Robot

Robot is a bot for Twitch.TV IRC that learns from people and responds to them with things that it has learned.

Tools for broadcasters and mods

Robot has a number of features for managing activity level and knowledge. Most are automatic: for example, by default, the bot is configured not to send more than one message per two seconds (although this can be changed), and it deletes recently learned information from users who get banned or timed out, or from messages that are individually deleted.

There are a few commands for more explicit management. All of these commands require admin priviliges (which are assigned automatically to the broadcaster and mods). The most relevant ones are:

• forget pattern deletes all recent messages containing the supplied pattern. E.g., if the bot's username is "Robot", then saying @Robot forget anime is trash makes the bot remove all messages received in the last fifteen minutes that contain "anime is trash" anywhere in the message.
• you're too active reduces the random response rate, making the bot speak less often when not addressed.
• set response probability to nn% sets the random response rate to a particular value. This is a good way to make the bot more talkative. Depending on the channel's activity level, the most reasonable values for this are usually somewhere around 2% to 10%.
• be quiet for 2 hours makes Robot neither learn from nor speak in the channel for two hours. You can use other amounts of time, but the bot limits the length to 12 hours – if you really need longer, contact the bot owner. If you don't provide an amount of time, it defaults to one hour.
• you may speak disables a previous use of the "be quiet" command.

For the exact syntax to use these commands, see the relevant section.

What information does Robot store?

Robot stores five types of information:

• Configuration details. This includes things like channels to connect to, how frequently to send messages, and who has certain privileges (including "privacy" privileges). For the most part, this information is relevant only to bot owners, broadcasters, and mods.
• Fifteen-minute history. Robot records all chat messages received in the last fifteen minutes, storing a hash specific to the sender, the channel it was sent to, the time it was received, and the full message text. Robot uses this information to delete messages it's learned under certain circumstances. Whenever Robot receives a new message, all records older than fifteen minutes are removed. Robot also records the messages it's generated in the last fifteen minutes.
• One-week privileged command audit. Robot records uses of most admin- and owner-level commands for seven days, including the user, the command that was used with the full message text, the channel in which it was used, and the time the message was received. For security reasons, there is no way to opt out of this data collection.
• Markov chain tuples. This is the majority of Robot's data, a simple list of prefix and suffix words tagged with the location that prefix and suffix may be used. This data is anonymous; Robot does not know who sent the messages that were used to obtain this information.
• Affection information. If you use the marriage command, Robot associates an "affection level" roughly based on how often you cause her to speak with your Twitch user ID (which is a number unrelated to your username).

If you want Robot not to record information from you for any reason, simply use the give me privacy command. Once you're set up to be private, none of your messages will enter her history or Markov chain data. You'll still be able to ask Robot for messages. If you'd like the bot to learn from you again after going private, use the learn from me again command.

How Robot works

Robot uses the mathematical concept of Markov chains, extended in some interesting ways, to learn from chat and apply its knowledge. Here's an example.

Let's say Robot receives this chat message: Can you provide a better example for me please? The first thing it will do is run some preliminary checks to make sure it's ok to learn from the message, e.g. no links, sender isn't a bot, &c.

This particular message is fine. Robot's next step is to break it up into a list of tokens – basically words, except that the English articles "a," "an," and "the" are usually combined with the next word as well, along with invisible tokens for the start and end of a message. The tokens here are Can, you, provide, a better, example, for, me, please?.

Robot is configured with an "order", a number governing how much context matters when learning from messages. Let's say in this case that the bot has an order of 4. This means Robot takes groups of five tokens at a time and learns that the first four can be followed by the fifth. So, the bot will learn:

• Can may be used to start a message.
• you may follow can at the start. (Robot learns the exact capitalization of the "to" word, but ignores it for the "from" words.)
• provide may follow can you at the start.
• a better may follow can you provide at the start.
• example may follow can you provide a better. (At this point, the start of the message is more than four tokens old, so she doesn't consider it anymore.)
• for may follow you provide a better example.
• me may follow provide a better example for.
• please? may follow a better example for me.
• A message may end after example for me please?.

