EmbeDB

Vector based image and text database

How it works

EmbeDB uses a vector based approach to store data. This means that the data is stored as a vector of numbers, called an embedding. This allows for fast retrieval of similar data.

What can you do with it?

• Similar image search
• Long term memory
• Web searching
• Much more!

Installation

EmbeDB requires Node.js and Python 3 to be installed.

Install the package using npm:

npm install embedb

Then, create a .env file in the root directory of your project and add the API keys for the models you want to use.

HUGGINGFACE_API_KEY=<your api key>
OPENAI_API_KEY=<your api key>

Usage

First, require the module and create a new instance of the database.

Memory(model<string, default='huggingface'>)

const Memory = require('embedb');

const memory = new Memory();

Inserting data

To memorize text, use the memorize method.

async Memory.memorize({
    key<string>,
    value<string>,
    model<string, default='huggingface'>
})

await memory.memorize({
	key: 'What is my name?',
	value: 'EmbeDB',
});

To memorize an image, you must pass in the image path and use an image model such as resnet50.

await memory.memorize({
	key: 'Matrix meme',
	value: './matrixMeme.png',
	model: 'resnet50',
});

To memorize multiple items, use the memorizeAll method.

async Memory.memorizeAll([
    {
        key<string>,
        value<string>,
    },
    {
        key<string>,
        value<string>,
    },
], model<string, default='huggingface'>)

await memory.memorizeAll([
  {
    key: "What is my name?",
    value: "EmbeDB",
  },
  {
    key: "Who is the president of the United States in 2023?"
    value: "Joe Biden",
  },
]);

Retrieving data

To retrieve the first most similar memory item, use the recall method.

async Memory.recall(key<string>, n<number> model<string, default='huggingface'>) -> MemoryItem{
    key<string>,
    value<string>,
    similarity<number>,
    prune<function>
}

const data = await memory.recall("What's my name?");

To retrieve the first n most similar memory items, use the recall method with the second parameter as n

const name = await memory.recall("What's my name?", 2);
/*
{
    key: "What is my name?",
    value: "EmbeDB",
    similarity: 0.9999999999999999,
    prune: [Function: prune]
},
{
    key: "What is your name?",
    value: "User",
    similarity: 0.3664122137402344,
    prune: [Function: prune]
}
*/

Deleting data

To delete a memory item, use the prune method on a returned memory item from recall.

MemoryItem.prune()

const name = await memory.recall("What's my name?");

await name.prune();

Loading saved data

Memory.load(memoryData<object>)

To load saved data, use the load method.

const fs = require('fs');
await memory.load(JSON.parse(await fs.promises.readFile('./memory.json')));

Embedding Models

Image Models

• resnet50

Text Models

• huggingface
• openai