File size: 8,670 Bytes
5d4f783
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
"""
This module provides functions for generating a highlighted PDF with important sentences.

The main function, `generate_highlighted_pdf`, takes an input PDF file and a pre-trained
sentence embedding model as input.

It splits the text of the PDF into sentences, computes sentence embeddings, and builds a
graph based on the cosine similarity between embeddings and at the same time split the
sentences to different clusters using clustering.

The sentences are then ranked using PageRank scores and a the middle of the cluster,
and important sentences are selected based on a threshold and clustering.

Finally, the selected sentences are highlighted in the PDF and the highlighted PDF content
is returned.

Other utility functions in this module include functions for loading a sentence embedding
model, encoding sentences, computing similarity matrices,building graphs, ranking sentences,
clustering sentence embeddings, and splitting text into sentences.

Note: This module requires the PyMuPDF, networkx, numpy, torch, sentence_transformers, and
sklearn libraries to be installed.
"""

import logging
from typing import BinaryIO, List, Tuple

import fitz  # PyMuPDF
import networkx as nx
import numpy as np
import torch
import torch.nn.functional as F
from sentence_transformers import SentenceTransformer
from sklearn.cluster import KMeans

# Constants
MAX_PAGE = 40
MAX_SENTENCES = 2000
PAGERANK_THRESHOLD_RATIO = 0.15
NUM_CLUSTERS_RATIO = 0.05
MIN_WORDS = 10

# Logger configuration
logging.basicConfig(level=logging.ERROR)
logger = logging.getLogger(__name__)


def load_sentence_model(revision: str = None) -> SentenceTransformer:
    """
    Load a pre-trained sentence embedding model.

    Args:
        revision (str): Optional parameter to specify the model revision.

    Returns:
        SentenceTransformer: A pre-trained sentence embedding model.
    """
    return SentenceTransformer("avsolatorio/GIST-Embedding-v0", revision=revision)


def encode_sentence(model: SentenceTransformer, sentence: str) -> torch.Tensor:
    """
    Encode a sentence into a fixed-dimensional vector representation.

    Args:
        model (SentenceTransformer): A pre-trained sentence embedding model.
        sentence (str): Input sentence.

    Returns:
        torch.Tensor: Encoded sentence vector.
    """

    model.eval()  # Set the model to evaluation mode

    # Check if GPU is available
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    with torch.no_grad():  # Disable gradient tracking
        return model.encode(sentence, convert_to_tensor=True).to(device)


def compute_similarity_matrix(embeddings: torch.Tensor) -> np.ndarray:
    """
    Compute the cosine similarity matrix between sentence embeddings.

    Args:
        embeddings (torch.Tensor): Sentence embeddings.

    Returns:
        np.ndarray: Cosine similarity matrix.
    """
    scores = F.cosine_similarity(
        embeddings.unsqueeze(1), embeddings.unsqueeze(0), dim=-1
    )
    similarity_matrix = scores.cpu().numpy()
    normalized_adjacency_matrix = similarity_matrix / similarity_matrix.sum(
        axis=1, keepdims=True
    )
    return normalized_adjacency_matrix


def build_graph(normalized_adjacency_matrix: np.ndarray) -> nx.DiGraph:
    """
    Build a directed graph from a normalized adjacency matrix.

    Args:
        normalized_adjacency_matrix (np.ndarray): Normalized adjacency matrix.

    Returns:
        nx.DiGraph: Directed graph.
    """
    return nx.DiGraph(normalized_adjacency_matrix)


def rank_sentences(graph: nx.DiGraph, sentences: List[str]) -> List[Tuple[str, float]]:
    """
    Rank sentences based on PageRank scores.

    Args:
        graph (nx.DiGraph): Directed graph.
        sentences (List[str]): List of sentences.

    Returns:
        List[Tuple[str, float]]: Ranked sentences with their PageRank scores.
    """
    pagerank_scores = nx.pagerank(graph)
    ranked_sentences = sorted(
        zip(sentences, pagerank_scores.values()),
        key=lambda x: x[1],
        reverse=True,
    )
    return ranked_sentences


def cluster_sentences(
    embeddings: torch.Tensor, num_clusters: int
) -> Tuple[np.ndarray, np.ndarray]:
    """
    Cluster sentence embeddings using K-means clustering.

