Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -1,446 +1,446 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
4 |
-
import os
|
5 |
-
import streamlit as st
|
6 |
-
import numpy as np
|
7 |
-
import fitz # PyMuPDF
|
8 |
-
from ultralytics import YOLO
|
9 |
-
from sklearn.cluster import KMeans
|
10 |
-
from sklearn.metrics.pairwise import cosine_similarity
|
11 |
-
from langchain_core.output_parsers import StrOutputParser
|
12 |
-
from langchain_community.document_loaders import PyMuPDFLoader
|
13 |
-
from langchain_openai import OpenAIEmbeddings
|
14 |
-
from langchain_text_splitters import RecursiveCharacterTextSplitter
|
15 |
-
from langchain_core.prompts import ChatPromptTemplate
|
16 |
-
from sklearn.decomposition import PCA
|
17 |
-
from langchain_openai import ChatOpenAI
|
18 |
-
import string
|
19 |
-
import re
|
20 |
-
|
21 |
-
|
22 |
-
# Load the trained model
|
23 |
-
model = YOLO("runs\\detect\\train7\\weights\\best.pt")
|
24 |
-
openai_api_key = os.environ.get("openai_api_key")
|
25 |
-
|
26 |
-
# Define the class indices for figures, tables, and text
|
27 |
-
figure_class_index = 4 # class index for figures
|
28 |
-
table_class_index = 3 # class index for tables
|
29 |
-
|
30 |
-
# Global variables to store embeddings and contents
|
31 |
-
global_embeddings = None
|
32 |
-
global_split_contents = None
|
33 |
-
|
34 |
-
def clean_text(text):
|
35 |
-
text = re.sub(r'\s+', ' ', text).strip()
|
36 |
-
return text
|
37 |
-
|
38 |
-
def remove_references(text):
|
39 |
-
reference_patterns = [
|
40 |
-
r'\bReferences\b', r'\breferences\b', r'\bBibliography\b', r'\bCitations\b',
|
41 |
-
r'\bWorks Cited\b', r'\bReference\b', r'\breference\b'
|
42 |
-
]
|
43 |
-
lines = text.split('\n')
|
44 |
-
for i, line in enumerate(lines):
|
45 |
-
if any(re.search(pattern, line, re.IGNORECASE) for pattern in reference_patterns):
|
46 |
-
return '\n'.join(lines[:i])
|
47 |
-
return text
|
48 |
-
|
49 |
-
def save_uploaded_file(uploaded_file):
|
50 |
-
with open(uploaded_file.name, 'wb') as f:
|
51 |
-
f.write(uploaded_file.getbuffer())
|
52 |
-
return uploaded_file.name
|
53 |
-
|
54 |
-
def summarize_pdf(pdf_file_path, num_clusters=10):
|
55 |
-
embeddings_model = OpenAIEmbeddings(model="text-embedding-3-small", api_key=openai_api_key)
|
56 |
-
llm = ChatOpenAI(model="gpt-3.5-turbo", api_key=openai_api_key, temperature=0.3)
|
57 |
-
prompt = ChatPromptTemplate.from_template(
|
58 |
-
"""Could you please provide a concise and comprehensive summary of the given Contexts?
|
59 |
-
The summary should capture the main points and key details of the text while conveying the author's intended meaning accurately.
|
60 |
-
Please ensure that the summary is well-organized and easy to read, with clear headings and subheadings to guide the reader through each section.
|
61 |
-
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.
|
62 |
-
example of summary:
|
63 |
-
## Summary:
|
64 |
-
## Key points:
|
65 |
-
Contexts: {topic}"""
|
66 |
-
)
|
67 |
-
output_parser = StrOutputParser()
|
68 |
-
chain = prompt | llm | output_parser
|
69 |
-
|
70 |
-
loader = PyMuPDFLoader(pdf_file_path)
|
71 |
-
docs = loader.load()
|
72 |
-
full_text = "\n".join(doc.page_content for doc in docs)
|
73 |
-
cleaned_full_text = remove_references(full_text)
|
74 |
-
cleaned_full_text = clean_text(cleaned_full_text)
|
75 |
-
|
76 |
-
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0,separators=["\n\n", "\n",".", " "])
|
77 |
-
split_contents = text_splitter.split_text(cleaned_full_text)
|
78 |
-
embeddings = embeddings_model.embed_documents(split_contents)
|
79 |
-
|
80 |
-
X = np.array(embeddings)
|
81 |
-
kmeans = KMeans(n_clusters=num_clusters, init='k-means++', random_state=0).fit(embeddings)
|
82 |
-
cluster_centers = kmeans.cluster_centers_
|
83 |
-
|
84 |
-
closest_point_indices = []
|
85 |
-
for center in cluster_centers:
|
86 |
-
distances = np.linalg.norm(embeddings - center, axis=1)
|
87 |
-
closest_point_indices.append(np.argmin(distances))
|
88 |
-
|
89 |
-
extracted_contents = [split_contents[idx] for idx in closest_point_indices]
|
90 |
-
results = chain.invoke({"topic": ' '.join(extracted_contents)})
|
91 |
-
|
92 |
-
summary_sentences = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s', results)
|
93 |
-
summary_embeddings = embeddings_model.embed_documents(summary_sentences)
|
94 |
-
extracted_embeddings = embeddings_model.embed_documents(extracted_contents)
|
95 |
-
similarity_matrix = cosine_similarity(summary_embeddings, extracted_embeddings)
|
96 |
-
|
97 |
-
cited_results = results
|
98 |
-
relevant_sources = []
|
99 |
-
source_mapping = {}
|
100 |
-
sentence_to_source = {}
|
101 |
-
similarity_threshold = 0.6
|
102 |
-
|
103 |
-
for i, sentence in enumerate(summary_sentences):
|
104 |
-
if sentence in sentence_to_source:
|
105 |
-
