RiveScript 2.00 Working Draft (2008/01/14)
This document details the standards for the RiveScript scripting language. The purpose of this document is that interpreters for RiveScript could be written in various programming languages that would meet the standards for the RiveScript language itself.
RiveScript is an interpreted scripting language for giving responses to chatterbots and other intelligent chatting entities in a complex trigger/reply format. The scripting language is intended to be simplistic and easy to learn and manage.
RiveScript is the name of the scripting language that this document explains.
The RiveScript interpreter is a program or library in another programming language that loads and parses a RiveScript document.
A RiveScript Document is a text file containing RiveScript code.
A Bot (short for robot) is the artificial entity that is represented by an instance of a RiveScript Interpreter object. That is, when you create a new Interpreter object and load a set of RiveScript Documents, that becomes the ``brain'' of the bot.
A variable that describes the bot, such as its name, age, or other details you want to define for the bot.
A variable that the bot keeps about a specific client, or user of the bot. Usually as the client tells the bot information about itself, the bot could save this information into Client Variables and recite it later.
A RiveScript document should be parsed line by line, and preferrably arranged in the interpreter's memory in an efficient way.
The first character on each line should be the command, and the rest of the
line is the command's arguments. The command should be a single character
that is not a number or a letter.
In its most simple form, a valid RiveScript trigger/response pair looks like this:
+ hello bot - Hello, human.
A RiveScript interpreter should ignore leading and trailing whitespace characters on any line. It should also ignore whitespace characters surrounding individual arguments of a RiveScript command, where applicable. That is to say, the following two lines should be interpreted as being exactly the same:
! global debug = 1 ! global debug= 1
The ! command is for defining variables within RiveScript. It's used to
define information about the bot, define global arrays that can be used in
multiple triggers, or override interpreter globals such as debug mode.
The format of the ! command is as follows:
! type name = value
Where type is one of version, global, var, array, sub, or person.
The name is the name of the variable being defined, and value is the
value of said variable.
Whitespace surrounding the = sign should be stripped out.
Setting a value to <undef> will undefine the variable (deleting it
or uninitializing it, depending on the implementation).
The variable types supported are detailed as follows:
It's highly recommended practice that new RiveScript documents explicitely define the version of RiveScript that they are following. RiveScript 2.00 has some compatibility issues with the old 1.x line (see REVERSE COMPATIBILITY). Newer RiveScript versions should encourag that RiveScript documents define their own version numbers.
! version 2.00
This should override a global variable at the interpreter level. The obvious variable name might be ``debug'' (to enable/disable debugging within the RiveScript interpreter).
The interpreter should take extra care not to allow reserved globals to be overridden by this command in ways that might break the interpreter.
Examples:
! global debug = 1
This should define a ``bot variable'' for the bot. This should only be used in an initialization sense; that is, as the interpreter loads the document, it should define the bot variable as it reads in this line. If you'd want to redefine or alter the value of a bot variable, you should do so using a tag inside of a RiveScript document (see TAGS).
Examples:
! var name = RiveScript Bot ! var age = 0 ! var gender = androgynous ! var location = Cyberspace ! var generator = RiveScript
This will create an array of strings, which can then be used later in triggers
(see + TRIGGER). If the array contains single words, separating the words
with a space character is fine. If the array contains items with multiple words
in them, separate the entries with a pipe symbol ("|").
Examples:
! array colors = red green blue cyan magenta yellow black white orange brown ! array be = is are was were ! array whatis = what is|what are|what was|what were
The sub variables are for defining substitutions that should be run against
the client's message before any attempts are made to match it to a reply.
The interpreter should do the minimum amount of formatting possible on the client's message until after it has been passed through all the substitution patterns.
NOTE: Spaces are allowed in both the variable name and the value fields.
Examples:
! sub what's = what is ! sub what're = what are ! sub what'd = what did ! sub a/s/l = age sex location ! sub brb = be right back ! sub afk = away from keyboard ! sub l o l = lol
The person variables work a lot like subs do, but these are run against
the bot's response, specifically within <person> tags (See TAGS).
