Gmail has a feature called "Smart Reply," where users can select from a list of recommended responses when replying to an email. But Gmail doesn't allow you to create your own smart replies. Smarter Reply for Gmail is a Chrome extension which does let you create custom replies, and will learn to recommend the right reply for a given email.

Installation

You can download the extension on the Chrome web store. Or you can install the development version of the extension by cloning the repo, navigating to chrome://extensions/, selecting "Load Unpacked" in the top-left corner, and uploading the repo.

Usage

Before using Smarter Reply, go to your Gmail settings and enable Smart Compose. It is usually enabled by default.

1. Creating a Smarter Reply

To create a Smarter Reply, open an email that already has smart replies. Click the "+" button and you should see a prompt for creating a new smart reply. Enter the title and body of your message and then click "Create."

2. Using a Smarter Reply

To use your Smarter Reply, simply click it as you would a regular smart reply. An email draft should be created and autofilled with your smart reply.

3. Deleting a Smarter Reply

To delete a Smarter Reply, click the "-" button and then select the Smarter Reply(ies) that you would like to delete. Then click the "-" button again.

How it Works

Smarter Reply is written in pure Javascript and has no dependencies. The event loop is pretty straightforward:

1. Background script (background.js) detects when a user opens an email and sends a signal to the content script (content.js)
2. Content script scrapes the email content and sends it back to the background script
3. Background script searches the Chrome cache for the best matching Smarter Replies (*) and sends them to the content script
4. Content script injects JS into the Gmail frontend to render the Smarter Replies
5. When a Smarter Reply is created, clicked, or deleted on the frontend, the content script sends a signal to the background to update the Chrome cache

* Smarter Replies are recommended by a unigram model, which calculates the similarity of the received email to the set of emails associated with each Smart Reply. Here's how it works:

1. An opened email is scraped and preprocessed (stopword and punctuation removal, uniform capitalization, and stemming)
2. For each Smarter Reply, a bag-of-words (BoW) model is fitted from the previous emails associated with that Smarter Reply
3. The BoW model produces a vector for the input email, x_i, and each associated email, y₀, y₁, ..., y_n
4. y₀, y₁, ..., y_n are summed to create y, a vector representing word frequencies across all associated emails
5. Cosine similarity between x_i and y is calculated, and the Smarter Replies with the highest similarity measures are returned to the frontend