Spaces:
Sleeping
Sleeping
File size: 6,336 Bytes
c2a146f |
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 |
# Imports for Transcript Loader
import os
import webvtt
import re
from datetime import datetime
from llama_index import Document
# Imports for Document Embedder
import gc
import re
class VTTTranscriptLoader:
"""
vtt file ingestion and cleaning. This was done because vtt files
are not recognized by llamaindex. The output should mirror that of
any document loader from llamaindex or langchain.
"""
def __init__(self, file_path):
self.fp = file_path
self.data = None
def open_vtt(self, file_path, plaintext=True):
"""Read VTT file."""
if plaintext:
with open(file_path, "r") as f:
data = f.readlines()
else:
data = webvtt.read(file_path)
return data
def extract_speaker_name(self, text):
"""Extracts the speaker name from a VTT caption."""
match = re.search(r"<v (.*?)>", text)
if match:
return match.group(1)
else:
return None
def extract_speaker_words(self, captions):
"""Extracts the speaker text from a VTT caption."""
return [caption.text for caption in captions]
def merge_speaker_words(self, words, speakers, split=True):
"""Joins speaker names with their words."""
# Extract speaker names
speaker_list = [self.extract_speaker_name(line) for line in speakers if self.extract_speaker_name(line)]
# Extract words
words_list = self.extract_speaker_words(words)
# Combine speaker names and words
combined_list = list(zip(speaker_list, words_list))
# Return the combined list as a single string if split is False
if not split:
combined_list = '\n'.join([f"{name}: '{text}'" for name, text in combined_list])
return combined_list, speaker_list
def get_metadata(self, speaker_list, file_path):
"""Generates metadata for the transcript."""
# Meeting length
time_format = "%H:%M:%S.%f"
sess = self.open_vtt(file_path, plaintext=False)
dt1 = datetime.strptime(sess[0].start, time_format)
dt2 = datetime.strptime(sess[-1].end, time_format)
minutes = (dt2 - dt1).seconds / 60
# Meeting date
match = re.search(r"\d{4}[-_]\d{2}[-_]\d{2}", file_path)
if match:
date_str = match.group().replace('_', '-')
date_obj = datetime.strptime(date_str, "%Y-%m-%d").date()
else:
date_obj = None
# Pull dictionary here
output = {
'title': file_path,
'duration': minutes,
'meeting_date': date_obj.strftime("%Y-%m-%d"),
'speakers': list(set(speaker_list)),
}
return output
def manual_document(self, output, metadata):
"""Create document manually"""
document = Document(text=output)
document.metadata = metadata
return document
def process_file(self, file_path):
"""Processes a single VTT file and returns the combined speaker names and words."""
# Get words as webvtt captions
words = self.open_vtt(file_path, plaintext=False)
# Get speaker lines as plaintext
speaker = self.open_vtt(file_path, plaintext=True)
# Combine speaker names and words
output, speaker_list = self.merge_speaker_words(words, speaker, split=False)
# Get session data as dictionary
metadata = self.get_metadata(speaker_list, file_path)
return self.manual_document(output, metadata)
def load(self):
"""Processes all VTT files in the given list of directories or files and returns a list of results."""
results = []
for path in self.fp:
if os.path.isdir(path):
for root, _, files in os.walk(path):
for file in files:
if file.endswith('.vtt'):
file_path = os.path.join(root, file)
transcript = self.process_file(file_path)
results.append(transcript)
else:
if path.endswith('.vtt'):
transcript = self.process_file(path)
results.append(transcript)
return results
class DocumentEmbedder:
"""
Process retrieves, cleans, embeds, and sends the documents to vector
store.
Currently supports hugginface embeddings only. Gotta keep things cheap.
"""
def __init__(self, pinecone_pipeline, files, embedder):
# pinecone
self.pipeline = pinecone_pipeline
# basic items
self.files = files
#self.interactive = interactive
def clean_text(self, content: str) -> str:
"""
Remove unwanted characters and patterns in text input.
:param content: Text input.
:return: Cleaned version of original text input.
"""
# Fix hyphenated words broken by newline
content = re.sub(r'(\w+)-\n(\w+)', r'\1\2', content)
# Remove specific unwanted patterns and characters
unwanted_patterns = [
"\\n", " —", "——————————", "—————————", "—————",
r'\\u[\dA-Fa-f]{4}', r'\uf075', r'\uf0b7'
]
for pattern in unwanted_patterns:
content = re.sub(pattern, "", content)
# Fix improperly spaced hyphenated words and normalize whitespace
content = re.sub(r'(\w)\s*-\s*(\w)', r'\1-\2', content)
content = re.sub(r'\s+', ' ', content)
return content
def embed(self):
"""stringing process above to embed and upsert directly to pinecone"""
# read_file
print("reading files")
results = self.files
# Call clean function
print("cleaning files")
for d in range(len(results)):
results[d].text = self.clean_text(results[d].text)
# run pinecone in batches (of 1) for memory preservation.
print("reading into pinecone db")
batchsize = 1
for i in range(0, len(results), batchsize):
gc.collect()
batch = self.pipeline.run(documents=results[i:i+batchsize])
print("completed batch %s" % ((i+batchsize)/batchsize))
|