Alanturner2's picture
Create app.py
7c1cc3d verified
raw
history blame
9.14 kB
import gradio as gr
from langchain.document_loaders import ArxivLoader
from PyPDF2 import PdfReader
from langchain_community.llms import HuggingFaceHub
from langchain.text_splitter import TokenTextSplitter
from langchain.chains.summarize import load_summarize_chain
from langchain.document_loaders import PyPDFLoader
from transformers import pipeline
from dotenv import load_dotenv
import os
load_dotenv()
hugging_api_key = os.getenv('HUGGING_API_KEY')
from groq import AsyncGroq
from groq import Groq
from langchain_groq import ChatGroq
from langchain.document_loaders import ArxivLoader
from langchain.vectorstores import Chroma
from langchain.chains import RetrievalQA
from langchain.embeddings.huggingface_hub import HuggingFaceHubEmbeddings
from huggingface_hub import login
login(hugging_api_key)
embedding_model = HuggingFaceHubEmbeddings(huggingfacehub_api_token=hugging_api_key)
llm = ChatGroq(temperature=0, model_name="llama3-70b-8192", api_key = "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")
def display_results(result):
return "\n".join(result) # Join each entry with double newlines for better readability
def summarize_pdf(pdf_file_path, max_length):
# summarizer = pipeline('summarization', model='allenai/led-large-16384-arxiv', min_length=100, max_length=max_length, device=0)
loader = PdfReader(pdf_file_path)
text = """ """
for page in loader.pages:
text += page.extract_text()
text_splitter = TokenTextSplitter(chunk_size=8192, chunk_overlap=1000)
chunks = text_splitter.split_text(text)
summary = ""
for i in range(len(chunks)):
# text = chunks[i].page_content
text = chunks[i]
summary += summarize_text(text)
# summary = str(max_length)
return summary
def summarize_text(text):
sum_client = Groq(api_key="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")
messages = []
# messages.append({"role": "system", "content": "You are arxiv paper summarizer. If I give you the doi number, you should only output summarization. Summarization should be more than 10% words of the paper. For example, in the paper there are 500 words, than summarization should be more than 50 words."})
messages.append({"role": "system", "content": "You are summarizer. If I give you the whole text you should summarize it. And you don't need the title and author"})
messages = messages + [
{
"role": "user",
"content": f"Summarize the paper. The whole text is {text}",
},
]
response = sum_client.chat.completions.create(
messages=messages,
model="llama3-70b-8192",
temperature=0,
max_tokens=8192,
top_p=1,
stop=None
)
text_summary = response.choices[0].message.content
return text_summary
def remove_first_sentence_and_title(text):
# Remove the first sentence
first_sentence_end = text.find('. ') + 2 # Find the end of the first sentence
text_without_first_sentence = text[first_sentence_end:]
# Remove the title
title_start = text_without_first_sentence.find('**Title:**')
if title_start != -1:
title_end = text_without_first_sentence.find('\n', title_start)
if title_end != -1:
text_without_title = text_without_first_sentence[:title_start] + text_without_first_sentence[title_end+1:]
else:
text_without_title = text_without_first_sentence[:title_start]
else:
text_without_title = text_without_first_sentence
return text_without_title.strip()
def summarize_arxiv_pdf(query):
loader = ArxivLoader(query=query, load_max_docs=10)
documents = loader.load()
text_splitter = TokenTextSplitter(chunk_size=5700, chunk_overlap=100)
chunks = text_splitter.split_documents(documents)
text = documents[0].page_content
ref_summary = ""
for i in range(len(chunks)):
text = chunks[i].page_content
ref_summary += summarize_text(text)
# ref_summary = ref_summary.split('paper:')[1]
# ref_summary = remove_first_sentence_and_title(ref_summary)
ref_summary = ref_summary.replace("Here is a summary of the paper:", "").strip()
