# 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 from llama_index.embeddings.huggingface import HuggingFaceEmbedding from pinecone.grpc import PineconeGRPC from pinecone import ServerlessSpec from llama_index.vector_stores import PineconeVectorStore from llama_index.node_parser import SemanticSplitterNodeParser from llama_index.ingestion import IngestionPipeline 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"", 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 directory or the single file and returns a list of results.""" results = [] if os.path.isdir(self.fp): for root, _, files in os.walk(self.fp): 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: transcript = self.process_file(self.fp) results.append(transcript) return results class DocumentEmbedder: """ Takes a document and embeds it directly into a pinecone data store. Process retrieves, cleans, embeds, and sends the documents to vector store. Currently supports hugginface embeddings only. Gotta keep things cheap. """ def __init__(self, api_keys, files, embedding, index_name): # api keys self.pinecone_api_key = api_keys['pinecone'] self.openai_api_key = api_keys['openai'] self.huggingface_api_key = api_keys['huggingface'] # pinecone self.embedding = embedding self.vector_db = index_name # 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 create_embedder(self): """Get the right embedding model""" embedding = HuggingFaceEmbedding(model_name=self.embedding) return embedding def pinecone_pipeline(self, embedding): """Initialize pinecone connection and vectorstore""" # connect pc = PineconeGRPC(api_key=self.pinecone_api_key) # Create your index if index does not exist indexes = [i.name for i in pc.list_indexes()] index_exists = any([self.vector_db in i for i in indexes]) if index_exists: print("Index already exists") else: print("Creating index") pc.create_index( self.vector_db, dimension=768, metric="cosine", spec=ServerlessSpec(cloud="aws", region="us-east-1"), ) # Initialize your index pinecone_index = pc.Index(self.vector_db) # Initialize VectorStore vector_store = PineconeVectorStore(pinecone_index=pinecone_index) # create pipeline (abstracts away the need to adaptively process and batch) pipeline = IngestionPipeline( transformations=[ # creating appropriate chunks and cutoffs (this needs to be worked on). SemanticSplitterNodeParser( buffer_size=10, # 1 = each sentence is a node breakpoint_percentile_threshold=95, embed_model=embedding, ), embedding, ], vector_store=vector_store ) return pipeline 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) # set up embedder print("retrieving embedder") embedder = self.create_embedder() # set up pinecone pipeline print("initializing pinecone db") pipeline = self.pinecone_pipeline(embedder) # 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 = pipeline.run(documents=results[i:i+batchsize]) print("completed batch %s" % ((i+batchsize)/batchsize))