@google-labs/breadboard NPM

Breadboard

A library for prototyping generative AI applications.

This library was inspired by the hardware maker community and their boundless creativity. They make amazing things with off-the-shelf parts and a breadboard, just wiring things together and trying this and that until it works.

Breadboard is an attempt to bring the same spirit of creativity and simplicity to making generative AI applications.

This library's design emphasizes two key properties:

:one: Ease and flexibility of wiring. Make wiring prototypes easy and fun.

:two: Modularity and composability. Easily share, remix, reuse, and compose prototypes.

Requirements

Breadboard requires Node version >=v19.0.0.

!WARNING The library is in active development, and we're transitioning to new syntax. Stuff below will likely be out of date. Please bear with us as we bring our words up to speed with our thoughts/actions.

Installing the library

To install breadboard, run:

npm install @google-labs/breadboard

You will also need the Core Kit:

npm install @google-labs/core-kit

Using breadboard

Just like for hardware makers, the wiring of a prototype begins with the Board.

import { Board } from "@google-labs/breadboard";

const board = new Board();

Breadboards are all nodes and wires. Nodes do useful things, and wires flow control and data between them.

Placing things on the board is simple. This example places an input and an output node on the board:

const input = board.input();
const output = board.output();

Wiring things is also simple:

input.wire("say->hear", output);

The statement above wires the say property of the input node to the hear property of the output node.

The wire method is chainable, so you can wire multiple wires at once. Wiring can also happen in both directions, allowing for more expressivity and flexibility.

Here's an example: a board that uses PaLM API to generate text:

const output = board.output();
board
  .input()
  .wire("say->", output)
  .wire(
    "say->text",
    kit
      .generateText()
      .wire("completion->hear", output)
      .wire("<-PALM_KEY", kit.secrets(["PALM_KEY"]))
  );

You can run boards using runOnce and run methods. The runOnce is the simplest; it takes inputs and produces a set of outputs:

const result = await board.runOnce({
  say: "Hi, how are you?",
});
console.log("result", result);

When run, the output of the sample board above will look something like this:

result { say: 'Hi, how are you?', hear: 'Doing alright.' }

The run method provides a lot more flexibility on how the board run happens, and is described in more detail Chapter 8: Continuous runs of Breadboard tutorial.

Breadboard is designed for modularity. You can easily save boards: they nicely serialize as JSON:

const json = JSON.stringify(board, null, 2);
await writeFile("./docs/tutorial/news-summarizer.json", json);

You can load this JSON from URLs:

const NEWS_BOARD_URL =
  "https://gist.githubusercontent.com/dglazkov/55db9bb36acd5ba5cfbd82d2901e7ced/raw/google-news-headlines.json";
const board = Board.load(NEWS_BOARD_URL);

You can include them into your own boards, similar to JS modules, and then treat them as nodes in your graph:

board
  .input()
  .wire(
    "say->text",
    board.invoke(NEWS_BOARD_URL).wire("text->hear", board.output())
  );

You can even create board templates by leaving "slots" in your board for others to fill in:

const input = board.input();
input.wire(
  "topic->",
  board.slot("news").wire(
    "headlines->",
    template.wire("topic<-", input).wire(
      "prompt->text",
      kit
        .generateText()
        .wire("<-PALM_KEY.", kit.secrets(["PALM_KEY"]))
        .wire("completion->summary", board.output())
    )
  )
);