Learning the message is finished. But, robots don't like learning things they'll never use.

When it's time for Robot to think of something to say, the bot does a "random walk" on everything it's learned. Starting with the invisible token for the end of a message, the bot picks out everything it knows can follow, then chooses one of those words entirely at random. Let's say it picks the word You. Robot records that the random walk went to You, then looks for everything that can follow you at the start. It might pick SHOULD; record it and look from you should, and maybe choose HAVE; then waited.

Now that the walk is at four words chosen, same as the order, Robot stop caring about the start again. So, it's looking for words that can follow you should have waited seen anywhere in a message. In my database at the time I write this, the only option that can follow is till. Since there are few options, and Robot wants to be clever, the bot tries to find matches with less context: where the previous three tokens were should have waited, but the token before that is not you. (Note, it's just a coincidence this happened right when the beginning-of-message token fell out of context. Robot could potentially look for extra matches this way at any point.)

It turns out there are no possibilities for you should have waited, but you should have waited has a few. Let's say so is next. It turns out should have waited so has zero matches, because it came from the search with the first two tokens eliminated. Searching should have waited so gives long! as the only option, and should have waited so only gives long (the idea of applying Markov chains to natural language is that regularities like this exist). If we pick long!, the only option we'll find with any of the searches we try will be the invisible end-of-message token. So, the generated message is You SHOULD HAVE waited so long!.

Commands

Robot acknowledges chat messages which start or end with the bot's username, ignoring case, possibly preceded by an @ character or followed by punctuation. For example, if the bot's name is "Robot", then it will recognize these as command messages:

• @Robot madoka
• madoka @rObOt
• robot madoka
• Robot: madoka
• madoka Robot ?

These are not recognized as commands:

• madoka @Robot homura
• ¡Robot madoka!

When Robot recognizes a command, it strips the triggering portion from the message, and the remainder is the command invocation. So, in all of the command examples above, the command invocation is madoka. If a message both starts and ends by addressing Robot, only the start is removed for the invocation; e.g. the invocation for @Robot madoka @Robot is madoka @Robot.

The command invocation is checked against the list of commands for which the user who sent the command message has appropriate privileges. Robot is designed to understand some amount of English for command invocations; there are usually multiple forms you can use to perform a given command.

Commands for everyone

• where is your source code? provides a link to this page, along with a short summary of prominent technologies leveraged.
• what information do you collect on me? provides a link to the section on privacy on this page.
• give me privacy makes the bot never record any information from your messages.
• learn from me again undoes give me privacy.
• generate something with starting chain generates a message that starts with starting chain. (Nothing happens if the bot doesn't know anything to say from there.)
• uwu genyewates an especiawwy uwu message.
• how are you? AAAAAAAAA A AAAAAAA AAA AAAA AAAAAAAA AA AA AAAAA.
• roar makes the bot go rawr ;3
• will you marry me? asks the bot to be your waifu, husbando, or whatever other label for a domestic partner is appropriate. Robot is choosy and capricious.

If a command invocation doesn't match any command, it instead prompts Robot to speak.

Commands for admins

• forget <pattern to forget> un-learns all recent messages that contain the text "pattern to forget".
• help <command-name> displays a brief help message on a command.
• invocation <command-name> displays the exact regular expression used to match a command.
• list commands displays all admin and regular command names.
• be quiet until tomorrow causes the bot to neither learn from nor randomly speak in the channel for twelve hours.
• be quiet for 2 hours is the same, but for 2 hours, or any other duration.
• you may speak undoes "be quiet" commands, allowing the bot to learn and speak again immediately.
• you're too active reduces Robot's random response rate.
• set response probability to <prob>% sets Robot's random response rate to a particular value.
• speak <n> times generates up to n messages at once, bypassing the bot's rate limit. The maximum for n is 5.
• raid generates five messages at once.
• echo <message> repeats an arbitrary message.