continue
|
106 |
-
max_similarity = max(similarity_matrix[i])
|
107 |
-
if max_similarity >= similarity_threshold:
|
108 |
-
most_similar_idx = np.argmax(similarity_matrix[i])
|
109 |
-
if most_similar_idx not in source_mapping:
|
110 |
-
source_mapping[most_similar_idx] = len(relevant_sources) + 1
|
111 |
-
relevant_sources.append((most_similar_idx, extracted_contents[most_similar_idx]))
|
112 |
-
citation_idx = source_mapping[most_similar_idx]
|
113 |
-
citation = f"([Source {citation_idx}](#source-{citation_idx}))"
|
114 |
-
cited_sentence = re.sub(r'([.!?])$', f" {citation}\\1", sentence)
|
115 |
-
sentence_to_source[sentence] = citation_idx
|
116 |
-
cited_results = cited_results.replace(sentence, cited_sentence)
|
117 |
-
|
118 |
-
sources_list = "\n\n## Sources:\n"
|
119 |
-
for idx, (original_idx, content) in enumerate(relevant_sources):
|
120 |
-
sources_list += f"""
|
121 |
-
<details style="margin: 10px 0; padding: 10px; border: 1px solid #ccc; border-radius: 5px; background-color: #f9f9f9;">
|
122 |
-
<summary style="font-weight: bold; cursor: pointer;">Source {idx + 1}</summary>
|
123 |
-
<pre style="white-space: pre-wrap; word-wrap: break-word; margin-top: 10px;">{content}</pre>
|
124 |
-
</details>
|
125 |
-
"""
|
126 |
-
cited_results += sources_list
|
127 |
-
return cited_results
|
128 |
-
|
129 |
-
def qa_pdf(pdf_file_path, query, num_clusters=5, similarity_threshold=0.6):
|
130 |
-
global global_embeddings, global_split_contents
|
131 |
-
|
132 |
-
# Initialize models and embeddings
|
133 |
-
embeddings_model = OpenAIEmbeddings(model="text-embedding-3-small", api_key=openai_api_key)
|
134 |
-
llm = ChatOpenAI(model="gpt-3.5-turbo", api_key=openai_api_key, temperature=0.3)
|
135 |
-
prompt = ChatPromptTemplate.from_template(
|
136 |
-
"""Please provide a detailed and accurate answer to the given question based on the provided contexts.
|
137 |
-
Ensure that the answer is comprehensive and directly addresses the query.
|
138 |
-
If necessary, include relevant examples or details from the text.
|
139 |
-
Question: {question}
|
140 |
-
Contexts: {contexts}"""
|
141 |
-
)
|
142 |
-
output_parser = StrOutputParser()
|
143 |
-
chain = prompt | llm | output_parser
|
144 |
-
|
145 |
-
# Load and process the PDF if not already loaded
|
146 |
-
if global_embeddings is None or global_split_contents is None:
|
147 |
-
loader = PyMuPDFLoader(pdf_file_path)
|
148 |
-
docs = loader.load()
|
149 |
-
full_text = "\n".join(doc.page_content for doc in docs)
|
150 |
-
cleaned_full_text = remove_references(full_text)
|
151 |
-
cleaned_full_text = clean_text(cleaned_full_text)
|
152 |
-
|
153 |
-
text_splitter = RecursiveCharacterTextSplitter(chunk_size=800, chunk_overlap=0, separators=["\n\n", "\n", ".", " "])
|
154 |
-
global_split_contents = text_splitter.split_text(cleaned_full_text)
|
155 |
-
global_embeddings = embeddings_model.embed_documents(global_split_contents)
|
156 |
-
|
157 |
-
# Embed the query and find the most relevant contexts
|
158 |
-
query_embedding = embeddings_model.embed_query(query)
|
159 |
-
similarity_scores = cosine_similarity([query_embedding], global_embeddings)[0]
|
160 |
-
top_indices = np.argsort(similarity_scores)[-num_clusters:]
|
161 |
-
relevant_contents = [global_split_contents[i] for i in top_indices]
|
162 |
-
|
163 |
-
# Generate the answer using the LLM chain
|
164 |
-
results = chain.invoke({"question": query, "contexts": ' '.join(relevant_contents)})
|
165 |
-
|
166 |
-
# Split the answer into sentences and embed them
|
167 |
-
answer_sentences = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s', results)
|
168 |
-
answer_embeddings = embeddings_model.embed_documents(answer_sentences)
|
169 |
-
relevant_embeddings = embeddings_model.embed_documents(relevant_contents)
|
170 |
-
similarity_matrix = cosine_similarity(answer_embeddings, relevant_embeddings)
|
171 |
-
|
172 |
-
# Map sentences to sources and create citations
|
173 |
-
cited_results = results
|
174 |
-
relevant_sources = []
|
175 |
-
source_mapping = {}
|
176 |
-
sentence_to_source = {}
|
177 |
-
|
178 |
-
for i, sentence in enumerate(answer_sentences):
|
179 |
-
if sentence in sentence_to_source:
|
180 |
-
continue
|
181 |
-
max_similarity = max(similarity_matrix[i])
|
182 |
-
if max_similarity >= similarity_threshold:
|
183 |
-
most_similar_idx = np.argmax(similarity_matrix[i])
|
184 |
-
if most_similar_idx not in source_mapping:
|
185 |
-
source_mapping[most_similar_idx] = len(relevant_sources) + 1
|
186 |
-
relevant_sources.append((most_similar_idx, relevant_contents[most_similar_idx]))
|
187 |
-
citation_idx = source_mapping[most_similar_idx]
|
188 |
-
citation = f"<strong style='color:blue;'>[Source {citation_idx}]</strong>"
|
189 |
-
cited_sentence = re.sub(r'([.!?])$', f" {citation}\\1", sentence)
|
190 |
-
