zliang's picture
Update app.py (#1)
0476da0 verified
import os
os.system("python -m spacy download en_core_web_sm")
import io
import base64
import streamlit as st
import numpy as np
import fitz # PyMuPDF
import tempfile
from ultralytics import YOLO
from sklearn.cluster import KMeans
from sklearn.metrics.pairwise import cosine_similarity
from langchain_core.output_parsers import StrOutputParser
from langchain_community.document_loaders import PyMuPDFLoader
from langchain_openai import OpenAIEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_text_splitters import SpacyTextSplitter
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
import re
from PIL import Image
from streamlit_chat import message
# Load the trained model
model = YOLO("best.pt")
openai_api_key = os.environ.get("openai_api_key")
# Define the class indices for figures, tables, and text
figure_class_index = 4
table_class_index = 3
# Utility functions
def clean_text(text):
return re.sub(r'\s+', ' ', text).strip()
def remove_references(text):
reference_patterns = [
r'\bReferences\b', r'\breferences\b', r'\bBibliography\b', r'\bCitations\b',
r'\bWorks Cited\b', r'\bReference\b', r'\breference\b'
]
lines = text.split('\n')
for i, line in enumerate(lines):
if any(re.search(pattern, line, re.IGNORECASE) for pattern in reference_patterns):
return '\n'.join(lines[:i])
return text
def save_uploaded_file(uploaded_file):
temp_file = tempfile.NamedTemporaryFile(delete=False)
temp_file.write(uploaded_file.getbuffer())
temp_file.close()
return temp_file.name
def summarize_pdf(pdf_file_path, num_clusters=10):
embeddings_model = OpenAIEmbeddings(model="text-embedding-3-small", api_key=openai_api_key)
llm = ChatOpenAI(model="gpt-4o-mini", api_key=openai_api_key, temperature=0.3)
prompt = ChatPromptTemplate.from_template(
"""Could you please provide a concise and comprehensive summary of the given Contexts?
The summary should capture the main points and key details of the text while conveying the author's intended meaning accurately.
Please ensure that the summary is well-organized and easy to read, with clear headings and subheadings to guide the reader through each section.
The length of the summary should be appropriate to capture the main points and key details of the text, without including unnecessary information or becoming overly long.
example of summary:
## Summary:
## Key points:
Contexts: {topic}"""
)
output_parser = StrOutputParser()
chain = prompt | llm | output_parser
loader = PyMuPDFLoader(pdf_file_path)
docs = loader.load()
full_text = "\n".join(doc.page_content for doc in docs)
cleaned_full_text = clean_text(remove_references(full_text))
text_splitter = SpacyTextSplitter(chunk_size=500)
#text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0, separators=["\n\n", "\n", ".", " "])
split_contents = text_splitter.split_text(cleaned_full_text)
embeddings = embeddings_model.embed_documents(split_contents)
kmeans = KMeans(n_clusters=num_clusters, init='k-means++', random_state=0).fit(embeddings)
closest_point_indices = [np.argmin(np.linalg.norm(embeddings - center, axis=1)) for center in kmeans.cluster_centers_]
extracted_contents = [split_contents[idx] for idx in closest_point_indices]
results = chain.invoke({"topic": ' '.join(extracted_contents)})
return generate_citations(results, extracted_contents)
def qa_pdf(pdf_file_path, query, num_clusters=5, similarity_threshold=0.6):
embeddings_model = OpenAIEmbeddings(model="text-embedding-3-small", api_key=openai_api_key)
llm = ChatOpenAI(model="gpt-4o-mini", api_key=openai_api_key, temperature=0.3)
prompt = ChatPromptTemplate.from_template(
"""Please provide a detailed and accurate answer to the given question based on the provided contexts.
Ensure that the answer is comprehensive and directly addresses the query.
If necessary, include relevant examples or details from the text.
