-
Notifications
You must be signed in to change notification settings - Fork 0
/
ingestor.py
38 lines (29 loc) · 1.51 KB
/
ingestor.py
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
import os
import warnings
import asyncio
from tqdm import tqdm
from langchain_community.vectorstores.chroma import Chroma
from langchain_community.document_loaders import (PyPDFLoader, DirectoryLoader)
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.embeddings import OllamaEmbeddings
warnings.simplefilter(action='ignore')
async def process_document(doc):
textSplitter = RecursiveCharacterTextSplitter(chunk_size=2000, chunk_overlap=200)
chunkedDocuments = textSplitter.split_documents([doc])
content = [chunk.page_content for chunk in chunkedDocuments]
metadatas = [chunk.metadata for chunk in chunkedDocuments]
return content, metadatas
async def create_vector_database():
pdfDirecLoader = DirectoryLoader("./files/", glob="*.pdf", loader_cls=PyPDFLoader)
loadedDocuments = pdfDirecLoader.load()
print(f"Loaded {len(loadedDocuments)} documents.")
# Process documents in parallel
results = await asyncio.gather(*[process_document(doc) for doc in loadedDocuments])
# Flatten results
content = [item for sublist in results for item in sublist[0]]
metadatas = [item for sublist in results for item in sublist[1]]
ollama_embeddings = OllamaEmbeddings(model="nomic-embed-text", show_progress=True, num_gpu=1)
vectorDB = Chroma.from_texts(texts=content, embedding=ollama_embeddings, metadatas=metadatas, persist_directory="./data")
vectorDB.persist()
if __name__ == "__main__":
asyncio.run(create_vector_database())