sentence_to_source[sentence] = citation_idx
|
191 |
-
cited_results = cited_results.replace(sentence, cited_sentence)
|
192 |
-
|
193 |
-
# Format the sources for markdown rendering
|
194 |
-
sources_list = "\n\n## Sources:\n"
|
195 |
-
for idx, (original_idx, content) in enumerate(relevant_sources):
|
196 |
-
sources_list += f"""
|
197 |
-
<details style="margin: 10px 0; padding: 10px; border: 1px solid #ccc; border-radius: 5px; background-color: #f9f9f9;">
|
198 |
-
<summary style="font-weight: bold; cursor: pointer;">Source {idx + 1}</summary>
|
199 |
-
<pre style="white-space: pre-wrap; word-wrap: break-word; margin-top: 10px;">{content}</pre>
|
200 |
-
</details>
|
201 |
-
"""
|
202 |
-
cited_results += sources_list
|
203 |
-
return cited_results
|
204 |
-
|
205 |
-
|
206 |
-
def infer_image_and_get_boxes(image, confidence_threshold=0.6):
|
207 |
-
results = model.predict(image)
|
208 |
-
boxes = [
|
209 |
-
(int(box.xyxy[0][0]), int(box.xyxy[0][1]), int(box.xyxy[0][2]), int(box.xyxy[0][3]), int(box.cls[0]))
|
210 |
-
for result in results for box in result.boxes
|
211 |
-
if int(box.cls[0]) in {figure_class_index, table_class_index} and box.conf[0] > confidence_threshold
|
212 |
-
]
|
213 |
-
return boxes
|
214 |
-
|
215 |
-
def crop_images_from_boxes(image, boxes, scale_factor):
|
216 |
-
figures = []
|
217 |
-
tables = []
|
218 |
-
for (x1, y1, x2, y2, cls) in boxes:
|
219 |
-
cropped_img = image[int(y1 * scale_factor):int(y2 * scale_factor), int(x1 * scale_factor):int(x2 * scale_factor)]
|
220 |
-
if cls == figure_class_index:
|
221 |
-
figures.append(cropped_img)
|
222 |
-
elif cls == table_class_index:
|
223 |
-
tables.append(cropped_img)
|
224 |
-
return figures, tables
|
225 |
-
|
226 |
-
|
227 |
-
def process_pdf(pdf_file_path):
|
228 |
-
doc = fitz.open(pdf_file_path)
|
229 |
-
all_figures = []
|
230 |
-
all_tables = []
|
231 |
-
low_dpi = 50
|
232 |
-
high_dpi = 300
|
233 |
-
scale_factor = high_dpi / low_dpi
|
234 |
-
low_res_pixmaps = [page.get_pixmap(dpi=low_dpi) for page in doc]
|
235 |
-
|
236 |
-
for page_num, low_res_pix in enumerate(low_res_pixmaps):
|
237 |
-
low_res_img = np.frombuffer(low_res_pix.samples, dtype=np.uint8).reshape(low_res_pix.height, low_res_pix.width, 3)
|
238 |
-
boxes = infer_image_and_get_boxes(low_res_img)
|
239 |
-
|
240 |
-
if boxes:
|
241 |
-
high_res_pix = doc[page_num].get_pixmap(dpi=high_dpi)
|
242 |
-
high_res_img = np.frombuffer(high_res_pix.samples, dtype=np.uint8).reshape(high_res_pix.height, high_res_pix.width, 3)
|
243 |
-
figures, tables = crop_images_from_boxes(high_res_img, boxes, scale_factor)
|
244 |
-
all_figures.extend(figures)
|
245 |
-
all_tables.extend(tables)
|
246 |
-
|
247 |
-
return all_figures, all_tables
|
248 |
-
|
249 |
-
# Set the page configuration for a modern look
|
250 |
-
|
251 |
-
# Set the page configuration for a modern look
|
252 |
-
# Set the page configuration for a modern look
|
253 |
-
st.set_page_config(page_title="PDF Reading Assistant", page_icon="π", layout="wide")
|
254 |
-
|
255 |
-
# Add some custom CSS for a modern look
|
256 |
-
st.markdown("""
|
257 |
-
<style>
|
258 |
-
/* Main background and padding */
|
259 |
-
.main {
|
260 |
-
background-color: #f8f9fa;
|
261 |
-
padding: 2rem;
|
262 |
-
font-family: 'Arial', sans-serif;
|
263 |
-
}
|
264 |
-
|
265 |
-
/* Section headers */
|
266 |
-
.section-header {
|
267 |
-
font-size: 2rem;
|
268 |
-
font-weight: bold;
|
269 |
-
color: #343a40;
|
270 |
-
margin-top: 2rem;
|
271 |
-
margin-bottom: 1rem;
|
272 |
-
text-align: center;
|
273 |
-
font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
|
274 |
-
}
|
275 |
-
|
276 |
-
/* Containers */
|
277 |
-
.uploaded-file-container, .chat-container, .summary-container, .extract-container {
|
278 |
-
padding: 2rem;
|
279 |
-
background-color: #ffffff;
|
280 |
-
border-radius: 10px;
|
281 |
-
margin-bottom: 2rem;
|
282 |
-
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.1);
|
283 |
-
}
|
284 |
-
|
285 |
-
/* Buttons */
|
286 |
-
.stButton>button {
|
287 |
-
background-color: #007bff;
|
288 |
-
color: white;
|
289 |
-
padding: 0.6rem 1.2rem;
|
290 |
-
border-radius: 5px;
|
291 |
-
border: none;
|
292 |
-
cursor: pointer;
|
293 |
-
font-size: 1rem;
|
294 |
-
transition: background-color 0.3s ease, transform 0.3s ease;
|
295 |
-
}
|
296 |
-
.stButton>button:hover {
|
297 |
-
background-color: #0056b3;
|
298 |
-
transform: translateY(-2px);
|
299 |
-
}
|
300 |
-
|
301 |
-
/* Chat messages */
|
302 |
-
.chat-message {
|
303 |
-
padding: 1rem;
|
304 |
-
border-radius: 10px;
|
305 |
-
margin-bottom: 1rem;
|
306 |
-
font-size: 1rem;
|
307 |
-
transition: all 0.3s ease;
|
308 |
-
box-shadow: 0 2px 8px rgba(0, 0, 0, 0.1);
|
309 |
-
}
|
310 |
-
.chat-message.user {
|
311 |
-
background-color: #e6f7ff;
|
312 |
-
border-left: 5px solid #007bff;
|
313 |
-
text-align: left;