arxiv_summary = loader.get_summaries_as_docs()
summaries = []
for doc in arxiv_summary:
title = doc.metadata.get("Title")
authors = doc.metadata.get("Authors")
url = doc.metadata.get("Entry ID")
summary = doc.page_content
summaries.append(f"**{title}**\n")
summaries.append(f"**Authors:** {authors}\n")
summaries.append(f"**View full paper:** [Link to paper]({url})\n")
summaries.append(f"**Summary:** {summary}\n")
summaries.append(f"**Lazyman Summary:**\n ")
summaries.append(f"{ref_summary}")
summaries = display_results(summaries)
print(summaries)
return summaries
client = AsyncGroq(api_key="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")
async def chat_with_replit(message, history):
messages = []
for chat in history:
user = str(chat[0])
assistant = str(chat[1])
messages.append({"role": "system", "content": "You are assistor. I will ask you some questions than you should answer!"})
messages.append({"role": 'user', "content": user})
messages.append({"role": 'assistant', "content": assistant})
messages = messages + [
{
"role": "user",
"content": str(message),
},
]
print(messages)
response_content = ""
stream = await client.chat.completions.create(
messages=messages,
model="llama3-70b-8192",
temperature=0,
max_tokens=1024,
top_p=1,
stop=None,
stream=True,
)
async for chunk in stream:
content = chunk.choices[0].delta.content
if content:
response_content += chunk.choices[0].delta.content
yield response_content
js = """<script src="https://replit.com/public/js/replit-badge-v2.js" theme="dark" position="bottom-right"></script>"""
async def chat_with_replit_pdf(message, history, doi_num):
messages = []
old_doi = "old"
if old_doi != doi_num:
loader = ArxivLoader(query=str(doi_num), load_max_docs=10)
documents = loader.load_and_split()
metadata = documents[0].metadata
vector_store = Chroma.from_documents(documents, embedding_model)
old_doi = doi_num
def retrieve_relevant_content(user_query):
results = vector_store.similarity_search(user_query, k=3)
relevant_content = "\n\n".join([doc.page_content for doc in results])
return relevant_content
relevant_content = retrieve_relevant_content(message)
messages = messages + [
{
"role": "user",
"content": str(message),
},
{
"role": "system",
"content": f"You should answer about this arxiv paper for {doi_num}.\n"
f"This is the metadata of the paper:{metadata}.\n"
f"This is relevant information of the paper:{relevant_content}.\n"
}
]
print(messages)
response_content = ""
stream = await client.chat.completions.create(
messages=messages,
model="llama3-70b-8192",
temperature=0,
max_tokens=1024,
top_p=1,
stop=None,
stream=False,
)
return stream.choices[0].message.content;
with gr.Blocks() as app:
with gr.Tab(label="Arxiv summarization"):
with gr.Column():
number = gr.Textbox(label="Enter your arxiv number")
sumarxiv_btn = gr.Button(value="summarize-arxiv")
with gr.Column():
outputs = gr.Markdown(label="Summary", height=1000)
sumarxiv_btn.click(summarize_arxiv_pdf, inputs=number, outputs=outputs)
with gr.Tab(label="Local summarization"):
with gr.Row():
with gr.Column():
input_path = gr.File(label="Upload PDF file")
with gr.Column():
# set_temperature = gr.Slider(0, 1, value=0, step=0.1, label="temperature")
set_max_length = gr.Slider(512, 4096, value=2048, step=512, label="max length")
sumlocal_btn = gr.Button(value="summarize-local")
with gr.Row():
output_local = gr.Markdown(label="summary", height=1000)
sumlocal_btn.click(summarize_pdf, inputs=[input_path, set_max_length], outputs=output_local)
with gr.Tab(label="ChatBot"):
gr.ChatInterface(chat_with_replit,
examples=[
"Explain about the attention is all you need",
"Who is the inventor of the GAN",
"What is the main idea style transfer?"
])
with gr.Tab(label="Chat with pdf"):
gr.ChatInterface(fn = chat_with_replit_pdf,
additional_inputs = [
gr.Textbox(label="doi", placeholder="Enter doi number")
],
type="messages")
app.launch()