svs 0.3.1

Stupid Vector Store (SVS): a vector database for the rest of us

Stupid Vector Store (SVS)

• 🤔 What is SVS?

  • Semantic search via deep-learning vector embeddings.
  • A stupid-simple library for storing and retrieving your documents.

• 💩 Why is it stupid?

  • Because it just uses SQLite and NumPy. Nothing fancy.
  • That is our core design choice. We want something stupid simple, yet reasonably fast.

• 🧠 Is it possibly... smart in any way though?

  • Maybe.
  • It will squeeze the most juice from your machine: 🍊
    • Optimized SQL
    • Cache-friendly memory access
    • Fast in the places that matter 🚀
    • All with a simple Python interface
  • Supports storing arbitrary metadata with each document. 🗃️
  • Supports storing and querying (optional) parent-child relationships between documents. 👪
    • Fully hierarchical - parents can have parents, children can have children, whatever you need...
  • Both sync and asyncio implementations:
    • use the synchronous impl (svs.KB) for scripts, notebooks, etc
    • use the asyncio impl (svs.AsyncKB) for web-services, etc
  • 100% Python type hints!

Overview

SVS is stupid yet can handle a million documents on commodity hardware, so it's probably perfect for you.

Should you use SVS? SVS is designed for the use-case where:

1. you have less than a million documents, and
2. you don't add/remove documents very often.

If that's you, then SVS will probably be the simples (and stupidest) way to manage your document vectors!

Table of Contents

• Installation
• Used By
• Quickstart
• Speed & Benchmarks
• Debug Logging
• License

Installation

pip install -U svs

Used By

SVS is used in production by:

Quickstart

Here is the most simple use-case; it just queries a pre-built knowledge base! This particular example queries a knowledge base of "Dad Jokes" 🤩.

(taken from ./examples/quickstart.py)

import svs   # <-- pip install -U svs

import os
from dotenv import load_dotenv; load_dotenv()
assert os.environ.get('OPENAI_API_KEY'), "You must set your OPENAI_API_KEY environment variable!"

#
# The database remembers which embeddings provider (e.g. OpenAI) was used.
#
# The "Dad Jokes" database below uses OpenAI embeddings, so that's why you had
# to set your OPENAI_API_KEY above!
#
# NOTE: The first time you run this script it will download this database,
#       so expect that to take a few seconds...
#
DB_URL = 'https://github.com/Rhobota/svs/raw/main/examples/dad_jokes/dad_jokes.sqlite.gz'


def demo() -> None:
    kb = svs.KB(DB_URL)

    records = kb.retrieve('chicken', n = 10)

    for record in records:
        score = record['score']
        text = record['doc']['text']
        print(f" 😆 score={score:.4f}: {text}\n")

    kb.close()


if __name__ == '__main__':
    demo()

⚠️ Want to see how that Dad Jokes knowledge base was created? See: ./examples/dad_jokes/Build Dad Jokes KB.ipynb

Speed & Benchmarks

SQLite and NumPy are fast, thus SVS is fast 🏎️. Our goal is to minimize the amount of work done at the Python-layer.

Also, your bottleneck will likely be the remote API calls to get document embeddings (e.g. calling out to OpenAI to get embeddings will be the slowest thing), so it's likely not critical to further optimize the Python-layer bits.

The following benchmarks were performed on 2018-era commodity hardware (Intel i3-8100):

Dataset Size (# of documents)	Load into SQLite	Obtain Embeddings (remote API call)	Compute Cosine Similarity + Sort + Retrieve Top-100 Documents §
10,548 short jokes †	0.07 seconds	80 seconds	0.5 seconds (first query) + 0.011 seconds (subsequent queries)
1,000,000 synthetic documents ‡	8 seconds	2 hours ¶	2 minutes (first query) + 0.24 seconds (subsequent queries)

† Dad jokes database from this notebook
‡ these one million synthetic documents have an average length of 1,200 characters, see this notebook
§ this time does not include the time it takes to obtain the query string's embedding from the external service (i.e. from OpenAI); rather, it includes the time it takes to compute the cosine similarity with the query string and all the documents (where embedding dimensionality is 1,536), then sort those results, and then retrieve the top-100 documents from the database; the first query is slow because it must load the vectors from disk into RAM; subsequent queries are fast since those vectors stay cached in RAM
¶ this is an estimate based on typical response times from OpenAI's embeddings API; for this test, we generate synthetic embeddings with dimensionality 1,536 to simulate the correct datasize and computation requirements

Debug Logging

This library logs using Python's builtin logging module. It logs mostly to INFO, so here's a snippet of code you can put in your app to see those traces:

import logging

logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
)

# ... now use SVS as you normally would, but you'll see extra log traces!

License

svs is distributed under the terms of the MIT license.