Effects

Robot can apply effects to randomly generated messages, modifying the actual output text. Effects can be configured per channel. The possible effects are:

• uwu: Transform using the uwu command.
• me: Use /me (CTCP ACTION) for the message.
• o: Replace vowels with o.

Privileges

Robot has six privilege levels:

• owner is a special privilege level for the bot owner.
• admin gives access to extra commands for moderating robot's activity levels and knowledge.
• regular is the default privilege level, for basic fun with the Markov chain features.
• ignore removes access to any commands, including Markov chain features. Robot also does not learn from ignored users.
• bot is a mix of admin and ignore privileges. Users with bot privileges can invoke admin commands, but Robot does not learn from their other messages.
• privacy is a mix of regular and ignore privileges. Users with privacy privileges can invoke regular-level commands, but Robot does not learn from their messages.

Robot scans a user's chat badges to assign default privileges. Unless overridden per user, broadcasters, mods, and Twitch staff have owner privileges, and everyone else (including VIPs and subscribers) has regular privileges.

Rate limits

When Robot wants to generate a message, whether randomly or through a command, the bot requests a ticket from its rate limiter for the channel it would send to. If there is no ticket available, Robot does not generate the message.

There are two knobs on Robot's rate limiters: the "rate" and the "burst size." I have done a poor job of explaining what these mean elsewhere, but essentially, the burst size is the maximum number of tickets Robot can take, and the rate is how many tickets regenerate per second.

Say a channel has a rate of 0.1 and a burst size of 2. Robot hasn't said anything for a couple minutes. Someone asks the bot to generate a message at 10:53:00; Robot takes a ticket, leaving two remaining, and sends a message. At 10:53:02, another person talks and triggers a random message; Robot takes another ticket, and now there are 1.2 tickets. Another person demands an uwu at 10:53:07, leaving 0.7 tickets. Someone in the channel needs a meme at 10:53:10; the rate limiter has just regenerated a full ticket, which Robot takes, leaving 0.0. Then, at 10:53:19, another person asks for an uwu, because they're beautiful, but the rate limiter has only 0.9 tickets, so Robot cannot take one and so does not speak.

In addition to the rate limiter described above, Robot has a global rate limiter that prevents her from trying to send more than 100 messages per thirty seconds on average, per Twitch's documentation. However, for that global limiter, she waits for a ticket to become available instead of giving up if there isn't one already.

Running your own instance

Robot is designed to be reasonable to install and use on your own. Doing so lets you control exactly when the bot runs and what information is in its database. If you can program in Go and SQL, you can even modify how Robot works to add new features or to remove existing ones, as long as you follow the GPLv3 license.

Installing and running

First, make sure you have the latest versions of Go and GCC installed. If you open a command prompt or terminal, entering go version should print something like go version go1.15.2 windows/amd64, and entering gcc --version should print something like gcc.exe (tdm64-1) 9.2.0. If either of these commands fails, follow the installation instructions for Go and, on Windows, TDM-GCC. (If you aren't on Windows, GCC is almost certainly installed already.)

Also recommended is to install an SQLite3 database interface, like SQLiteStudio.

With at least Go and GCC installed, simply enter go get github.com/AbsyntheSyne/robot/.... This installs the robot, robot-convert, robot-init, and robot-talk commands. At this point, you should be able to enter robot -help to see a basic help message.

Before you can run Robot, you'll need to use robot-init to initialize a database. You'll probably want to copy and modify the example configuration, then do robot-init -conf modified.json -source robot.sqlite3. See the README for robot-init for more information.

You'll also need an OAuth token for Twitch, with at least chat:read and chat:edit scopes. If you don't already have one, you can get one through the TMI OAuth generator.

Finally, run Robot using robot -source robot.sqlite3 -token y0UrOAuth70Ken.