Question: {question}
Contexts: {contexts}"""
)
output_parser = StrOutputParser()
chain = prompt | llm | output_parser
loader = PyMuPDFLoader(pdf_file_path)
docs = loader.load()
full_text = "\n".join(doc.page_content for doc in docs)
cleaned_full_text = clean_text(remove_references(full_text))
text_splitter = SpacyTextSplitter(chunk_size=500)
#text_splitter = RecursiveCharacterTextSplitter(chunk_size=800, chunk_overlap=0, separators=["\n\n", "\n", ".", " "])
split_contents = text_splitter.split_text(cleaned_full_text)
embeddings = embeddings_model.embed_documents(split_contents)
query_embedding = embeddings_model.embed_query(query)
similarity_scores = cosine_similarity([query_embedding], embeddings)[0]
top_indices = np.argsort(similarity_scores)[-num_clusters:]
relevant_contents = [split_contents[i] for i in top_indices]
results = chain.invoke({"question": query, "contexts": ' '.join(relevant_contents)})
return generate_citations(results, relevant_contents, similarity_threshold)
def generate_citations(text, contents, similarity_threshold=0.6):
embeddings_model = OpenAIEmbeddings(model="text-embedding-3-small", api_key=openai_api_key)
text_sentences = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s', text)
text_embeddings = embeddings_model.embed_documents(text_sentences)
content_embeddings = embeddings_model.embed_documents(contents)
similarity_matrix = cosine_similarity(text_embeddings, content_embeddings)
cited_text = text
relevant_sources = []
source_mapping = {}
sentence_to_source = {}
for i, sentence in enumerate(text_sentences):
if sentence in sentence_to_source:
continue
max_similarity = max(similarity_matrix[i])
if max_similarity >= similarity_threshold:
most_similar_idx = np.argmax(similarity_matrix[i])
if most_similar_idx not in source_mapping:
source_mapping[most_similar_idx] = len(relevant_sources) + 1
relevant_sources.append((most_similar_idx, contents[most_similar_idx]))
citation_idx = source_mapping[most_similar_idx]
citation = f"([Source {citation_idx}](#source-{citation_idx}))"
cited_sentence = re.sub(r'([.!?])$', f" {citation}\\1", sentence)
sentence_to_source[sentence] = citation_idx
cited_text = cited_text.replace(sentence, cited_sentence)
sources_list = "\n\n## Sources:\n"
for idx, (original_idx, content) in enumerate(relevant_sources):
sources_list += f"""
<details style="margin: 1px 0; padding: 5px; border: 1px solid #ccc; border-radius: 8px; background-color: #f9f9f9; transition: all 0.3s ease;">
<summary style="font-weight: bold; cursor: pointer; outline: none; padding: 5px 0; transition: color 0.3s ease;">Source {idx + 1}</summary>
<pre style="white-space: pre-wrap; word-wrap: break-word; margin: 1px 0; padding: 10px; background-color: #fff; border-radius: 5px; border: 1px solid #ddd; box-shadow: 0 2px 5px rgba(0, 0, 0, 0.1);">{content}</pre>
</details>
"""
# Add dummy blanks after the last source
dummy_blanks = """
<div style="margin: 20px 0;"></div>
<div style="margin: 20px 0;"></div>
<div style="margin: 20px 0;"></div>
<div style="margin: 20px 0;"></div>
<div style="margin: 20px 0;"></div>
"""
cited_text += sources_list + dummy_blanks
return cited_text
def infer_image_and_get_boxes(image, confidence_threshold=0.8):
results = model.predict(image)
return [
(int(box.xyxy[0][0]), int(box.xyxy[0][1]), int(box.xyxy[0][2]), int(box.xyxy[0][3]), int(box.cls[0]))
for result in results for box in result.boxes
if int(box.cls[0]) in {figure_class_index, table_class_index} and box.conf[0] > confidence_threshold
]
def crop_images_from_boxes(image, boxes, scale_factor):
figures = []
tables = []
for (x1, y1, x2, y2, cls) in boxes:
cropped_img = image[int(y1 * scale_factor):int(y2 * scale_factor), int(x1 * scale_factor):int(x2 * scale_factor)]
if cls == figure_class_index:
figures.append(cropped_img)
elif cls == table_class_index:
tables.append(cropped_img)
return figures, tables
def process_pdf(pdf_file_path):
doc = fitz.open(pdf_file_path)
all_figures = []
all_tables = []
low_dpi = 50
high_dpi = 300
scale_factor = high_dpi / low_dpi
low_res_pixmaps = [page.get_pixmap(dpi=low_dpi) for page in doc]
for page_num, low_res_pix in enumerate(low_res_pixmaps):
low_res_img = np.frombuffer(low_res_pix.samples, dtype=np.uint8).reshape(low_res_pix.height, low_res_pix.width, 3)
boxes = infer_image_and_get_boxes(low_res_img)
if boxes:
high_res_pix = doc[page_num].get_pixmap(dpi=high_dpi)
high_res_img = np.frombuffer(high_res_pix.samples, dtype=np.uint8).reshape(high_res_pix.height, high_res_pix.width, 3)
figures, tables = crop_images_from_boxes(high_res_img, boxes, scale_factor)
all_figures.extend(figures)
all_tables.extend(tables)
return all_figures, all_tables
def image_to_base64(img):
buffered = io.BytesIO()
img = Image.fromarray(img)
img.save(buffered, format="PNG")
return base64.b64encode(buffered.getvalue()).decode()
def on_btn_click():
del st.session_state.chat_history[:]
# Streamlit interface
# Custom CSS for the file uploader
uploadercss='''
<style>
[data-testid='stFileUploader'] {
width: max-content;
}
[data-testid='stFileUploader'] section {
padding: 0;
float: left;
}
[data-testid='stFileUploader'] section > input + div {
display: none;
}
[data-testid='stFileUploader'] section + div {
float: right;
padding-top: 0;
}
</style>
'''
st.set_page_config(page_title="PDF Reading Assistant", page_icon="πŸ“„")
# Initialize chat history in session state if not already present
if 'chat_history' not in st.session_state:
st.session_state.chat_history = []
st.title("πŸ“„ PDF Reading Assistant")
st.markdown("### Extract tables, figures, summaries, and answers from your PDF files easily.")