Person substitutions should swap first- and second-person pronouns. This is so that ex. if the client asks the bot a direct question using ``you'' when addressing the bot, if the bot uses the client's message in the response it should swap ``you'' for ``I''.
Examples:
! sub you are = I am ! sub i am = you are ! sub you = I ! sub i = you
The > and < commands are for defining a subset of your code under
a certain label. The label command takes between one and three arguments. The
first argument defines the type of the label, which is one of begin, topic,
or object. The various types are as follows.
This is a special label used with the BEGIN block. Every message the
client sends to the bot gets passed through the Begin Statement first, and the
response in there determines whether or not to get an actual reply.
Here's a full example of the Begin Statement.
> begin
+ request
- {ok}
< begin
A topic is a smaller set of responses to which the client will be bound until
the topic is changed to something else. The default topic is random.
The topic label only requires one additional argument, which is the name of
the topic. The topic's name should be one word and lowercase.
Example:
+ i hate you
- Well then, I won't talk to you until you take that back.{topic=apology}
> topic apology
+ *
- I won't listen to you until you apologize for being mean to me.
- I have nothing to say until you say you're sorry.
+ (sorry|i apologize)
- Okay. I guess I'll forgive you then.{topic=random}
< topic
Objects are bits of program code that the interpreter should try to process. The programming language that the interpreter was written in will determine whether or not it will attempt to process the object.
See OBJECT MACROS for more information on objects.
The object label should have two arguments: a lowercase single-word name for
the object, and the programming language that the object should be interpreted
by, which should also be lowercase.
Example:
> object encode perl
my ($rs,$method,$args) = @_;
use Digest::MD5 qw(md5_hex);
use MIME::Base64 qw(encode_base64);
if ($method eq 'md5') {
return md5_hex ($args);
}
elsif ($method eq 'base64') {
return encode_base64 ($args);
}
return "Unrecognized encoding type!";
< object
The + command is the basis for all things that actually do stuff within a
RiveScript document. The trigger command is what matches the user's message to
a response.
The trigger's text should be entirely lowercase and not contain any symbols
(except those used for matching complicated messages). That is, a trigger that
wants to match ``what's your name'' shoudldn't be used; you should use a
substitution to convert what's into what is ahead of time.
Example:
+ are you a bot - How did you know I'm a robot?
An atomic trigger is a trigger that matches nothing but plain text. It doesn't
contain any wildcards (*) or optionals, but it may contain alternations.
Atomic triggers should take higher priority for matching a client's message
than should triggers containing wildcards and optionals.
Examples:
+ hello bot + what is your name + what is your (home|office) phone number + who is george w bush
Using an asterisk (*) in the trigger will make it act as a wildcard. Anything
the user says in place of the wildcard may still match the trigger. For example:
+ my name is * - Pleased to meet you, <star>.
The values matched by the wildcards can be retrieved in the responses by using
the tags <star1>, <star2>, <star3>, etc. in the
order that the wildcard appeared. <star> is an alias for <star1>.
An alternation in a trigger is a sub-set of strings, in which any one of the strings will still match the trigger. For example, the following trigger should match both ``are you okay'' and ``are you alright'':
+ are you (okay|alright)
Alternations can contain spaces in them, too.
+ (are you|you) (okay|alright)
That would match all of the following questions from the client:
are you okay are you alright you okay you alright
Alternations match the same as wildcards do; they can be retrieved via the
<star> tags.
Triggers can contain optional words as well. Optionals are written similarly to alternations, but they use square braces. The following example would match both ``what is your phone number'' as well as ``what is your U<home> phone number''
+ what is your [home] phone number
Optionals do NOT match like wildcards do. They do NOT go into the
<star> tags. The reason for this is that optionals are optional, and
won't always match anything if the client didn't actually say the optional word(s).
Arrays defined via the ! DEFINITION array commands can be used within a trigger. This is the only place where arrays are used, and they're added as a convenience feature.
For example, you can make an array of color names, and then use that array in multiple triggers, without having to copy a whole bunch of alternation code between triggers.
! array colors = red green blue cyan magenta yellow black white orange brown
+ i am wearing a (@colors) shirt - I don't know if I have a shirt that's colored <star>.
