marius/project_bookworm
1
0
Fork 0
mirror of https://github.com/norandom/project_bookworm.git synced 2024-11-21 16:23:44 +00:00
Code Issues Packages Projects Releases Wiki Activity
c422d14692
project_bookworm/EverNote_To_OpenAI.ipynb
Marius Ciepluch e56c87a71d moved from sqlite to faiss for concurrency reasons
2024-04-04 10:38:12 +00:00

825 lines
27 KiB
Plaintext
Raw Blame History

{
"cells": [
{
"cell_type": "markdown",
"id": "18d62071e34b0d53",
"metadata": {
"collapsed": false
},
"source": [
"# This is an experiment: create vectorized embeddings out of an EverNote DB (PDF, DOCX, HTML, TXT)\n",
"\n",
"## Features\n",
"\n",
"* vectorize text, html files, pdfs and docx into one vector DB, split in tables (sqlite vss)\n",
"* use local self-hosted embeddings (CPU or GPU computed)\n",
" * for sentences \n",
"* query a local sqlite vss vector db, use cache from LangChain (sqlite)\n",
"* use OpenAI API and (Ollama on-prem self-hosted) Mistral for the response processing\n",
"* compare with LLMware Bling \n",
"\n",
"## Anti-Features\n",
"\n",
"* due to cost reasons the OpenAI embeddings don't get used. So sorry :p"
]
},
{
"cell_type": "markdown",
"id": "94517a27e3148ff4",
"metadata": {
"collapsed": false
},
"source": [
"# Configuration"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "fd9747a54ea8fcef",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:35:54.949214Z",
"start_time": "2024-04-04T10:35:54.945013Z"
},
"collapsed": false
},
"outputs": [],
"source": [
"import sys\n",
"import os\n",
"IN_COLAB = 'google.colab' in sys.modules\n",
"\n",
"if not IN_COLAB:\n",
" # The Evernote DB path containing the extracted data.\n",
" extracted_evernote_db = \"/home/marius/data/it-sec-research-extracted/IT sec research\"\n",
"\n",
" # Output paths containing the Evernote text notes or documents data.\n",
" # These get generated by the data extraction process\n",
" output_path_extracted_notes = \"/home/marius/source/bookworm/export.txt\"\n",
" output_path_extracted_docs = \"/home/marius/source/bookworm/export.documents.txt\"\n",
"\n",
" # Resulting DB or vector store path.\n",
" result_db = \"/home/marius/source/bookworm/evernote.db\"\n",
"\n",
"else:\n",
" # For the Goog Colab env\n",
" output_path_extracted_notes = \"/content/export.txt\"\n",
" output_path_extracted_docs = \"/content/export.documents.txt\"\n",
" result_db = \"/content/evernote.db\"\n",
"\n",
"# To suppress some warnings\n",
"import os\n",
"os.environ[\"TOKENIZERS_PARALLELISM\"] = \"True\""
]
},
{
"cell_type": "markdown",
"id": "a8c8692786d83c00",
"metadata": {
"collapsed": false
},
"source": [
"## Dependencies\n",
"\n",
"* Cryptography is used to handle some PDF functions here (signatures)"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "bb34db1ea75a1edf",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:08:32.520341Z",
"start_time": "2024-04-04T10:08:30.353678Z"
},
"collapsed": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Name: cryptography\r\n",
"Version: 42.0.5\r\n",
"Summary: cryptography is a package which provides cryptographic recipes and primitives to Python developers.\r\n",
"Home-page: \r\n",
"Author: \r\n",
"Author-email: The Python Cryptographic Authority and individual contributors <cryptography-dev@python.org>\r\n",
"License: Apache-2.0 OR BSD-3-Clause\r\n",
"Location: /home/marius/miniconda3/envs/llm_langchain/lib/python3.11/site-packages\r\n",
"Requires: cffi\r\n",