chat_placeholder = st.empty()
# File uploader for PDF
uploaded_file = st.file_uploader("Upload a PDF", type="pdf")
st.markdown(uploadercss, unsafe_allow_html=True)
if uploaded_file:
file_path = save_uploaded_file(uploaded_file)
# Chat container where all messages will be displayed
chat_container = st.container()
user_input = st.chat_input("Ask a question about the pdf......", key="user_input")
with chat_container:
# Scrollable chat messages
for idx, chat in enumerate(st.session_state.chat_history):
if chat.get("user"):
message(chat["user"], is_user=True, allow_html=True, key=f"user_{idx}", avatar_style="initials", seed="user")
if chat.get("bot"):
message(chat["bot"], is_user=False, allow_html=True, key=f"bot_{idx}",seed="bot")
# Input area and buttons for user interaction
with st.form(key="chat_form", clear_on_submit=True,border=False):
col1, col2, col3 = st.columns([1, 1, 1])
with col1:
summary_button = st.form_submit_button("Generate Summary")
with col2:
extract_button = st.form_submit_button("Extract Tables and Figures")
with col3:
st.form_submit_button("Clear message", on_click=on_btn_click)
# Handle responses based on user input and button presses
if summary_button:
with st.spinner("Generating summary..."):
summary = summarize_pdf(file_path)
st.session_state.chat_history.append({"user": "Generate Summary", "bot": summary})
st.rerun()
if extract_button:
with st.spinner("Extracting tables and figures..."):
figures, tables = process_pdf(file_path)
if figures:
st.session_state.chat_history.append({"user": "Figures"})
for idx, figure in enumerate(figures):
figure_base64 = image_to_base64(figure)
result_html = f'<img src="data:image/png;base64,{figure_base64}" style="width:100%; display:block;" alt="Figure {idx+1}"/>'
st.session_state.chat_history.append({"bot": f"Figure {idx+1} {result_html}"})
if tables:
st.session_state.chat_history.append({"user": "Tables"})
for idx, table in enumerate(tables):
table_base64 = image_to_base64(table)
result_html = f'<img src="data:image/png;base64,{table_base64}" style="width:100%; display:block;" alt="Table {idx+1}"/>'
st.session_state.chat_history.append({"bot": f"Table {idx+1} {result_html}"})
st.rerun()
if user_input:
st.session_state.chat_history.append({"user": user_input, "bot": None})
with st.spinner("Processing..."):
answer = qa_pdf(file_path, user_input)
st.session_state.chat_history[-1]["bot"] = answer
st.rerun()
# Additional CSS and JavaScript to ensure the chat container is scrollable and scrolls to the bottom
st.markdown("""
<style>
#chat-container {
max-height: 500px;
overflow-y: auto;
padding: 1rem;
border: 1px solid #ddd;
border-radius: 8px;
background-color: #fefefe;
box-shadow: 0 2px 4px rgba(0, 0, 0, 0.1);
transition: background-color 0.3s ease;
}
#chat-container:hover {
background-color: #f9f9f9;
}
.stChatMessage {
padding: 0.75rem;
margin: 0.75rem 0;
border-radius: 8px;
box-shadow: 0 1px 3px rgba(0, 0, 0, 0.1);
transition: background-color 0.3s ease;
}
.stChatMessage--user {
background-color: #E3F2FD;
}
.stChatMessage--user:hover {
background-color: #BBDEFB;
}
.stChatMessage--bot {
background-color: #EDE7F6;
}
.stChatMessage--bot:hover {
background-color: #D1C4E9;
}
textarea {
width: 100%;
padding: 1rem;
border: 1px solid #ddd;
border-radius: 8px;
box-shadow: inset 0 1px 3px rgba(0, 0, 0, 0.1);
transition: border-color 0.3s ease, box-shadow 0.3s ease;
}
textarea:focus {
border-color: #4CAF50;
box-shadow: 0 0 5px rgba(76, 175, 80, 0.5);
}
.stButton > button {
width: 100%;
background-color: #4CAF50;
color: white;
border: none;
border-radius: 8px;
padding: 0.75rem;
font-size: 16px;
box-shadow: 0 2px 4px rgba(0, 0, 0, 0.1);
transition: background-color 0.3s ease, box-shadow 0.3s ease;
}
.stButton > button:hover {
background-color: #45A049;
box-shadow: 0 4px 8px rgba(0, 0, 0, 0.1);
}
</style>
<script>
const chatContainer = document.getElementById('chat-container');
chatContainer.scrollTop = chatContainer.scrollHeight;
</script>
""", unsafe_allow_html=True)