This project aims to develop a browser extension that runs AI models locally on the browser to answer questions about the content of the page the user is visiting.
The extension is available for Chrome, Firefox and Edge. You can install it from the following links:
Upon installation, the extension downloads the models, after its done you can start using it by clicking on the extension icon or by pressing Alt+A. A popup will appear with a text input where you can write your question. After writing your question, press Enter or click on the Sumbmit button to get the answer.
A few settings can be changed after pressing the Settings button. These are the following:
- Embeddings model: the embeddings model to use to embed the text extracted from the websites into vector embeddings. The default is
multi-qa-MiniLM-L6-cos-v1-quantized. - Main generative model: the generative model used to generate an answer given a context. The default is
LaMini-Flan-T5-248M-quantized. - Secondary generative model: the generative model used to generate a temporary answer without context, useful to find more relevant embeddings. The default is
LaMini-Flan-T5-77M-quantized. - Chunk size: the number of characters to use when splitting up the text extracted from the websites to be saved into embeddings. The default is
280. - Results number: the number of embeddings to return from the similarity search to be used as a context to answer the question. The default is
5.
Install Node.js through fnm and pnpm.
curl -fsSL https://fnm.vercel.app/install | bash
curl -fsSL https://get.pnpm.io/install.sh | sh -
fnm useInstall the project dependencies:
pnpm installStart a development server:
pnpm devOpen your browser and load the appropriate development build. For example, if you are developing for the chrome browser, using manifest v3, use: build/chrome-mv3-dev.
To create a production version of the extension, run the following:
pnpm buildThis should create a production bundle for the extension, ready to be zipped and published to the stores. The builds can be packaged into zip files with pnpm package.
- Alberto García ([email protected])