"Required-by: \r\n",
"Note: you may need to restart the kernel to use updated packages.\n"
]
}
],
"source": [
"%pip show cryptography"
]
},
{
"cell_type": "markdown",
"id": "297746c807e95fbf",
"metadata": {
"collapsed": false
},
"source": [
"* pikepdf is used to repair some PDFs"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "ebc8af0183532fc2",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:08:34.665865Z",
"start_time": "2024-04-04T10:08:32.522020Z"
},
"collapsed": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Name: pikepdf\r\n",
"Version: 8.13.0\r\n",
"Summary: Read and write PDFs with Python, powered by qpdf\r\n",
"Home-page: \r\n",
"Author: \r\n",
"Author-email: \"James R. Barlow\" <james@purplerock.ca>\r\n",
"License: MPL-2.0\r\n",
"Location: /home/marius/miniconda3/envs/llm_langchain/lib/python3.11/site-packages\r\n",
"Requires: Deprecated, lxml, packaging, Pillow\r\n",
"Required-by: \r\n",
"Note: you may need to restart the kernel to use updated packages.\n"
]
}
],
"source": [
"%pip show pikepdf"
]
},
{
"cell_type": "markdown",
"id": "7c7a7f6b0db3719e",
"metadata": {
"collapsed": false
},
"source": [
"* pypdf with all features is needed because this DB consists of 100+ PDFs "
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "779f81e2ab00f73c",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:08:37.436449Z",
"start_time": "2024-04-04T10:08:35.269255Z"
},
"collapsed": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Name: pypdf\r\n",
"Version: 4.0.2\r\n",
"Summary: A pure-python PDF library capable of splitting, merging, cropping, and transforming PDF files\r\n",
"Home-page: \r\n",
"Author: \r\n",
"Author-email: Mathieu Fenniak <biziqe@mathieu.fenniak.net>\r\n",
"License: \r\n",
"Location: /home/marius/miniconda3/envs/llm_langchain/lib/python3.11/site-packages\r\n",
"Requires: \r\n",
"Required-by: \r\n",
"Note: you may need to restart the kernel to use updated packages.\n"
]
}
],
"source": [
"%pip show \"pypdf\""
]
},
{
"cell_type": "code",
"execution_count": 12,
"id": "de3f715519fda6c4",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:08:39.729429Z",
"start_time": "2024-04-04T10:08:37.438498Z"
},
"collapsed": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Name: torch\r\n",
"Version: 2.2.1+cpu\r\n",
"Summary: Tensors and Dynamic neural networks in Python with strong GPU acceleration\r\n",
"Home-page: https://pytorch.org/\r\n",
"Author: PyTorch Team\r\n",
"Author-email: packages@pytorch.org\r\n",
"License: BSD-3\r\n",
"Location: /home/marius/miniconda3/envs/llm_langchain/lib/python3.11/site-packages\r\n",
"Requires: filelock, fsspec, jinja2, networkx, sympy, typing-extensions\r\n",
"Required-by: sentence-transformers, torchaudio, torchvision\r\n",
"Note: you may need to restart the kernel to use updated packages.\n"
]
}
],
"source": [
"%pip show torch"
]
},
{
"cell_type": "markdown",
"id": "ce1350d2d6e3ed63",
"metadata": {
"collapsed": false
},
"source": [
"## Text extraction\n",
"\n",
"* Here the html and text data is extracted into one txt file\n",
"* The PDF and DOCX data is extracted into another txt file\n",
"\n",
"This will be used for weighted data fusion later.\n",
"\n",
"* the texts are normalized: \n",
" * unicode normalization\n",
" * surrogate characters get replaced\n",
" * html gets converted to text\n",
" * pdfs get repaired\n",
" * docx files get read\n",
"\n",