    Args:
        embeddings (torch.Tensor): Sentence embeddings.
        num_clusters (int): Number of clusters.

    Returns:
        Tuple[np.ndarray, np.ndarray]: Cluster assignments and cluster centers.
    """
    kmeans = KMeans(n_clusters=num_clusters, random_state=42)
    cluster_assignments = kmeans.fit_predict(embeddings.cpu())
    cluster_centers = kmeans.cluster_centers_
    return cluster_assignments, cluster_centers


def get_middle_sentence(cluster_indices: np.ndarray, sentences: List[str]) -> List[str]:
    """
    Get the middle sentence from each cluster.

    Args:
        cluster_indices (np.ndarray): Cluster assignments.
        sentences (List[str]): List of sentences.

    Returns:
        List[str]: Middle sentences from each cluster.
    """
    middle_indices = [
        int(np.median(np.where(cluster_indices == i)[0]))
        for i in range(max(cluster_indices) + 1)
    ]
    middle_sentences = [sentences[i] for i in middle_indices]
    return middle_sentences


def split_text_into_sentences(text: str, min_words: int = MIN_WORDS) -> List[str]:
    """
    Split text into sentences.

    Args:
        text (str): Input text.
        min_words (int): Minimum number of words for a valid sentence.

    Returns:
        List[str]: List of sentences.
    """
    sentences = []
    for s in text.split("."):
        s = s.strip()
        # filtering out short sentences and sentences that contain more than 40% digits
        if (
            s
            and len(s.split()) >= min_words
            and (sum(c.isdigit() for c in s) / len(s)) < 0.4
        ):
            sentences.append(s)
    return sentences


def extract_text_from_pages(doc):
    """Generator to yield text per page from the PDF, for memory efficiency for large PDFs."""
    for page_num in range(len(doc)):
        yield doc[page_num].get_text()


def generate_highlighted_pdf(
    input_pdf_file: BinaryIO, model=load_sentence_model()
) -> bytes:
    """
    Generate a highlighted PDF with important sentences.

    Args:
        input_pdf_file: Input PDF file object.
        model (SentenceTransformer): Pre-trained sentence embedding model.

    Returns:
        bytes: Highlighted PDF content.
    """
    with fitz.open(stream=input_pdf_file.read(), filetype="pdf") as doc:
        num_pages = doc.page_count

        if num_pages > MAX_PAGE:
            # It will show the error message for the user.
            return f"The PDF file exceeds the maximum limit of {MAX_PAGE} pages."

        sentences = []
        for page_text in extract_text_from_pages(doc):  # Memory efficient
            sentences.extend(split_text_into_sentences(page_text))

        len_sentences = len(sentences)

        print(len_sentences)

        if len_sentences > MAX_SENTENCES:
            # It will show the error message for the user.
            return (
                f"The PDF file exceeds the maximum limit of {MAX_SENTENCES} sentences."
            )

        embeddings = encode_sentence(model, sentences)
        similarity_matrix = compute_similarity_matrix(embeddings)
        graph = build_graph(similarity_matrix)
        ranked_sentences = rank_sentences(graph, sentences)

        pagerank_threshold = int(len(ranked_sentences) * PAGERANK_THRESHOLD_RATIO) + 1
        top_pagerank_sentences = [
            sentence[0] for sentence in ranked_sentences[:pagerank_threshold]
        ]

        num_clusters = int(len_sentences * NUM_CLUSTERS_RATIO) + 1
        cluster_assignments, _ = cluster_sentences(embeddings, num_clusters)

        center_sentences = get_middle_sentence(cluster_assignments, sentences)
        important_sentences = list(set(top_pagerank_sentences + center_sentences))

        for i in range(num_pages):
            try:
                page = doc[i]

                for sentence in important_sentences:
                    rects = page.search_for(sentence)
                    colors = (fitz.pdfcolor["yellow"], fitz.pdfcolor["green"])

                    for i, rect in enumerate(rects):
                        color = colors[i % 2]
                        annot = page.add_highlight_annot(rect)
                        annot.set_colors(stroke=color)
                        annot.update()
            except Exception as e:
                logger.error(f"Error processing page {i}: {e}")

        output_pdf = doc.write()

    return output_pdf