+ my favorite color is (@colors) - I like <star> too.
+ i have a @colors colored * - Have you thought about getting a <star> in a different color?
When an array is called within parenthesis, it should be matched into a
<star> tag. When the parenthesis are absent, however, it should not
be matched into a <star> tag.
The - tag is used to indicate a response to a matched trigger. A single
response to a single trigger is called an ``atomic response.'' When more than one
response is given to a single trigger, the collection of responses become a
``random response,'' where a response is chosen randomly from the list. Random
responses can also use a {weight} tag to improve the likelihood of one response
being randomly chosen over another.
A single response to a single trigger makes an Atomic Response. The bot will respond pretty much the same way each time the trigger is matched.
Examples:
+ hello bot - Hello human.
+ my name is * - Nice to meet you, <star>.
+ i have a (@colors) shirt - You're not the only one that has a <star> shirt.
Multiple responses to a single trigger will be chosen randomly.
+ hello - Hey there! - Hello! - Hi, how are you?
+ my name is * - Nice to meet you, <star>. - Hi, <star>, my name is <bot name>. - <star>, nice to meet you.
When using random responses, it's possible to give weight to them to change the likelihood that a response will be chosen. In this example, the response of ``Hello there'' will be much more likely to be chosen than would the response of ``Hi''.
+ hello
- Hello there!{weight=50}
- Hi.
When the {weight} tag isn't used, a default weight of 1 is implied for that
response. The {weight} should always be a number greater than zero and must
be an integer (no decimal point).
The % command is for drawing the user back to finish a short discussion. Its
behavior is similar to using topics, but is implied automatically and used for
short-term things. It's also less strict than topics are; if the client replies
in a way that doesn't match, a normal reply is given anyway. For example:
+ knock knock - Who's there?
+ * % who is there - <star> who?
+ * % * who - lol! <star>! That's hilarious!
The text of the % command looks similar to the text next to the trigger. In
essence, they work the same; the only difference is that the % command matches
the last thing that the bot sent to you.
Here's another example:
+ i have a dog - What color is it?
+ (@colors) % what color is it - That's an odd color for a dog.
In that case, if the client says ``I have a dog,'' the bot will reply asking what
color it is. Now, if I tell it the color in my next message, it will reply back
and tell me what an odd color that is. However, if I change the topic instead and
say something else to the bot, it will answer my new question anyway. This is in
contrast to using topics, where I'd be stuck inside of the topic until the bot
resets the topic to random.
Note: while the % command can contain wildcards and other matchable things,
these values are not matched and put into any <star> tags like they
are for triggers. They're simply discarded.
The ^ command is used to continue the text of a lengthy previous command down
to the new line. It can be used to extend any other command. Example:
+ tell me a poem - Little Miss Muffit sat on her tuffet\n ^ in a nonchalant sort of way.\n ^ With her forcefield around her,\n ^ the Spider, the bounder,\n ^ Is not in the picture today.
Note that when the ^ command continues the previous command, no spaces or
line breaks are implied at the joining of the two lines. The \s and \n
tags must be explicitely defined where needed.
The @ command is used to redirect an entire response to appear as though the
client asked an entirely different question. For example:
+ my name is * - Nice to meet you, <star>.
+ call me * @ my name is <star>
If the client says ``call me John'', the bot will redirect it as though the client actually said ``my name is John'' and give the response of ``Nice to meet you, John.''
The * command is used with conditionals when replying to a trigger. Put simply,
they compare two values, and when the comparison is true the associated response
is given. The syntax is as follows:
* value symbol value => response
The following inequality symbols may be used:
== equal to eq equal to (alias) != not equal to ne not equal to (alias) <> not equal to (alias) < less than <= less than or equal to > greater than >= greater than or equal to
In each of the value places, tags can be used to i.e. insert client or bot variables.
Examples:
+ am i a boy or a girl * <get gender> eq male => You told me you were a boy. * <get gender> eq female => You told me you were a girl. - You never told me what you were.
+ am i your master * <id> eq <bot master> => Yes, you are. - No, you're not my master.