|
314 |
-
}
|
315 |
-
.chat-message.bot {
|
316 |
-
background-color: #fff0f1;
|
317 |
-
border-left: 5px solid #dc3545;
|
318 |
-
text-align: left;
|
319 |
-
}
|
320 |
-
|
321 |
-
/* Input area */
|
322 |
-
.input-container {
|
323 |
-
display: flex;
|
324 |
-
align-items: center;
|
325 |
-
gap: 10px;
|
326 |
-
margin-top: 1rem;
|
327 |
-
}
|
328 |
-
.input-container textarea {
|
329 |
-
border: 2px solid #ccc;
|
330 |
-
border-radius: 10px;
|
331 |
-
padding: 10px;
|
332 |
-
width: 100%;
|
333 |
-
background-color: #fff;
|
334 |
-
transition: border-color 0.3s ease;
|
335 |
-
margin: 0;
|
336 |
-
font-size: 1rem;
|
337 |
-
}
|
338 |
-
.input-container textarea:focus {
|
339 |
-
border-color: #007bff;
|
340 |
-
outline: none;
|
341 |
-
}
|
342 |
-
.input-container button {
|
343 |
-
background-color: #007bff;
|
344 |
-
color: white;
|
345 |
-
padding: 0.6rem 1.2rem;
|
346 |
-
border-radius: 5px;
|
347 |
-
border: none;
|
348 |
-
cursor: pointer;
|
349 |
-
font-size: 1rem;
|
350 |
-
transition: background-color 0.3s ease, transform 0.3s ease;
|
351 |
-
}
|
352 |
-
.input-container button:hover {
|
353 |
-
background-color: #0056b3;
|
354 |
-
transform: translateY(-2px);
|
355 |
-
}
|
356 |
-
|
357 |
-
/* Expander */
|
358 |
-
.st-expander {
|
359 |
-
border: none;
|
360 |
-
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.1);
|
361 |
-
margin-bottom: 2rem;
|
362 |
-
}
|
363 |
-
|
364 |
-
/* Markdown elements */
|
365 |
-
.stMarkdown {
|
366 |
-
font-size: 1rem;
|
367 |
-
color: #343a40;
|
368 |
-
line-height: 1.6;
|
369 |
-
}
|
370 |
-
|
371 |
-
/* Titles and subtitles */
|
372 |
-
.stTitle {
|
373 |
-
color: #343a40;
|
374 |
-
text-align: center;
|
375 |
-
margin-bottom: 1rem;
|
376 |
-
font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
|
377 |
-
}
|
378 |
-
.stSubtitle {
|
379 |
-
color: #6c757d;
|
380 |
-
text-align: center;
|
381 |
-
margin-bottom: 1rem;
|
382 |
-
font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
|
383 |
-
}
|
384 |
-
</style>
|
385 |
-
""", unsafe_allow_html=True)
|
386 |
-
|
387 |
-
# Streamlit interface
|
388 |
-
st.title("π PDF Reading Assistant")
|
389 |
-
st.markdown("### Extract tables, figures, summaries, and answers from your PDF files easily.")
|
390 |
-
|
391 |
-
uploaded_file = st.file_uploader("Upload a PDF", type="pdf")
|
392 |
-
if uploaded_file:
|
393 |
-
file_path = save_uploaded_file(uploaded_file)
|
394 |
-
with st.container():
|
395 |
-
st.markdown("<div class='section-header'>Extract Tables and Figures</div>", unsafe_allow_html=True)
|
396 |
-
with st.expander("Click to Extract Tables and Figures", expanded=True):
|
397 |
-
with st.container():
|
398 |
-
extract_button = st.button("Extract")
|
399 |
-
if extract_button:
|
400 |
-
figures, tables = process_pdf(file_path)
|
401 |
-
col1, col2 = st.columns(2)
|
402 |
-
with col1:
|
403 |
-
st.write("### Figures")
|
404 |
-
if figures:
|
405 |
-
for figure in figures:
|
406 |
-
st.image(figure, use_column_width=True)
|
407 |
-
else:
|
408 |
-
st.write("No figures found.")
|
409 |
-
with col2:
|
410 |
-
st.write("### Tables")
|
411 |
-
if tables:
|
412 |
-
for table in tables:
|
413 |
-
st.image(table, use_column_width=True)
|
414 |
-
else:
|
415 |
-
st.write("No tables found.")
|
416 |
-
|
417 |
-
with st.container():
|
418 |
-
st.markdown("<div class='section-header'>Get Summary</div>", unsafe_allow_html=True)
|
419 |
-
with st.expander("Click to Generate Summary", expanded=True):
|
420 |
-
with st.container():
|
421 |
-
summary_button = st.button("Generate Summary")
|
422 |
-
if summary_button:
|
423 |
-
summary = summarize_pdf(file_path)
|
424 |
-
st.markdown(summary, unsafe_allow_html=True)
|
425 |
-
|
426 |
-
with st.container():
|
427 |
-
st.markdown("<div class='section-header'>Chat with your PDF</div>", unsafe_allow_html=True)
|
428 |
-
st.write("### Chat with your PDF")
|
429 |
-
if 'chat_history' not in st.session_state:
|
430 |
-
st.session_state['chat_history'] = []
|
431 |
-
|
432 |
-
for chat in st.session_state['chat_history']:
|
433 |
-
chat_user_class = "user" if chat["user"] else ""
|
434 |
-
chat_bot_class = "bot" if chat["bot"] else ""
|
435 |
-
st.markdown(f"<div class='chat-message {chat_user_class}'>{chat['user']}</div>", unsafe_allow_html=True)
|
436 |
-
st.markdown(f"<div class='chat-message {chat_bot_class}'>{chat['bot']}</div>", unsafe_allow_html=True)
|
437 |
-
|
438 |
-
with st.form(key="chat_form", clear_on_submit=True):
|
439 |
-
user_input = st.text_area("Ask a question about the PDF:", key="user_input")
|
440 |
-
submit_button = st.form_submit_button(label="Send")
|
441 |
-
|
442 |
-
if submit_button and user_input:
|
443 |
-
st.session_state['chat_history'].append({"user": user_input, "bot": None})
|
444 |
-
answer = qa_pdf(file_path, user_input)
|
445 |
-
st.session_state['chat_history'][-1]["bot"] = answer