"* exceptions get handled (UTF-16 issues, PDF reference errors)"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "b557444b8b1d4839",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T09:25:39.388933Z",
"start_time": "2024-04-04T09:25:39.320902Z"
},
"collapsed": false
},
"outputs": [
{
"ename": "ModuleNotFoundError",
"evalue": "No module named 'html2text'",
"output_type": "error",
"traceback": [
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
"\u001b[0;31mModuleNotFoundError\u001b[0m Traceback (most recent call last)",
"Cell \u001b[0;32mIn[7], line 5\u001b[0m\n\u001b[1;32m 2\u001b[0m \u001b[38;5;28;01mimport\u001b[39;00m \u001b[38;5;21;01mos\u001b[39;00m\n\u001b[1;32m 4\u001b[0m \u001b[38;5;28;01mimport\u001b[39;00m \u001b[38;5;21;01municodedata\u001b[39;00m \u001b[38;5;66;03m# to normalize text\u001b[39;00m\n\u001b[0;32m----> 5\u001b[0m \u001b[38;5;28;01mimport\u001b[39;00m \u001b[38;5;21;01mhtml2text\u001b[39;00m \u001b[38;5;66;03m# to convert html to text\u001b[39;00m\n\u001b[1;32m 6\u001b[0m \u001b[38;5;28;01mfrom\u001b[39;00m \u001b[38;5;21;01mlangchain\u001b[39;00m\u001b[38;5;21;01m.\u001b[39;00m\u001b[38;5;21;01mdocument_loaders\u001b[39;00m \u001b[38;5;28;01mimport\u001b[39;00m PyPDFLoader, Docx2txtLoader\n\u001b[1;32m 7\u001b[0m \u001b[38;5;28;01mimport\u001b[39;00m \u001b[38;5;21;01mpikepdf\u001b[39;00m \u001b[38;5;66;03m# to repair PDFs\u001b[39;00m\n",
"\u001b[0;31mModuleNotFoundError\u001b[0m: No module named 'html2text'"
]
}
],
"source": [
"import glob\n",
"import os\n",
"\n",
"import unicodedata # to normalize text\n",
"import html2text # to convert html to text\n",
"from langchain.document_loaders import PyPDFLoader, Docx2txtLoader\n",
"import pikepdf # to repair PDFs\n",
"from pathlib import Path\n",
"from tqdm.notebook import tqdm\n",
"from concurrent.futures import ThreadPoolExecutor, as_completed\n",
"\n",
"def convert_html_to_text(html_blob: str) -> str:\n",
" \"\"\"\n",
" Converts a html blob into a string.\n",
" \"\"\"\n",
" h = html2text.HTML2Text()\n",
" h.mark_code = True\n",
" h.escape_snob = True\n",
" h.unicode_snob = True\n",
" # h.use_automatic_links = True \n",
" h.images_as_html = True\n",
" h.single_line_break = True\n",
" h.ignore_links = True\n",
" return h.handle(html_blob)\n",
"\n",
"def normalize_text(txt_blob: str) -> str:\n",
" \"\"\"\n",
" Normalize a text blob using NFKD normalization.\n",
" \"\"\"\n",
" return unicodedata.normalize(\"NFKD\", txt_blob)\n",
"\n",
"def repair_pdf(file_path: str) -> bool:\n",
" \"\"\"\n",
" Attempts to repair a PDF file using pikepdf.\n",
" \"\"\"\n",
" try:\n",
" with pikepdf.open(file_path, allow_overwriting_input=True) as pdf:\n",
" pdf.save(file_path)\n",
" return True\n",
" except pikepdf.PdfError as e:\n",
" print(f\"Failed to repair PDF {file_path}: {e}\")\n",
" return False\n",
"\n",
"def read_and_convert_file(file_path: str, is_html: bool, is_pdf: bool, is_docx: bool) -> str:\n",
" \"\"\"\n",
" Reads and converts a file from HTML, PDF, DOCX, or plain text to text.\n",
" :param file_path: \n",
" :param is_html: \n",
" :param is_pdf: \n",
" :param is_docx: \n",
" :return: \n",
" \"\"\"\n",
" \n",
" content = \"\"\n",
" if is_html:\n",
" try:\n",
" with open(file_path, 'r', encoding='utf-8') as file:\n",
" content = file.read()\n",
" return convert_html_to_text(content)\n",
" except Exception as e:\n",
" print(f\"Error reading {file_path}: {e}\")\n",
" return \"\"\n",
"\n",
" elif is_pdf:\n",
" try:\n",
" loader = PyPDFLoader(file_path)\n",
