NewFriend

This is a Slackbot that can be trained on your Slack history to emulate your friends and coworkers to great amusement for all.

How does it work?

It trains a Markov chain and produces conversations based on those in your workspace. It uses the https://github.com/jsvine/markovify library as core Markov chain functionality.

How do I do it?

You need to be an administrator of your Slack workspace to use NewFriend.

Install

First, clone this repo and install:

git clone https://github.com/perciplex/new-friend.git
cd new-friend
pip install .

Train

To train, you'll need download the conversation history for your Slack workspace. You can do this by following the instructions here: https://slack.com/help/articles/201658943-Export-your-workspace-data

Extract the resulting zip file into a directory. It should contain (among other things) a sub-directory for each channel and a user.json containing data about each user. If there's anyone humorless in your workspace, you should delete their entry from users.json to prevent their replacement from speaking.

To train the Markov model run

python3 -m new_friend train --data <export_data_path> --out model.p

Test

Test the model with

python -m new_friend run --model model.p --dry-run

Setup the bot

To run the model you need create a Slack app and add a bot token with the following scopes:

chat:write -- write messages in chats
chat:write.customize -- write messages with a custom name and icon (an app badge will appear to distinguish from real users)
chat:write.public -- write messages in any channel

For info on how to create a Slack app with bot permissions, check here: https://slack.com/help/articles/115005265703-Create-a-bot-for-your-workspace

Run

Grab the Bot User OAuth Token and press go:

python3 -m new_friend run --model model.p --channel <channel> --token <bot_user_oauth_token>

We recommend posting directly to #general without testing it or warning anyone.