+ my name is * * <get name> eq <star> => I know, you told me that already. * <get name> ne undefined => Did you get a name change?<set name=<star>> - <set name=<star>>Nice to meet you, <star>.
It's recommended practice to always include at least one response in case all of the conditionals return false.
NOTE: Conditionals are tried in the order they appear in the RiveScript document, and the next condition is tried when the previous ones are false.
The # command is for putting comments into your RiveScript document. The
C-style // and /* */ syntax is also supported for commenting.
Comments on their own line should be ignored by all interpreters. Inline comments
(comments next to RiveScript commands) should be ignored only if the comment
symbol is not touching the text of the RiveScript command. There should be at
least one space before the # or // command.
To explicitely use the characters # or // in a RiveScript command, you must
escape them with a \. Example: \# or \//.
An object macro is a piece of program code that is processed by the interpreter
to give a little more ``kick'' to the RiveScript. All objects are required to define
the programming language they use. Ones that don't should result in vociferous
warnings by the interpreter.
Objects should be able to be declared inline within RiveScript code, however they may also be defined by the program utilizing the interpreter as well. All objects should receive, at a minimum, some kind of reference to the RiveScript interpreter object that called them.
Here is an example of a simple Perl object that encodes a bit of text into MD5 or Base64.
> object encode perl
my ($rs,$method,$args) = @_;
use Digest::MD5 qw(md5_hex);
use MIME::Base64 qw(encode_base64);
if ($method eq 'md5') {
return md5_hex ($args);
}
elsif ($method eq 'base64') {
return encode_base64 ($args);
}
return "Unrecognized encoding type!";
< object
To call an object within a response, call it in the format of:
&object_name.method_name(arguments)
For example:
+ encode * in md5 - The MD5 hash of "<star>" is: &encode.md5(<star>)
+ encode * in base64 - The Base64 hash of "<star>" is: &encode.base64(<star>)
In the above examples, &encode calls on the object named ``encode'', which we
defined above; md5 and base64 calls on the method name, which is received
by the object as $method. Finally, $args as received by the object would
be the value of <star> in this example.
$rs in this example would be a reference to the RiveScript interpreter.
Tags are bits of text inserted within the argument space of a RiveScript command. As a general rule of thumb, tags with <angle brackets> are for setting and getting a variable or for inserting text. Tags with {curly brackets} modify the text around them, such as to change the formatting of enclosed text.
No tags can be used within ! Definition and > Label under any
circumstances.
Unless otherwise specified, all of the tags can be used within every RiveScript command.
The <star> tags are used for matching responses. See
+ TRIGGER for usage examples.
The <star> tags can NOT be used within + Trigger.
The input and reply tags insert the previous 1 to 9 things the client said, and
the last 1 to 9 things the bot said, respectively. When these tags are used with
+ Trigger, they should be formatted against substitutions first. This way, the
bot might be able to detect when the client is repeating themself or when they're
repeating the bot's replies.
+ <reply1> - Don't repeat what I say.
+ <input1>
* <input1> eq <input2> => That's the second time you've repeated yourself.
* <input1> eq <input3> => If you repeat yourself again I'll stop talking to you.
* <input1> eq <input4> => That's it. I'm done talking to you.{topic=blocked}
- Please don't repeat yourself.
<input> and <reply> are aliases for <input1> and
<reply1>, respectively.
The <id> tag inserts the client's ID, as told to the RiveScript
interpreter when the client's ID and message were passed in.
Insert a bot variable, which was previously defined via the ! Definition
var commands.
+ what is your name - I am <bot name>, a chatterbot created by <bot company>.
+ my name is <bot name> - <set name=<bot name>>What a coincidence, that's my name too!
Get and set a client variable. These variables are local to the user ID that is chatting with the bot.
+ my name is * - <set name=<star>>Nice to meet you, <star>.
<get> can be used within + Trigger, but <set> can not.
Add, subtract, multiply, and divide a numeric client variable, respectively.
+ give me 5 points - <add points=5>I've added 5 points to your account.
These tags can not be used within + Trigger.
Change the client's topic. This tag can only be used with * Condition and
- Response.
Assign a weight to a random response. This tag can only be used with
- Response.