|
446 |
st.experimental_rerun()
|
|
|
1 |
+
|
2 |
+
|
3 |
+
|
4 |
+
import os
|
5 |
+
import streamlit as st
|
6 |
+
import numpy as np
|
7 |
+
import fitz # PyMuPDF
|
8 |
+
from ultralytics import YOLO
|
9 |
+
from sklearn.cluster import KMeans
|
10 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
11 |
+
from langchain_core.output_parsers import StrOutputParser
|
12 |
+
from langchain_community.document_loaders import PyMuPDFLoader
|
13 |
+
from langchain_openai import OpenAIEmbeddings
|
14 |
+
from langchain_text_splitters import RecursiveCharacterTextSplitter
|
15 |
+
from langchain_core.prompts import ChatPromptTemplate
|
16 |
+
from sklearn.decomposition import PCA
|
17 |
+
from langchain_openai import ChatOpenAI
|
18 |
+
import string
|
19 |
+
import re
|
20 |
+
|
21 |
+
|
22 |
+
# Load the trained model
|
23 |
+
model = YOLO("runs\\detect\\train7\\weights\\best.pt")
|
24 |
+
openai_api_key = os.environ.get("openai_api_key")
|
25 |
+
|
26 |
+
# Define the class indices for figures, tables, and text
|
27 |
+
figure_class_index = 4 # class index for figures
|
28 |
+
table_class_index = 3 # class index for tables
|
29 |
+
|
30 |
+
# Global variables to store embeddings and contents
|
31 |
+
global_embeddings = None
|
32 |
+
global_split_contents = None
|
33 |
+
|
34 |
+
def clean_text(text):
|
35 |
+
text = re.sub(r'\s+', ' ', text).strip()
|
36 |
+
return text
|
37 |
+
|
38 |
+
def remove_references(text):
|
39 |
+
reference_patterns = [
|
40 |
+
r'\bReferences\b', r'\breferences\b', r'\bBibliography\b', r'\bCitations\b',
|
41 |
+
r'\bWorks Cited\b', r'\bReference\b', r'\breference\b'
|
42 |
+
]
|
43 |
+
lines = text.split('\n')
|
44 |
+
for i, line in enumerate(lines):
|
45 |
+
if any(re.search(pattern, line, re.IGNORECASE) for pattern in reference_patterns):
|
46 |
+
return '\n'.join(lines[:i])
|
47 |
+
return text
|
48 |
+
|
49 |
+
def save_uploaded_file(uploaded_file):
|
50 |
+
with open(uploaded_file.name, 'wb') as f:
|
51 |
+
f.write(uploaded_file.getbuffer())
|
52 |
+
return uploaded_file.name
|
53 |
+
|
54 |
+
def summarize_pdf(pdf_file_path, num_clusters=10):
|
55 |
+
embeddings_model = OpenAIEmbeddings(model="text-embedding-3-small", api_key=openai_api_key)
|
56 |
+
llm = ChatOpenAI(model="gpt-3.5-turbo", api_key=openai_api_key, temperature=0.3)
|
57 |
+
prompt = ChatPromptTemplate.from_template(
|
58 |
+
"""Could you please provide a concise and comprehensive summary of the given Contexts?
|
59 |
+
The summary should capture the main points and key details of the text while conveying the author's intended meaning accurately.
|
60 |
+
Please ensure that the summary is well-organized and easy to read, with clear headings and subheadings to guide the reader through each section.
|
61 |
+
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.
|
62 |
+
example of summary:
|
63 |
+
## Summary:
|
64 |
+
## Key points:
|
65 |
+
Contexts: {topic}"""
|
66 |
+
)
|
67 |
+
output_parser = StrOutputParser()
|
68 |
+
chain = prompt | llm | output_parser
|
69 |
+
|
70 |
+
loader = PyMuPDFLoader(pdf_file_path)
|
71 |
+
docs = loader.load()
|
72 |
+
full_text = "\n".join(doc.page_content for doc in docs)
|
73 |
+
cleaned_full_text = remove_references(full_text)
|
74 |
+
cleaned_full_text = clean_text(cleaned_full_text)
|
75 |
+
|
76 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0,separators=["\n\n", "\n",".", " "])
|
77 |
+
split_contents = text_splitter.split_text(cleaned_full_text)
|
78 |
+
embeddings = embeddings_model.embed_documents(split_contents)
|
79 |
+
|
80 |
+
X = np.array(embeddings)
|
81 |
+
kmeans = KMeans(n_clusters=num_clusters, init='k-means++', random_state=0).fit(embeddings)
|
82 |
+
cluster_centers = kmeans.cluster_centers_
|
83 |
+
|
84 |
+
closest_point_indices = []
|
85 |
+
for center in cluster_centers:
|
86 |
+
distances = np.linalg.norm(embeddings - center, axis=1)
|
87 |
+
closest_point_indices.append(np.argmin(distances))
|
88 |
+
|
89 |
+
extracted_contents = [split_contents[idx] for idx in closest_point_indices]
|
90 |
+
results = chain.invoke({"topic": ' '.join(extracted_contents)})
|
91 |
+
|
92 |
+
summary_sentences = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s', results)
|
93 |
+
summary_embeddings = embeddings_model.embed_documents(summary_sentences)
|
94 |
+
extracted_embeddings = embeddings_model.embed_documents(extracted_contents)
|
95 |
+
similarity_matrix = cosine_similarity(summary_embeddings, extracted_embeddings)
|
96 |
+
|
97 |
+
cited_results = results
|
98 |
+
relevant_sources = []
|
99 |
+
source_mapping = {}
|
100 |
+
sentence_to_source = {}
|
101 |
+
similarity_threshold = 0.6
|
102 |
+
|
103 |
+
for i, sentence in enumerate(summary_sentences):
|
104 |
+
if sentence in sentence_to_source:
|
105 |
+
continue
|
106 |
+
max_similarity = max(similarity_matrix[i])