Database structure

Robot's database tables are:

• audit - log of uses of most admin- and owner-level commands
  • time - time the command was received
  • chan - channel in which the command was received
  • sender - username of the sender
  • cmd - command name that was executed
  • msg - full message text, including the command activation
• chans - configuration for channels known to the bot.
  • name - primary key, name of the channel; must begin with '#' and be all lower case, otherwise Twitch will silently ignore the bot trying to join.
  • learn - tag to use for learned messages
  • send - tag to select from to generate messages
  • lim - maximum length of generated messages
  • prob - probability that a non-command message will trigger generating a message; must be between 0 and 1
  • rate - maximum average messages per second to send
  • burst - maximum number of messages to send in a burst
  • block - regular expression matching messages to block, in addition to the block expression in config
  • respond - whether commands can generate messages, separate from the random speaking chance
  • silence - datetime before which the bot will not learn or speak
  • echo - whether to report an echo directory for this channel (part of a personal experiment)
• config - global configuration, only row 1 is used
  • me - bot's username, used as nick
  • pfix - Markov chain order, as described above
  • block - regular expression matching messages to block
• copypasta - copypasta detection configuration
  • chan - channel to which this configuration applies
  • min - number of messages required to trigger copypasta detection
  • lim - time in seconds to consider messages for copypasta
• effects - global and per-channel effects to apply to messages
  • tag - send tag where used, or everywhere if null
  • effect - effect name, see above
  • weight - integer weight; higher values mean more likely to select
• emotes - global and per-channel emotes to append to messages
  • tag - send tag where used, or everywhere if null
  • emote - emote text
  • weight - integer weight; higher values mean more likely to select
• generated - messages generated in the last fifteen minutes
  • time - timestamp of generated message
  • tag - tag used to generate the message
  • msg - generated message text
• history - messages learned from in the last fifteen minutes
  • tid - Twitch IRC message ID
  • time - timestamp of message receipt
  • tags - formatted tags on the message, with display-name and user-id stripped
  • senderh - hash corresponding to the message sender
  • chan - channel received in
  • tag - tag used to learn the message
  • msg - message text
• marriages - users the bot is currently "married" to
  • chan - channel to which this marriage applies
  • userid - Twitch user ID for the partner in this channel
  • time - time at which the partnership was affirmed
• memes - copypasta messages in the last fifteen minutes
  • time - timestamp of copypaste
  • chan - channel message was copypasted in
  • msg - copypasta message text
• privs - global and per-channel user priviliges
  • user - username receiving this privilege
  • chan - channel where applicable, or NULL if a global default
  • priv - privilege type, one of "owner", "admin", "bot", "privacy", or "ignore". (Regular privs are implied by not being in the table, or can be forced by setting this to the empty string.)
• scores - "affection" levels for marriages
  • chan - channel to which this affection level applies
  • userid - Twitch user ID
  • score - affection level for this user
• tuplesn, where n is a number - Markov chain data n prefix words
  • tag - tag with which this chain was learned
  • p0, p1, ... - prefix words; null means before start of message
  • suffix - suffix word; null means end of message

Owner commands

• warranty prints a brief "NO WARRANTY" message, extracted from the GPLv3, on the bot's terminal.
• disable <command-name> globally disables a command.
• enable <command-name> globally re-enables a command.
• resync synchronizes the bot's channel configurations with what is in the database. Use this after modifying the chans, emotes, or privs tables.
• EXEC <query> executes an arbitrary mutating SQL query.
• raw <command> <params> :<trailing> sends a raw IRC message to the server.
• join <channel> [<learn-tag> <send-tag>] joins a channel and adds its configuration to the database. If the tags are not given, the inserted values will be NULL.
• give <user> {owner|admin|bot|regular|ignore} privileges [in <channel>|everywhere] sets a user's privileges level, in the current channel if omitted.
• quit disconnects from the IRC server and closes the bot.
• reconnect disconnects from and then reconnects to the IRC server.
• list commands (overriding the admin version) lists all commands, including owner-only and disabled ones. The latter are marked by a *.
• debug [<channel>] lists (and prints to terminal) the in-memory configuration of the given channel, or the current one if omitted.