[](https://chattyui.com/)

Chatty

Chatty is your private AI that leverages WebGPU to run large language models (LLMs) natively & privately in your browser, bringing you the most feature rich in-browser AI experience.

Features ✨

• In-browser privacy: All AI models run locally (client side) on your hardware, ensuring that your data is processed only on your pc. No server-side processing!
• Offline: Once the initial download of a model is processed, you'll be able to use it without an active internet connection.
• Chat history: Access and manage your conversation history.
• Supports new open-source models: Chat with popular open-source models such as Gemma, Llama2 & 3 and Mistral!
• Responsive design: If your phone supports WebGl, you'll be able to use Chatty just as you would on desktop.
• Intuitive UI: Inspired by popular AI interfaces such as Gemini and ChatGPT to enhance similarity in the user experience.
• Markdown & code highlight: Messages returned as markdown will be displayed as such & messages that include code, will be highlighted for easy access.
• Chat with files: Load files (pdf & all non-binary files supported - even code files) and ask the models questions about them - fully local! Your documents never gets processed outside of your local environment, thanks to XenovaTransformerEmbeddings & MemoryVectorStore
• Custom memory support: Add custom instructions/memory to allow the AI to provide better and more personalized responses.
• Export chat messages: Seamlessly generate and save your chat messages in either json or markdown format.
• Voice input support: Use voice interactions to interact with the models.
• Regenerate responses: Not quite the response you were hoping for? Quickly regenerate it without having to write out your prompt again.
• Light & Dark mode: Switch between light & dark mode.

Preview

https://github.com/addyosmani/chatty/assets/114422072/a994cc5c-a99d-4fd2-9eab-c2d4267fcfd3

Why?

This project is meant to be the closest attempt at bringing the familarity & functionality from popular AI interfaces such as ChatGPT and Gemini into a in-browser experience.

Browser support

By default, WebGPU is enabled and supported in both Chrome and Edge. However, it is possible to enable it in Firefox and Firefox Nightly. Check the browser compatibility for more information.

How to Install

If you just want to try out the app, it's live on this website.

This is a Next.js application and requires Node.js (18+) and npm installed to run the project locally.

Install from source

If you want to setup and run the project locally, follow the steps below:

1. Clone the repository to a directory on your pc via command prompt:

git clone https://github.com/addyosmani/chatty

2. Open the folder:

cd chatty

3. Install dependencies:

npm install

4. Start the development server:

npm run dev

5. Go to localhost:3000 and start chatting!

Docker

[!NOTE]
The Dockerfile has not yet been optimized for a production environment. If you wish to do so yourself, checkout the Nextjs example

docker build -t chattyui .
docker run -d -p 3000:3000 chattyui

Or use docker-compose:

docker compose up

If you've made changes and want to rebuild, you can simply run docker-compose up --build

Roadmap

• [ ] Multiple file embeddings: The ability to embed multiple files instead of one at a time for each session.
• [ ] Prompt management system: Select from and add different system prompts to quickly use in a session.

Contributing

Contributions are more than welcome! However, please make sure to read the contributing guidelines first :)

Hardware requirements

[!NOTE]
To run the models efficiently, you'll need a GPU with enough memory. 7B models require a GPU with about 6GB memory whilst 3B models require around 3GB.

Smaller models might not be able to process file embeddings as efficient as larger ones.

Acknowledgements & credits

Chatty is built using the WebLLM project, utilizing HuggingFace, open source LLMs and LangChain. We want to acknowledge their great work and thank the open source community.

Authors

Chatty is created and maintained by Addy Osmani & Jakob Hoeg Mørk.