|
107 |
+
if max_similarity >= similarity_threshold:
|
108 |
+
most_similar_idx = np.argmax(similarity_matrix[i])
|
109 |
+
if most_similar_idx not in source_mapping:
|
110 |
+
source_mapping[most_similar_idx] = len(relevant_sources) + 1
|
111 |
+
relevant_sources.append((most_similar_idx, extracted_contents[most_similar_idx]))
|
112 |
+
citation_idx = source_mapping[most_similar_idx]
|
113 |
+
citation = f"([Source {citation_idx}](#source-{citation_idx}))"
|
114 |
+
cited_sentence = re.sub(r'([.!?])$', f" {citation}\\1", sentence)
|
115 |
+
sentence_to_source[sentence] = citation_idx
|
116 |
+
cited_results = cited_results.replace(sentence, cited_sentence)
|
117 |
+
|
118 |
+
sources_list = "\n\n## Sources:\n"
|
119 |
+
for idx, (original_idx, content) in enumerate(relevant_sources):
|
120 |
+
sources_list += f"""
|
121 |
+
<details style="margin: 10px 0; padding: 10px; border: 1px solid #ccc; border-radius: 5px; background-color: #f9f9f9;">
|
122 |
+
<summary style="font-weight: bold; cursor: pointer;">Source {idx + 1}</summary>
|
123 |
+
<pre style="white-space: pre-wrap; word-wrap: break-word; margin-top: 10px;">{content}</pre>
|
124 |
+
</details>
|
125 |
+
"""
|
126 |
+
cited_results += sources_list
|
127 |
+
return cited_results
|
128 |
+
|
129 |
+
def qa_pdf(pdf_file_path, query, num_clusters=5, similarity_threshold=0.6):
|
130 |
+
global global_embeddings, global_split_contents
|
131 |
+
|
132 |
+
# Initialize models and embeddings
|
133 |
+
embeddings_model = OpenAIEmbeddings(model="text-embedding-3-small", api_key=openai_api_key)
|
134 |
+
llm = ChatOpenAI(model="gpt-3.5-turbo", api_key=openai_api_key, temperature=0.3)
|
135 |
+
prompt = ChatPromptTemplate.from_template(
|
136 |
+
"""Please provide a detailed and accurate answer to the given question based on the provided contexts.
|
137 |
+
Ensure that the answer is comprehensive and directly addresses the query.
|
138 |
+
If necessary, include relevant examples or details from the text.
|
139 |
+
Question: {question}
|
140 |
+
Contexts: {contexts}"""
|
141 |
+
)
|
142 |
+
output_parser = StrOutputParser()
|
143 |
+
chain = prompt | llm | output_parser
|
144 |
+
|
145 |
+
# Load and process the PDF if not already loaded
|
146 |
+
if global_embeddings is None or global_split_contents is None:
|
147 |
+
loader = PyMuPDFLoader(pdf_file_path)
|
148 |
+
docs = loader.load()
|
149 |
+
full_text = "\n".join(doc.page_content for doc in docs)
|
150 |
+
cleaned_full_text = remove_references(full_text)
|
151 |
+
cleaned_full_text = clean_text(cleaned_full_text)
|
152 |
+
|
153 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=800, chunk_overlap=0, separators=["\n\n", "\n", ".", " "])
|
154 |
+
global_split_contents = text_splitter.split_text(cleaned_full_text)
|
155 |
+
global_embeddings = embeddings_model.embed_documents(global_split_contents)
|
156 |
+
|
157 |
+
# Embed the query and find the most relevant contexts
|
158 |
+
query_embedding = embeddings_model.embed_query(query)
|
159 |
+
similarity_scores = cosine_similarity([query_embedding], global_embeddings)[0]
|
160 |
+
top_indices = np.argsort(similarity_scores)[-num_clusters:]
|
161 |
+
relevant_contents = [global_split_contents[i] for i in top_indices]
|
162 |
+
|
163 |
+
# Generate the answer using the LLM chain
|
164 |
+
results = chain.invoke({"question": query, "contexts": ' '.join(relevant_contents)})
|
165 |
+
|
166 |
+
# Split the answer into sentences and embed them
|
167 |
+
answer_sentences = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s', results)
|
168 |
+
answer_embeddings = embeddings_model.embed_documents(answer_sentences)
|
169 |
+
relevant_embeddings = embeddings_model.embed_documents(relevant_contents)
|
170 |
+
similarity_matrix = cosine_similarity(answer_embeddings, relevant_embeddings)
|
171 |
+
|
172 |
+
# Map sentences to sources and create citations
|
173 |
+
cited_results = results
|
174 |
+
relevant_sources = []
|
175 |
+
source_mapping = {}
|
176 |
+
sentence_to_source = {}
|
177 |
+
|
178 |
+
for i, sentence in enumerate(answer_sentences):
|
179 |
+
if sentence in sentence_to_source:
|
180 |
+
continue
|
181 |
+
max_similarity = max(similarity_matrix[i])
|
182 |
+
if max_similarity >= similarity_threshold:
|
183 |
+
most_similar_idx = np.argmax(similarity_matrix[i])
|
184 |
+
if most_similar_idx not in source_mapping:
|
185 |
+
source_mapping[most_similar_idx] = len(relevant_sources) + 1
|
186 |
+
relevant_sources.append((most_similar_idx, relevant_contents[most_similar_idx]))
|
187 |
+
citation_idx = source_mapping[most_similar_idx]
|
188 |
+
citation = f"<strong style='color:blue;'>[Source {citation_idx}]</strong>"
|
189 |
+
cited_sentence = re.sub(r'([.!?])$', f" {citation}\\1", sentence)
|
190 |
+
sentence_to_source[sentence] = citation_idx
|
191 |
+