" # ... fixes \"Multiple definitions in dictionary at byte 0xb32 for key /ExtGState\" error\n",
" documents = loader.load()\n",
" content = \"\\n\".join(doc.page_content for doc in documents if hasattr(doc, 'page_content'))\n",
" except Exception as e:\n",
" print(f\"Error loading PDF {file_path}: {e}. Attempting to repair...\")\n",
" if repair_pdf(file_path):\n",
" try:\n",
" loader = PyPDFLoader(file_path)\n",
" documents = loader.load()\n",
" content = \"\\n\".join(doc.page_content for doc in documents if hasattr(doc, 'page_content'))\n",
" except Exception as e:\n",
" print(f\"Failed to process PDF {file_path} after repair: {e}\")\n",
" return \"\"\n",
" return normalize_text(content)\n",
"\n",
" elif is_docx:\n",
" try:\n",
" loader = Docx2txtLoader(file_path)\n",
" content = loader.load()\n",
" if isinstance(content, list):\n",
" content = \"\\n\".join(content)\n",
" except Exception as e:\n",
" print(f\"Error reading DOCX {file_path}: {e}\")\n",
" return \"\"\n",
" return normalize_text(content)\n",
"\n",
" else: # For plain text files\n",
" try:\n",
" with open(file_path, 'r', encoding='utf-8') as file:\n",
" return normalize_text(file.read())\n",
" except Exception as e:\n",
" print(f\"Error reading {file_path}: {e}\")\n",
" return \"\"\n",
"\n",
"def sanitize_text(text):\n",
" \"\"\"\n",
" Removes or replaces surrogate characters from a string.\n",
" \"\"\"\n",
" return text.encode('utf-8', 'replace').decode('utf-8')\n",
"\n",
"def append_to_output(data: str, is_pdf: bool, is_docx: bool, output_path: str):\n",
" \"\"\"\n",
" Appends sanitized data to an output file.\n",
" \"\"\"\n",
" sanitized_data = sanitize_text(data)\n",
" if is_pdf or is_docx:\n",
" output_path = str(Path(output_path).with_suffix('')) + \".documents.txt\"\n",
" \n",
" with open(output_path, \"a\", encoding='utf-8') as output_file:\n",
" output_file.write(sanitized_data)\n",
"\n",
"def process_file(file):\n",
" is_html = file.endswith('.html')\n",
" is_pdf = file.endswith('.pdf')\n",
" is_docx = file.endswith('.docx')\n",
" \n",
" file_content = read_and_convert_file(file, is_html, is_pdf, is_docx)\n",
" append_to_output(file_content, is_pdf, is_docx, output_path=output_path)\n",
"\n",
"def process_files_in_directory(directory: str):\n",
" txt_html_files = glob.glob(os.path.join(directory, \"*.txt\")) + glob.glob(os.path.join(directory, \"*.html\"))\n",
" pdf_docx_files = glob.glob(os.path.join(directory, \"img\", \"*.pdf\")) + glob.glob(os.path.join(directory, \"img\", \"*.docx\"))\n",
" all_files = txt_html_files + pdf_docx_files\n",
"\n",
" # Initialize the progress bar\n",
" pbar = tqdm(total=len(all_files), desc=\"Processing files\")\n",
"\n",
" with ThreadPoolExecutor(max_workers=3) as executor:\n",
" # Submit all files to the executor and store future objects\n",
" futures = [executor.submit(process_file, file) for file in all_files]\n",
" \n",
" # As tasks complete, update the progress bar\n",
" for future in as_completed(futures):\n",
" pbar.update(1) # Update the progress bar by one for each task completed\n",
"\n",
" # Ensure the progress bar is closed upon completion\n",
" pbar.close()\n",
"\n",
"process_files_in_directory(extracted_evernote_db)"
]
},
{
"cell_type": "markdown",
"id": "e1bcc07f980c865f",
"metadata": {
"collapsed": false
},
"source": [
"## Chunking of the texts\n",
"\n",
"The texts need to get chunked (pre-processing) before the embedding process."
