A blazing fast, lightweight, and simple vector database written in less than 200 lines of code. Originally written by sdan, there were CRUD functionalities missing (UD). This is a fork to add the support and fix several issues.

I'm not sure of the comparison methods originally used, but with the CRUD additions and the reworks associated with them, it's possible these comparitive metrics don't hold up. I do not plan on redoing any speed tests.

from vlite import VLite

db = VLite() # default mps

# db = VLite(device="cpu") # to run on cpu

db.memorize(["hello world"]*5)

db.remember("adele")

To install this fork (installs as vlite):

pip install git+https://github.com/Automacene/vlite

To install a specific branch:

pip install git+https://github.com/Automacene/vlite@codie/name_of_branch

pip uninstall vlite

VLite is a vector database built for agents, ChatGPT Plugins, and other AI apps that need a fast and simple database to store vectors.

I built it to support the millions of embeddings I generate , index, and sort with ChatWith+ ChatGPT Plugins which run for millions of users. Most vector databases either repeatedly crashed on a daily basis or was too expensive for the throughput I was putting through.

It uses Apple's Metal Performance Shaders via Pytorch to accelerate vector loading and uses CPU threading to accelerate vector queries to reduce time spent copying vectors from the GPU(MPS) to the CPU.

taken from OpenAI's tiktoken repo, I added a visualize_tokens() function to visualize BPE tokens, i made visualize_tokens to handle the output of the tokenizer.encode() function since the currently supported embeddings are based on BERT and don't use the same tokenization as GPT-4.

You can check out the roadmap to see what will be changing in the near future.