cited_results = cited_results.replace(sentence, cited_sentence)
|
192 |
+
|
193 |
+
# Format the sources for markdown rendering
|
194 |
+
sources_list = "\n\n## Sources:\n"
|
195 |
+
for idx, (original_idx, content) in enumerate(relevant_sources):
|
196 |
+
sources_list += f"""
|
197 |
+
<details style="margin: 10px 0; padding: 10px; border: 1px solid #ccc; border-radius: 5px; background-color: #f9f9f9;">
|
198 |
+
<summary style="font-weight: bold; cursor: pointer;">Source {idx + 1}</summary>
|
199 |
+
<pre style="white-space: pre-wrap; word-wrap: break-word; margin-top: 10px;">{content}</pre>
|
200 |
+
</details>
|
201 |
+
"""
|
202 |
+
cited_results += sources_list
|
203 |
+
return cited_results
|
204 |
+
|
205 |
+
|
206 |
+
def infer_image_and_get_boxes(image, confidence_threshold=0.6):
|
207 |
+
results = model.predict(image)
|
208 |
+
boxes = [
|
209 |
+
(int(box.xyxy[0][0]), int(box.xyxy[0][1]), int(box.xyxy[0][2]), int(box.xyxy[0][3]), int(box.cls[0]))
|
210 |
+
for result in results for box in result.boxes
|
211 |
+
if int(box.cls[0]) in {figure_class_index, table_class_index} and box.conf[0] > confidence_threshold
|
212 |
+
]
|
213 |
+
return boxes
|
214 |
+
|
215 |
+
def crop_images_from_boxes(image, boxes, scale_factor):
|
216 |
+
figures = []
|
217 |
+
tables = []
|
218 |
+
for (x1, y1, x2, y2, cls) in boxes:
|
219 |
+
cropped_img = image[int(y1 * scale_factor):int(y2 * scale_factor), int(x1 * scale_factor):int(x2 * scale_factor)]
|
220 |
+
if cls == figure_class_index:
|
221 |
+
figures.append(cropped_img)
|
222 |
+
elif cls == table_class_index:
|
223 |
+
tables.append(cropped_img)
|
224 |
+
return figures, tables
|
225 |
+
|
226 |
+
|
227 |
+
def process_pdf(pdf_file_path):
|
228 |
+
doc = fitz.open(pdf_file_path)
|
229 |
+
all_figures = []
|
230 |
+
all_tables = []
|
231 |
+
low_dpi = 50
|
232 |
+
high_dpi = 300
|
233 |
+
scale_factor = high_dpi / low_dpi
|
234 |
+
low_res_pixmaps = [page.get_pixmap(dpi=low_dpi) for page in doc]
|
235 |
+
|
236 |
+
for page_num, low_res_pix in enumerate(low_res_pixmaps):
|
237 |
+
low_res_img = np.frombuffer(low_res_pix.samples, dtype=np.uint8).reshape(low_res_pix.height, low_res_pix.width, 3)
|
238 |
+
boxes = infer_image_and_get_boxes(low_res_img)
|
239 |
+
|
240 |
+
if boxes:
|
241 |
+
high_res_pix = doc[page_num].get_pixmap(dpi=high_dpi)
|
242 |
+
high_res_img = np.frombuffer(high_res_pix.samples, dtype=np.uint8).reshape(high_res_pix.height, high_res_pix.width, 3)
|
243 |
+
figures, tables = crop_images_from_boxes(high_res_img, boxes, scale_factor)
|
244 |
+
all_figures.extend(figures)
|
245 |
+
all_tables.extend(tables)
|
246 |
+
|
247 |
+
return all_figures, all_tables
|
248 |
+
|
249 |
+
# Set the page configuration for a modern look
|
250 |
+
|
251 |
+
# Set the page configuration for a modern look
|
252 |
+
# Set the page configuration for a modern look
|
253 |
+
st.set_page_config(page_title="PDF Reading Assistant", page_icon="π", layout="wide")
|
254 |
+
|
255 |
+
# Add some custom CSS for a modern look
|
256 |
+
st.markdown("""
|
257 |
+
<style>
|
258 |
+
/* Main background and padding */
|
259 |
+
.main {
|
260 |
+
background-color: #f8f9fa;
|
261 |
+
padding: 2rem;
|
262 |
+
font-family: 'Arial', sans-serif;
|
263 |
+
}
|
264 |
+
|
265 |
+
/* Section headers */
|
266 |
+
.section-header {
|
267 |
+
font-size: 2rem;
|
268 |
+
font-weight: bold;
|
269 |
+
color: #343a40;
|
270 |
+
margin-top: 2rem;
|
271 |
+
margin-bottom: 1rem;
|
272 |
+
text-align: center;
|
273 |
+
font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
|
274 |
+
}
|
275 |
+
|
276 |
+
/* Containers */
|
277 |
+
.uploaded-file-container, .chat-container, .summary-container, .extract-container {
|
278 |
+
padding: 2rem;
|
279 |
+
background-color: #ffffff;
|
280 |
+
border-radius: 10px;
|
281 |
+
margin-bottom: 2rem;
|
282 |
+
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.1);
|
283 |
+
}
|
284 |
+
|
285 |
+
/* Buttons */
|
286 |
+
.stButton>button {
|
287 |
+
background-color: #007bff;
|
288 |
+
color: white;
|
289 |
+
padding: 0.6rem 1.2rem;
|
290 |
+
border-radius: 5px;
|
291 |
+
border: none;
|
292 |
+
cursor: pointer;
|
293 |
+
font-size: 1rem;
|
294 |
+
transition: background-color 0.3s ease, transform 0.3s ease;
|
295 |
+
}
|
296 |
+
.stButton>button:hover {
|
297 |
+
background-color: #0056b3;
|
298 |
+
transform: translateY(-2px);
|
299 |
+
}
|
300 |
+
|
301 |
+
/* Chat messages */
|
302 |
+
.chat-message {
|
303 |
+
padding: 1rem;
|
304 |
+
border-radius: 10px;
|
305 |
+
margin-bottom: 1rem;
|
306 |
+
font-size: 1rem;
|
307 |
+
transition: all 0.3s ease;
|
308 |
+
box-shadow: 0 2px 8px rgba(0, 0, 0, 0.1);
|
309 |
+
}
|
310 |
+
.chat-message.user {
|
311 |
+
background-color: #e6f7ff;
|
312 |
+
border-left: 5px solid #007bff;
|
313 |
+
text-align: left;
|
314 |
+
}
|
315 |
+
.chat-message.bot {
|
316 |
