Wordcraft is an LLM-powered text editor with an emphasis on short story writing.
Wordcraft is a tool built by researchers at Google PAIR for writing stories with AI. The application is powered by LLMs such as PaLM, one of the latest generation of large language models. At its core, LLMs are simple machines — it's trained to predict the most likely next word given a textual prompt. But because the model is so large and has been trained on a massive amount of text, it's able to learn higher-level concepts. It also demonstrates a fascinating emergent capability often referred to as in-context learning. By carefully designing input prompts, the model can be instructed to perform an incredibly wide range of tasks.
However this process (often referred to as prompt engineering) is finicky and difficult even for experienced practitioners. We built Wordcraft with the goal of exploring how far we could push this technique through a carefully crafted user interface, and to empower writers by giving them access to these state-of-the-art tools.
npm i
npm run dev
In order to run Wordcraft, you'll need a PaLM API key. Please follow the instructions at developers.generativeai.google/tutorials/setup. Once you have your API key, create a .env file and add the key!
touch .env
echo "PALM_API_KEY=\"<INSERT_PALM_API_KEY>\"" > .env
Remember, use your API keys securely. Do not share them with others, or embed them directly in code that's exposed to the public! This application stores/loads API keys on the client for ease of development, but these should be removed in all production apps!
You can find more information about the PaLM 2 API at developers.generativeai.google
Wordcraft can be customized by adding additional models or adding operations/controls. The basic architecture allows for a great deal of flexibility in the
Defines the underlying data/examples that will be used to construct few-shot instructions to the underlying language model. This example data can be customized to fit a particular style or genre.
Defines how the user's intent is combined with the document state, manages updating the text editor, and handles user choices.
Defines how the data from the Context
is combined with an Operation
state to
construct text that will be sent to a mode, and parses model output.
The Wordcraft application uses few-shot examples for constructing prompts to
send to the model. The style of the generated text is influenced by these
examples, and you can customize Wordcraft's style or genre by editing these
examples. These examples are found in /app/context/json
, and follow a schema
defined in /app/context/schema
.
To add a new custom control (e.g. a button that translates into pig latin):
- Create a new
pigLatinSchema
in/app/context/schema.ts
- Create a new
pig_latin_examples.json
in/app/context/json/
- Register the examples int the
WordCraftContext
constructor (/app/context/index.ts
) - Create a corresponding prompt handler in
/app/models/palm/prompts
- Register that prompt handler with the underlying
Model
class in/app/models/palm/index.ts
- Create a new
PigLatinOperation
in/app/core/operations
- Register the operation in
main.ts
This is not an officially supported Google product