_embeddings(query)

    # Indexing
    vector_db = VectorDatabase()

    sentence_counter = 1

    for pdf_data in pdf_metadata:
        text = pdf_data['text']
        title = pdf_data['title']

        sentences = [ s['sentence'] for s in structurize_text(text)]

        for sentence in sentences:
            sentence_embedding = model.extract_embeddings(sentence)
            vector_db.store_embedding(
                sentence_counter,
                sentence_embedding,
                {
                    'sentence': sentence,
                    'title': title
                }
            )
            sentence_counter += 1
    
    embedding_time = time.time() - start

    # Retrieval
    start = time.time()
    search_results = vector_db.find_most_similar(query_embedding, k = 15)
    search_metadata = search_results[2]
    retrieval_time = time.time() - start

    retrieved_contents = {}
    for ret_cont in search_metadata:
        title = ret_cont['title']
        if title not in retrieved_contents:
            retrieved_contents[title] = []
        retrieved_contents[title].append(ret_cont['sentence'])
    
    retrieved_contents = {k: '\n'.join(v) for k, v in retrieved_contents.items() if len(v) > 2}

    return retrieved_contents, embedding_time, retrieval_time

def retrieval_pipeline(query, question):
    start = time.time()
    links = fetch_arxiv_links(query)
    websearch_time = time.time() - start

    start = time.time()
    pdf_metadata = fetch_data_from_pdfs(links)
    webcrawl_time = time.time() - start

    retrieved_contents, embedding_time, retrieval_time = index_and_search(question, pdf_metadata)

    return retrieved_contents, websearch_time, webcrawl_time, embedding_time, retrieval_time, links


async def predict(message, history):
    # message is in format: "Search: <query>; Question: <question>"
    # we need to parse both parts into variables
    message = message.split(';')

    query = message[0].split(':')[-1].strip()
    question = message[1].split(':')[-1].strip()

    retrieved_contents, websearch_time, webcrawl_time, embedding_time, retrieval_time, links = retrieval_pipeline(query, question)

    if detect_language(message) == Language.ptbr:
        context = ""
        for title, content in retrieved_contents.items():
            context += f'Artigo "{title}"\nConteúdo:\n{content}\n\n'
        prompt = f'{context.strip()}\n\nBaseado nos conteúdos dos artigos, responda: "{question}"\n\nPor favor, mencione a fonte da sua resposta.\nResponda somente em português brasileiro'
    else:
        context = ""
        for title, content in retrieved_contents.items():
            context += f'Article "{title}"\nContent:\n{content}\n\n'
        prompt = f'{context.strip()}\n\nBased on the article\'s contents, answer: "{question}"\n\nPlease, mention the source of your answer.'

    print(prompt)

    url = "https://openrouter.ai/api/v1/chat/completions"
    headers = { "Content-Type": "application/json",
                "Authorization": f"Bearer {openrouter_key}" }
    body = { "stream": True,
             "model": "deepseek/deepseek-chat",
             "max_tokens": 1024,
             "messages": [
                 {"role": "user", "content": prompt}
             ] }

    full_response = ""
    async with aiohttp.ClientSession() as session:
        async with session.post(url, headers=headers, json=body) as response:
            buffer = ""  # A buffer to hold incomplete lines of data
            async for chunk in response.content.iter_any():
                buffer += chunk.decode()
                while "\n" in buffer:  # Process as long as there are complete lines in the buffer
                    line, buffer = buffer.split("\n", 1)

                    if line.startswith("data: "):
                        event_data = line[len("data: "):]
                        if event_data != '[DONE]':
                            try:
                                current_text = json.loads(event_data)['choices'][0]['delta']['content']
                                full_response += current_text
                                yield full_response
                                await asyncio.sleep(0.01)
                            except Exception:
                                try:
                                    current_text = json.loads(event_data)['choices'][0]['text']
                                    full_response += current_text
                                    yield full_response
                                    await asyncio.sleep(0.01)
                                except Exception:
                                    pass
    
    final_metadata_block = ""

    final_metadata_block += f"Links visited:\n"
    for link in links:
        final_metadata_block += f"{link['title']} ({link['link']})\n"
    final_metadata_block += f"\nWeb search time: {websearch_time:.4f} seconds\n"
    final_metadata_block += f"\nText extraction: {webcrawl_time:.4f} seconds\n"
    final_metadata_block += f"\nEmbedding time: {embedding_time:.4f} seconds\n"
    final_metadata_block += f"\nRetrieval from VectorDB time: {retrieval_time:.4f} seconds"

    yield f"{full_response}\n\n{final_metadata_block}"

gr.ChatInterface(
    predict,
    title="Automated Arxiv Paper Search and Question Answering",
    description="Provide a search term and a question to find relevant papers and answer questions about them.",
    retry_btn=None,
    undo_btn=None,
    examples=[
        'Search: RAG LLM; Question: What are some challenges of implementing a system of RAG with LLMs?',
        'Search: LLM Self-Play; Question: What are the benefits of using self-play with LLMs?',
        'Search: Portable Blockchain; Question: How can a portable blockchain device be implemented?',
        'Search: 1.58 bit LLMs; Question: How do 1.58 bit LLMs work? Is there an available model to test?',
        'Search: Programação Robocode; Question: Como posso utilizar o robocode no contexto de aprendizagem de programação?',
        'Search: Pensamento Computacional; Question: Explique os conceitos do pensamento computacional.'
    ]
).launch()