+
background-color: #fff0f1;
|
317 |
+
border-left: 5px solid #dc3545;
|
318 |
+
text-align: left;
|
319 |
+
}
|
320 |
+
|
321 |
+
/* Input area */
|
322 |
+
.input-container {
|
323 |
+
display: flex;
|
324 |
+
align-items: center;
|
325 |
+
gap: 10px;
|
326 |
+
margin-top: 1rem;
|
327 |
+
}
|
328 |
+
.input-container textarea {
|
329 |
+
border: 2px solid #ccc;
|
330 |
+
border-radius: 10px;
|
331 |
+
padding: 10px;
|
332 |
+
width: 100%;
|
333 |
+
background-color: #fff;
|
334 |
+
transition: border-color 0.3s ease;
|
335 |
+
margin: 0;
|
336 |
+
font-size: 1rem;
|
337 |
+
}
|
338 |
+
.input-container textarea:focus {
|
339 |
+
border-color: #007bff;
|
340 |
+
outline: none;
|
341 |
+
}
|
342 |
+
.input-container button {
|
343 |
+
background-color: #007bff;
|
344 |
+
color: white;
|
345 |
+
padding: 0.6rem 1.2rem;
|
346 |
+
border-radius: 5px;
|
347 |
+
border: none;
|
348 |
+
cursor: pointer;
|
349 |
+
font-size: 1rem;
|
350 |
+
transition: background-color 0.3s ease, transform 0.3s ease;
|
351 |
+
}
|
352 |
+
.input-container button:hover {
|
353 |
+
background-color: #0056b3;
|
354 |
+
transform: translateY(-2px);
|
355 |
+
}
|
356 |
+
|
357 |
+
/* Expander */
|
358 |
+
.st-expander {
|
359 |
+
border: none;
|
360 |
+
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.1);
|
361 |
+
margin-bottom: 2rem;
|
362 |
+
}
|
363 |
+
|
364 |
+
/* Markdown elements */
|
365 |
+
.stMarkdown {
|
366 |
+
font-size: 1rem;
|
367 |
+
color: #343a40;
|
368 |
+
line-height: 1.6;
|
369 |
+
}
|
370 |
+
|
371 |
+
/* Titles and subtitles */
|
372 |
+
.stTitle {
|
373 |
+
color: #343a40;
|
374 |
+
text-align: center;
|
375 |
+
margin-bottom: 1rem;
|
376 |
+
font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
|
377 |
+
}
|
378 |
+
.stSubtitle {
|
379 |
+
color: #6c757d;
|
380 |
+
text-align: center;
|
381 |
+
margin-bottom: 1rem;
|
382 |
+
font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
|
383 |
+
}
|
384 |
+
</style>
|
385 |
+
""", unsafe_allow_html=True)
|
386 |
+
|
387 |
+
# Streamlit interface
|
388 |
+
st.title("π PDF Reading Assistant")
|
389 |
+
st.markdown("### Extract tables, figures, summaries, and answers from your PDF files easily.")
|
390 |
+
|
391 |
+
uploaded_file = st.file_uploader("Upload a PDF", type="pdf")
|
392 |
+
if uploaded_file:
|
393 |
+
file_path = save_uploaded_file(uploaded_file)
|
394 |
+
with st.container():
|
395 |
+
st.markdown("<div class='section-header'>Extract Tables and Figures</div>", unsafe_allow_html=True)
|
396 |
+
with st.expander("Click to Extract Tables and Figures", expanded=True):
|
397 |
+
with st.container():
|
398 |
+
extract_button = st.button("Extract")
|
399 |
+
if extract_button:
|
400 |
+
figures, tables = process_pdf(file_path)
|
401 |
+
col1, col2 = st.columns(2)
|
402 |
+
with col1:
|
403 |
+
st.write("### Figures")
|
404 |
+
if figures:
|
405 |
+
for figure in figures:
|
406 |
+
st.image(figure, use_column_width=True)
|
407 |
+
else:
|
408 |
+
st.write("No figures found.")
|
409 |
+
with col2:
|
410 |
+
st.write("### Tables")
|
411 |
+
if tables:
|
412 |
+
for table in tables:
|
413 |
+
st.image(table, use_column_width=True)
|
414 |
+
else:
|
415 |
+
st.write("No tables found.")
|
416 |
+
|
417 |
+
with st.container():
|
418 |
+
st.markdown("<div class='section-header'>Get Summary</div>", unsafe_allow_html=True)
|
419 |
+
with st.expander("Click to Generate Summary", expanded=True):
|
420 |
+
with st.container():
|
421 |
+
summary_button = st.button("Generate Summary")
|
422 |
+
if summary_button:
|
423 |
+
summary = summarize_pdf(file_path)
|
424 |
+
st.markdown(summary, unsafe_allow_html=True)
|
425 |
+
|
426 |
+
with st.container():
|
427 |
+
st.markdown("<div class='section-header'>Chat with your PDF</div>", unsafe_allow_html=True)
|
428 |
+
st.write("### Chat with your PDF")
|
429 |
+
if 'chat_history' not in st.session_state:
|
430 |
+
st.session_state['chat_history'] = []
|
431 |
+
|
432 |
+
for chat in st.session_state['chat_history']:
|
433 |
+
chat_user_class = "user" if chat["user"] else ""
|
434 |
+
chat_bot_class = "bot" if chat["bot"] else ""
|
435 |
+
st.markdown(f"<div class='chat-message {chat_user_class}'>{chat['user']}</div>", unsafe_allow_html=True)
|
436 |
+
st.markdown(f"<div class='chat-message {chat_bot_class}'>{chat['bot']}</div>", unsafe_allow_html=True)
|
437 |
+
|
438 |
+
with st.form(key="chat_form", clear_on_submit=True):
|
439 |
+
user_input = st.text_area("Ask a question about the PDF:", key="user_input")
|
440 |
+
submit_button = st.form_submit_button(label="Send")
|
441 |
+
|
442 |
+
if submit_button and user_input:
|
443 |
+
st.session_state['chat_history'].append({"user": user_input, "bot": None})
|
444 |
+
answer = qa_pdf(file_path, user_input)
|
445 |
+
st.session_state['chat_history'][-1]["bot"] = answer
|
446 |
st.experimental_rerun()
|