Try without Firebase authentication (temporary solution): https://github.com/vignshwarar/AI-Employe/issues/2#issuecomment-1880328518
Our stack consists of Next.js, Rust, Postgres, MeiliSearch, and Firebase Auth for authentication. Please sign up for a Firebase account and create a project.
In Firebase, navigate to Project settings -> Service accounts, generate a private key, and save it inside firebaseAdmin/cert/dev.json if it's for development or prod.json if it's for production.
After that, make sure you install the dependencies before starting the app.
npm installnpm run db:deploynpm run dev (for development)npm run build (for production)npm run start (for production)Once you have run 'dev' or 'build', you will find the extension built inside the ./client/extension/build folder. You can then load this folder as an unpacked extension in your browser.
There are several problems with current browser agents. Here, we explain the problems and how we have solved them.
There are several techniques for this, ranging from sending a shortened form of HTML to GPT-3, creating a bounding box with IDs and sending it to GPT-4-vision to take actions, or directly asking GPT-4-vision to obtain the X and Y coordinates of the element. However, none of these methods were reliable; they all led to hallucinations.
To address this, we developed a new technique where we index the entire DOM in MeiliSearch, allowing GPT-4-vision to generate commands for which element's inner text to click, copy, or perform other actions. We then search the index with the generated text and retrieve the element ID to send back to the browser to take action. There are a few limitations here, but we have implemented some techniques to overcome them, such as dealing with the same text in multiple elements or clicking on an icon (we are still working on this).
To prevent GPT from derailing from tasks, we use a technique that is akin to retrieval-augmented generation, but we kind of call it Actions Augmented Generation. Essentially, when a user creates a workflow, we don't record the screen, microphone, or camera, but we do record the DOM element changes for every action (clicking, typing, etc.) the user takes. We then use the workflow title, objective, and recorded actions to generate a set of tasks. Whenever we execute a task, we embed all the actions the user took on that particular domain with the prompt. This way, GPT stays on track with the task, even if the user has not provided a very brief title and objective; their actions will guide GPT to complete the task.