]
},
{
"cell_type": "code",
"execution_count": 13,
"id": "de8d9f18d8342c57",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:09:23.408646Z",
"start_time": "2024-04-04T10:08:56.104045Z"
},
"collapsed": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Now you have 723845 chunks in /home/marius/source/bookworm/export.txt\n",
"Now you have 151259 chunks in /home/marius/source/bookworm/export.documents.txt\n"
]
}
],
"source": [
"from langchain.text_splitter import RecursiveCharacterTextSplitter\n",
"\n",
"def chunk_text_data(txt_file=output_path_extracted_notes):\n",
" \n",
" with open(txt_file) as f:\n",
" text_notes = f.read()\n",
" \n",
" text_splitter = RecursiveCharacterTextSplitter(\n",
" chunk_size=100,\n",
" chunk_overlap=20,\n",
" length_function=len\n",
" )\n",
" \n",
" chunks = text_splitter.create_documents([text_notes])\n",
" print(f'Now you have {len(chunks)} chunks in {txt_file}')\n",
" return chunks\n",
" \n",
"# chunk individual text file containing the data\n",
"text_chunks = chunk_text_data(txt_file=output_path_extracted_notes)\n",
"doc_chunks = chunk_text_data(txt_file=output_path_extracted_docs)"
]
},
{
"cell_type": "markdown",
"id": "aea7ceb111fed5f3",
"metadata": {
"collapsed": false
},
"source": [
"### Embedding costs - why no OpenAI?\n",
"\n",
"The OpenAI API has a cost for the embeddings.\n",
"At this point there seems to be no way to pre-estimate the costs reliably. \n",
"The following calculation is probably flawed:"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "afb2c8feb9ca0bb4",
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"def print_embedding_cost(texts):\n",
" import tiktoken\n",
" enc = tiktoken.encoding_for_model('gpt-4')\n",
" total_tokens = sum([len(enc.encode(page.page_content)) for page in texts])\n",
" print(f'Total Tokens: {total_tokens}')\n",
" print(f'Embedding Cost in USD: { (0.03 / 1_000) * total_tokens}')\n",
" \n",
"print_embedding_cost(text_chunks)"
]
},
{
"cell_type": "markdown",
"id": "8012516604037e2f",
"metadata": {
"collapsed": false
},
"source": [
"## Use Hugging Face Embeddings Sentence Transformers\n",
"\n",
"Here we:\n",
"\n",
"* use a self-hosted on-premises model for the embedding and vectorization\n",
"* configure it for the use with the CPU or GPU\n",
"\n",
"This model is from the Beijing Academy of Artificial Intelligence\n",
"* https://huggingface.co/BAAI/bge-large-en-v1.5 \n",
"* It uses: https://huggingface.co/docs/transformers/model_doc/auto \n",
"\n",
"It will produce embeddings of 1024 dimensions, roughly 500 less than OpenAI Embeddings."
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "3081256c9cf22780",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:09:29.687485Z",
"start_time": "2024-04-04T10:09:23.410187Z"
},
"collapsed": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"No CUDA available\n"
]
}
],
"source": [
"import torch\n",
"use_cuda = torch.cuda.is_available()\n",
"\n",
"USE_GPU=False\n",
"\n",
"if use_cuda:\n",
" print('__CUDNN VERSION:', torch.backends.cudnn.version())\n",
" print('__Number CUDA Devices:', torch.cuda.device_count())\n",
" print('__CUDA Device Name:',torch.cuda.get_device_name(0))\n",
" print('__CUDA Device Total Memory [GB]:',torch.cuda.get_device_properties(0).total_memory/1e9)\n",
" USE_GPU=True\n",
" print(\"GPU enabled\")\n",
" \n",
"if not use_cuda:\n",
" print('No CUDA available')"
]
},
{
"cell_type": "code",
"execution_count": 15,
"id": "c1ca979bbc1610bb",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:09:29.889360Z",
"start_time": "2024-04-04T10:09:29.688832Z"
},
"collapsed": false
},
"outputs": [],
"source": [
"from langchain.embeddings import HuggingFaceEmbeddings\n",
"\n",
"# pre-trained model path\n",
"modelPath = \"BAAI/bge-large-en-v1.5\"\n",
"\n",
"# Create a dictionary with model configuration options, specifying to use the CPU or GPU for computations\n",
"if not USE_GPU:\n",
" model_kwargs = {'device':'cpu'}\n",
"else:\n",
" model_kwargs = {}\n",
"\n",
"# Create a dictionary with encoding options, specifically setting 'normalize_embeddings' to True\n",
"encode_kwargs = {'normalize_embeddings': True}"
]
},
{
"cell_type": "code",
"execution_count": 16,
"id": "3c2b9cd67f161714",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:09:55.733575Z",
"start_time": "2024-04-04T10:09:34.059018Z"
},
"collapsed": false
},
"outputs": [],
"source": [
"# Initialize an instance of HuggingFaceEmbeddings with the specified parameters\n",
"# this model requires sentence_transformers\n",
"\n",
"embeddings = HuggingFaceEmbeddings(\n",
" model_name=modelPath, # Provide the pre-trained model's path\n",
" model_kwargs=model_kwargs, # Pass the model configuration options\n",
" encode_kwargs=encode_kwargs # Pass the encoding options\n",
" )"
]
},
{
"cell_type": "code",
"execution_count": 17,
"id": "3b9ff8cad49442cf",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:10:01.717769Z",
"start_time": "2024-04-04T10:09:58.740831Z"
},
"collapsed": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"1024 dimensions are going to be used\n"
]
}
],
"source": [
"vector = embeddings.embed_query(text_chunks[0].page_content)\n",
"# print(vector)\n",
"n_dimensions = len(vector)\n",
"print(n_dimensions, \" dimensions are going to be used\")"
]
},
{
"cell_type": "markdown",
"id": "b347fb5ee68daf60",
"metadata": {
"collapsed": false
},
"source": [
"## Batch process the embedding\n",
"\n",
"Many data-science tasks require to split a larger processing operation into batch jobs.\n",
"Like in the good old Mainframe days.\n",
"\n",
"The vector DB: https://github.com/asg017/sqlite-vss \n",
"Basis: https://faiss.ai/ - a library for efficient similarity search and clustering of dense vectors.\n",
"\n",
"We add vectors of 1024 dimensions per chunk (sentence, line break delimited) to the DB.\n",
"The processing is done in batches of 50 chunks, using 3 threads."
