app.py

import gradio as gr
import paperqa
import pickle
from pathlib import Path
import requests
import zipfile
import io
import tempfile
import os


docs = None


def request_pathname(files):
    if files is None:
        return [[]]
    return [[file.name, file.name.split('/')[-1], None] for file in files],  [[len(files), 0]]


def validate_dataset(dataset, openapi):
    global docs
    docs = None  # clear it out if dataset is modified
    docs_ready = dataset.iloc[-1, 0] != ""
    if docs_ready and type(openapi) is str and len(openapi) > 0:
        return "✨Ready✨"
    elif docs_ready:
        return "⚠️Waiting for key..."
    elif type(openapi) is str and len(openapi) > 0:
        return "⚠️Waiting for documents..."
    else:
        return "⚠️Waiting for documents and key..."


def make_stats(docs):
    return [[len(docs.doc_previews), sum([x[0] for x in docs.doc_previews])]]


# , progress=gr.Progress()):
def do_ask(question, button, openapi, dataset, length, do_marg, k, max_sources, docs):
    passages = ""
    docs_ready = dataset.iloc[-1, 0] != ""
    if button == "✨Ready✨" and type(openapi) is str and len(openapi) > 0 and docs_ready:
        if docs is None:  # don't want to rebuild index if it's already built
            import os
            os.environ['OPENAI_API_KEY'] = openapi.strip()
            docs = paperqa.Docs()
            # dataset is pandas dataframe
            for _, row in dataset.iterrows():
                try:
                    docs.add(row['filepath'], row['citation string'],
                             key=row['key'], disable_check=True)
                    yield "", "", "", docs, make_stats(docs)
                except Exception as e:
                    pass
    else:
        yield "", "", "", docs, [[0, 0]]
    #progress(0, "Building Index...")
    docs._build_faiss_index()
    #progress(0.25, "Querying...")
    for i, result in enumerate(docs.query_gen(question,
                                              length_prompt=f'use {length:d} words',
                                              marginal_relevance=do_marg,
                                              k=k, max_sources=max_sources)):
        #progress(0.25 + 0.1 * i, "Generating Context" + str(i))
        yield result.formatted_answer, result.context, passages, docs,  make_stats(docs)
    #progress(1.0, "Done!")
    # format the passages
    for i, (key, passage) in enumerate(result.passages.items()):
        passages += f'{i+1}. {key}\n\n >{passage} \n\n'
    yield result.formatted_answer, result.context, passages, docs,  make_stats(docs)


def download_repo(gh_repo, pbar=gr.Progress()):
    # download zipped version of repo
    r = requests.get(f'https://api.github.com/repos/{gh_repo}/zipball')
    files = []
    if r.status_code == 200:
        pbar(1, 'Downloaded')

        # iterate through files in zip
        with zipfile.ZipFile(io.BytesIO(r.content)) as z:
            for i, f in enumerate(z.namelist()):
                # skip directories
                if f.endswith('/'):
                    continue
                # try to read as plaintext (skip binary files)
                try:
                    text = z.read(f).decode('utf-8')
                except UnicodeDecodeError:
                    continue
                # check if it's bigger than 1MB or smaller than 10 bytes
                if len(text) > 1e6 or len(text) < 10:
                    continue
                # have to save to temporary file so we have a path
                with tempfile.NamedTemporaryFile(delete=False) as tmp:
                    tmp.write(text.encode('utf-8'))
                    tmp.flush()
                    path = tmp.name
                    # strip off the first directory of f
                    rel_path = '/'.join(f.split('/')[1:])
                    key = os.path.basename(f)
                    citation = f'[{rel_path}](https://github.com/{gh_repo}/tree/main/{rel_path})'
                    files.append([path, citation, key])
                    yield files, [[len(files), 0]]
                pbar(int((i+1)/len(z.namelist()) * 99),
                     f'Added {f}')
        pbar(100, 'Done')
    else:
        raise ValueError('Unknown Github Repo')


with gr.Blocks() as demo:

    docs = gr.State(None)
    openai_api_key = gr.State('')

    gr.Markdown(f"""
    # Document Question and Answer (v{paperqa.__version__})

    *By Andrew White ([@andrewwhite01](https://twitter.com/andrewwhite01))*

    This tool will enable asking questions of your uploaded text, PDF documents,
    or scrape github repos.
    It uses OpenAI's GPT models and thus you must enter your API key below. This
    tool is under active development and currently uses many tokens - up to 10,000
    for a single query. That is $0.10-0.20 per query, so please be careful!

    * [PaperQA](https://github.com/whitead/paper-qa) is the code used to build this tool.
    * [langchain](https://github.com/hwchase17/langchain) is the main library this tool utilizes.

    1. Enter API Key ([What is that?](https://platform.openai.com/account/api-keys))
    2. Upload your documents and modify citation strings if you want (to look prettier in answer)
    """)
    openai_api_key = gr.Textbox(
        label="OpenAI API Key", placeholder="sk-...", type="password")
    with gr.Tab('File Upload'):
        uploaded_files = gr.File(
            label="Your Documents Upload (PDF or txt)", file_count="multiple", )
    with gr.Tab('Github Repo'):
        gh_repo = gr.Textbox(
            label="Github Repo", placeholder="whitead/paper-qa")
        download = gr.Button("Download Repo")

    with gr.Accordion("See Docs:", open=False):
        dataset = gr.Dataframe(
            headers=["filepath", "citation string", "key"],
            datatype=["str", "str", "str"],
            col_count=(3, "fixed"),
            interactive=True,
            label="Documents and Citations",
            overflow_row_behaviour='paginate',
            max_rows=5
        )
    buildb = gr.Textbox("⚠️Waiting for documents and key...",
                        label="Status", interactive=False, show_label=True,
                        max_lines=1)
    stats = gr.Dataframe(headers=['Docs', 'Chunks'],
                         datatype=['number', 'number'],
                         col_count=(2, "fixed"),
                         interactive=False,
                         label="Doc Stats")
    openai_api_key.change(validate_dataset, inputs=[
                          dataset, openai_api_key], outputs=[buildb])
    dataset.change(validate_dataset, inputs=[
                   dataset, openai_api_key], outputs=[buildb])
    uploaded_files.change(request_pathname, inputs=[
                          uploaded_files], outputs=[dataset, stats])
    download.click(fn=download_repo, inputs=[
                   gh_repo], outputs=[dataset, stats])
    query = gr.Textbox(
        placeholder="Enter your question here...", label="Question")
    with gr.Row():
        length = gr.Slider(25, 200, value=100, step=5,
                           label='Words in answer')
        marg = gr.Checkbox(True, label='Max marginal relevance')
        k = gr.Slider(1, 20, value=10, step=1,
                      label='Chunks to examine')
        sources = gr.Slider(1, 10, value=5, step=1,
                            label='Contexts to include')

    ask = gr.Button("Ask Question")
    gr.Markdown("## Answer")
    answer = gr.Markdown(label="Answer")
    with gr.Accordion("Context", open=True):
        gr.Markdown(
            "### Context\n\nThe following context was used to generate the answer:")
        context = gr.Markdown(label="Context")

    with gr.Accordion("Raw Text", open=False):
        gr.Markdown(
            "### Raw Text\n\nThe following raw text was used to generate the answer:")
        passages = gr.Markdown(label="Passages")
    ask.click(fn=do_ask, inputs=[query, buildb,
                                 openai_api_key, dataset,
                                 length, marg, k, sources,
                                 docs], outputs=[answer, context, passages, docs, stats])

demo.queue(concurrency_count=20)
demo.launch(show_error=True)