Perform an inline redirection. This should work like a regular redirection but
is embedded within another response. This tag can only be used with
- Response.
<@> is an alias for {@<star>}
+ your *
- I think you meant to say "you are" or "you're", not "your". {@you are <star>}
Perform an inline definition. This can be used just like the normal
! Definition command from within a reply. This tag can only be used
with - Response.
+ set bot mood to *
- {! var mood = <star>}Bot mood set to <star>.
Insert a sub-set of random text. This tag can NOT be used with + Trigger. Use
the same array syntax as when defining arrays (separate single-word groups with
spaces and multi-word groups with pipes).
+ say something random
- This {random}sentence statement{/random} has a random {random}set of words|gang of vocabulary{/random}.
Process person substitutions on a group of text.
+ say * - Umm... "<person>"
In that example, if the client says ``say you are a robot'', the bot should reply, ``Umm... ''I am a robot.``''
<person> is an alias for {person}<star>{/person}.
Formalize A String Of Text (Capitalize Every First Letter Of Every Word).
+ my name is * - Nice to meet you, <formal>.
<formal> is an alias for {formal}<star>{/formal}.
Format a string of text in sentence-case (capitilizing only the first letter of the first word of each sentence).
<sentence> is an alias for {sentence}<star>{/sentence}.
FORMAT A STRING OF TEXT INTO UPPERCASE.
<uppercase> is an alias for {uppercase}<star>{/uppercase}.
format a string of text into lowercase.
<lowercase> is an alias for {lowercase}<star>{/lowercase}.
This is used only with the ``request'' trigger within the BEGIN block. It tells the interpreter that it's okay to go and get a real response to the client's message.
Inserts a white space character. This is useful with the ^ Continue
command.
Inserts a line break character.
Inserts a forward slash.
Inserts a pound symbol.
Interpreters of RiveScript should follow these guidelines when interpreting RiveScript code. This details some of the priorities for processing tags and sorting internal data structures. This part of the document should be programming-language-independent.
Interpreters should parse all of the RiveScript documents ahead of time and store them in an efficient way in which replies can be looked up quickly.
Triggers should be sorted in a ``most specific first'' order. That is:
1. Atomic triggers first. Sort them so that the triggers with the most amount
of words are on top. For multiple triggers with the same amount of words,
sort them by length, and then alphabetically if there are still matches
in length.
2. Sort triggers that contain optionals in their triggers next. Sort them in
the same manner as the atomic triggers.
3. Sort triggers containing wildcards next. Sort them by the number of words
that aren't wildcards.
4. The very bottom of the list will be a trigger that simply matches * by
itself, if it exists.
% Previous triggers should be sorted in the same manner as + Triggers, and
associated with the reply group that they belong to (creating pseudotopics for
each % Previous is a good way to go).
It will be helpful if the interpreter also offers syntax checking and will give verbose warnings when it tries to parse something that doesn't follow standards. When possible, it should try to correct the error, but should still emit a warning so that the author might fix it.
It would also be good practice to keep track of file names and line numbers of each parsed command, so that syntax warnings can direct the author to the exact location where the problem occurred.
When attempting to get a response to a client's message, the interpreter should support the sending of a ``sender ID'' along with the message. This would preferably be a screen name or handle of the client who is sending the message, and the interpreter should be able to keep different groups of user variables for each user ID. The <id> tag should substitute for the user's ID.
If the BEGIN block was defined in any of the loaded RiveScript documents, it should be tried for the ``request'' trigger. That is, this trigger should be matched:
> begin
+ request
- {ok}
< begin
The interpreter should make the request for that trigger in the context of the calling user, and allow it to change the user's topic or set a user variable immediately. Do not process any other tags that are present in the response (see TAG PRIORITY).
If the response contains the {ok} tag, then make a second request to try to
match the client's actual message. When a response was found, substitute the
{ok} tag from the BEGIN response with the text of the actual response the
client wanted, and then process any remaining tags in the BEGIN response.
Finally, return the reply to the client.
When fetching responses, the following order of events should happen.
1. Build in a system of recursion prevention. Since replies can redirect to
other replies, there's the possibility of deep recursion. The first thing
that the reply fetching routine should do is prevent this from getting out
of control.