]
},
{
"cell_type": "code",
"execution_count": 18,
"id": "b03bfcb6c666db1",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:10:08.134514Z",
"start_time": "2024-04-04T10:10:07.895943Z"
},
"collapsed": false
},
"outputs": [],
"source": [
"from concurrent.futures import ThreadPoolExecutor, as_completed\n",
"import os\n",
"\n",
"from tqdm.notebook import tqdm # Import tqdm for notebook\n",
"from typing import List\n",
"from langchain.schema.document import Document\n",
"\n",
"from langchain_community.vectorstores import FAISS"
]
},
{
"cell_type": "code",
"execution_count": 26,
"id": "e6ffc345c26298ad",
"metadata": {
"ExecuteTime": {
"end_time": "2024-04-04T10:32:48.905517Z",
"start_time": "2024-04-04T10:30:48.115043Z"
},
"collapsed": false
},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "bbc83a7dfc1945eb9c967ce0cc85a6b0",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Processing batches: 0%| | 0/14477 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"\n",
"KeyboardInterrupt\n",
"\n"
]
}
],
"source": [
"def add_texts_in_batches(batch: List[Document], sqlite_table: str = \"evernote\", embeddings=embeddings) -> None:\n",
" \"\"\"\n",
" Using type hints is a good idea here, because error messages get swallowed by the ThreadPoolExecutor.\n",
"\n",
" The exception handling serves the same purpose.\n",
" \"\"\"\n",
"\n",
" try:\n",
" db = FAISS.from_documents(batch, embeddings)\n",
" return db\n",
"\n",
" except Exception as e:\n",
" print(f\"Exception occurred in add_texts_in_batches: {e}\")\n",
"\n",
"def divide_chunks(chunks, n):\n",
" \"\"\"\n",
" Divide and conquer\n",
" \"\"\"\n",
" for i in range(0, len(chunks), n):\n",
" yield chunks[i:i + n]\n",
"\n",
"\n",
"def vectorize_data_in_batches(chunks, embeddings):\n",
" num_workers = 3\n",
" batch_size = 50 # Adjust based on your needs and memory constraints\n",
"\n",
" batches = list(divide_chunks(chunks, batch_size))\n",
" faiss_db = None # List to collect the returned db objects\n",
"\n",
" with ThreadPoolExecutor(max_workers=num_workers) as executor:\n",
" # Submit all the batches for processing\n",
" futures = {executor.submit(add_texts_in_batches, batch, embeddings=embeddings): batch for batch in batches}\n",
"\n",
" # Setup the tqdm progress bar\n",
" progress_bar = tqdm(total=len(futures), desc=\"Processing batches\")\n",
"\n",
" for future in as_completed(futures):\n",
" # Each time a future completes, update the progress and collect the result\n",
" progress_bar.update(1)\n",
" try:\n",
" db_result = future.result() # This is where you get the returned value from add_texts_in_batches\n",
" if faiss_db is not None:\n",
" faiss_db = db_result.merge_from(faiss_db)\n",
" else:\n",
" faiss_db = db_result\n",
" \n",
" except Exception as e:\n",
" print(f\"An error occurred: {e}\")\n",
"\n",
" progress_bar.close() # Ensure the progress bar is closed at the end\n",
"\n",
" print(\"All texts have been added to the database.\")\n",
" faiss_db.save_local(\"faiss_index\")\n",
" \n",
"vectorize_data_in_batches(chunks=text_chunks, embeddings=embeddings)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 2
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython2",
"version": "2.7.6"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
Reference in New Issue View Git Blame Copy Permalink