2. Dig through the triggers under the client's current topic. Check to see if
there are any %Previous commands on any of these topics and see if they
match the bot's last message to the client. If so, make sure the client's
current message matches the trigger in question. If so, we have a response
set; skip to step 4.
3. Find a trigger that matches the client's message. If one is found, we have
a response set; continue to step 4.
4. If we found a reply set, process the reply. First check if this reply set
has a "solid redirection" (an @ command). If so, recurse the response
routine with the redirection trigger and resume from step 1. Break when an
eventual response was returned.
5. Process conditionals if they exist in order. As soon as one of them returns
true, we have a response and break. If none are true, continue to step 6.
6. See if there is more than one response to this trigger. If any of the random
responses has a {weight}, take that into account as a random response is
chosen. If we have a reply now, break.
7. If there is still no reply, insert a generic "no reply" error message.
When a reply was obtained, then tags should be executed on them in the order defined under TAG PRIORITY.
Within the ``request'' response of the BEGIN block, the following tags can be executed prior to getting a real response for the client's message:
{topic}
<set>
All other tags, especially modifier tags, must be held off until the final
response has been given. Substitute {ok} for the final response, and then
process the other tags.
Things like this should be able to work:
> begin
+ request
* <get name> eq undefined => {topic=new_user}{ok}
* <bot mood> eq happy => {ok}
* <bot mood> eq angry => {uppercase}{ok}{/uppercase}
* <bot mood> eq sad => {lowercase}{ok}{/lowercase}
- {ok}
< begin
All tags that appear within the context of + Trigger must be processed prior
to any attempts to match on the trigger.
The order that the tags should be processed within a response or anywhere else that a tag is allowed is as follows:
<star> # Static text macros
<input> #
<reply> #
<id> #
\s #
\n #
\\ #
\# #
{!} # Redefine variables
<set> # User variable modifiers
<add> #
<sub> #
<mult> #
<div> #
<get> # Get user and bot variables
<bot> #
<person> # String modifiers
<formal> #
<sentence> #
<uppercase> #
<lowercase> #
{random} # Random text insertion (which may contain other tags)
{topic} # Set user topic
<@> # Inline redirection
&objects() # Object macros.
RiveScript 2.00 will have limited backwards compatibility with RiveScript 1.x documents. Here is a full breakdown of the differences:
RiveScript Changes from 1.02 to 2.00 ------------------------------------
REMOVED:
- Variants of !DEFINITION
- ! addpath
- ! include
- ! syslib
- RiveScript Libraries (RSL files)
- RiveScript Packages (RSP files)
- These made code management messy. Keep your own
brain's files together!
COMPATIBLE CHANGES:
- Object macros now require the programming language to be defined.
- Old way: > object encode
- New way: > object encode perl
- The ^CONTINUE command can extend every command.
- Most tags can be used with almost every command.
INCOMPATIBLE CHANGES:
- Conditionals work differently now. Instead of comparing variables to values,
they compare values to values, and each value can <get> variables to compare.
- Old way: * name = Bob => Hello Bob!
- New way: * <get name> eq Bob => Hello Bob!
- Conditionals no longer use a single = for "equal to" comparison. Replace it
with either == or "eq".
- Object macros will receive a reference to the RiveScript object as their first
argument.
NEW THINGS:
- Conditionals have more inequality comparisons now:
"==" and "eq" : equal to
"!=", "ne", and "<>" : not equal to
Nice interpreters might be able to fix some old RiveScript code to make them work. For example, if a condition is found that has one equals sign instead of two, it could print a warning that it's detected RiveScript 1.x code in action and automatically adjust it to 2.x standards, and perhaps reparse the entire file or group of files, assuming that they are RiveScript 1.x code and fix these inconsistencies altogether.
Or perhaps there will just be a converter tool created that would go through code that it already assumes will be RiveScript 1.x and update it to 2.x standards.
Note that this document is only a working draft of the RiveScript 2.00 specification and may undergo numerous changes before a final standard is agreed on. Changes to this document after the creation date on January 14, 2008 will be noted in a change log.