prague

Ideas

Spell out "magic" functions:

arg => first(...)(arg)

Do "simple" Prague:

Prague

A rule system handy for games and conversational user interfaces. I thought of it as I walked around the city of Prague on a sunny Spring day. This is not an official Microsoft project.

Major features of Prague:

• strongly-typed when using TypeScript (but you don't have to use TypeScript)
• deeply asynchronous via RxJS (but you don't have to use RxJS)
• utilizes and promotes functional programming (you do actually have to use functional programming)

Some types of applications you could build with Prague:

• OS shell
• Chat bot
• Games

Building Prague

• clone this repo
• npm install
• npm run build (or npm run watch to build on file changes)

To add to your app

• npm install prague

Prague Essentials

Transform

The fundamental unit of Prague is a special type of function called a Transform:

type Transform<ARGS extends any[], OUTPUT extends Result | null> = (...args: ARGS) => Observable<OUTPUT>;

A Transform function is called with arguments as normal. But instead of returning a result directly, it returns an object called an Observable. You subscribe to that object to get the result. If you're new to Observables, you may want to read Observables and Promises, which has both a quick introduction to Observables and also shows how to ignore them and just work with Promises.

A Transform emits either null or a subclass of Result.

Result

Result is an abstract base class. The following three subclasses of Result are "core" to Prague:

• Value<VALUE> - contains a value of type VALUE
• Action - contains an action (function) to potentially execute at a future time

Prague also includes and makes use of these subclasses of Result:

• ActionReference - contains the (serializable) name and arguments of a function to potentially execute at a future time
• Multiple - contains an array of Results

from

The from function allows you to write Tranforms more simply, by returning a value instead a Value, a function instead of an Action, undefined instead of null, or a Promise or a synchronous result instead of an Observable:

const repeat = from((a: string) => a.repeat(5))

const confirm = from((a: number) => () => console.log(`You picked ${a.toString()}`));

const getName = from((a: string) => fetch(`url/${a}`).then(r => r.json()).then(r => r.name));

are equivalent to:

const repeat = (a: string) => Rx.of(new Value(a.repeat(5)));

const confirm = (a: number) => Rx.of(new Action(() => console.log(`You picked ${a.toString()}`)));

const getName = (a: string) => Rx.from(fetch(`url/${a}`).then(r => r.json()).then(r => new Value(r.name)));

For your convenience, from is automatically called every place a Transform is expected. For example:

first(
    (t: string) => t === "Bill" ? "Bill Barnes" : null,
    t => t === "Hao" ? "Hao Luo" : null,
    t => t === "Kevin" ? "Kevin Leung" : null,
)

is equivalent to:

first(
    from((t: string) => t === "Bill" ? "Bill Barnes" : null),
    from(t => t === "Hao" ? "Hao Luo" : null),
    from(t => t === "Kevin" ? "Kevin Leung" : null),
)

As a result you never need to explicitly call from unless you are writing your own helper function.

Composition via helpers

You can compose Transforms together into a new Transform using a variety of high-order functions included in Prague, or you can create your own.

first

first returns a new Transform which calls each of the supplied Transforms in turn. If one emits a Result, it stops and emits that. If all emit null, it emits null.

import { first } from 'prague';

const fullName = first(
    (t: string) => t === "Bill" ? "Bill Barnes" : null,
    t => t === "Hao" ? "Hao Luo" : null,
    t => t === "Kevin" ? "Kevin Leung" : null,
);

fullName("Bill").subscribe(console.log);    // Value{ value: "Bill Barnes" }
fullName("Hao").subscribe(console.log);     // Value{ value: "Hao Luo" }
fullName("Yomi").subscribe(console.log);    // null

Note that all the Transforms have the same argument types. However you only need to declare the argument types for the first Transform. TypeScript will use those for the rest, and for the resultant Transform, automatically. It will also complain if your Transforms have incompatibile argument types.

pipe

pipe returns a new Transform which calls each of the supplied Transforms in turn. You supply the arguments for the first. If it emits a Result, that becomes the argument for the second, and so on. If any of the Transforms emit null, the new Transform stops and emits null. Otherwise the new Transform emits the Result emitted by the last Transform.

import { pipe } from 'prague';

const someAssemblyRequired = pipe(
    (a: string, b: string) => a + b,
    fullName,
);

someAssemblyRequired("Kev", "in").subscribe(console.log);      // Value{ value: "Kevin Leung." }
someAssemblyRequired("Yo", "mi").subscribe(console.log);       // null

Note that you only need to declare the argument types for the first transform. TypeScript will infer the argument types for the rest (and for the resultant Transform) automatically.

match

match(getValue, onValue, onNull) returns a new Transform that calls getValue. If that emits a Value, it calls onValue with that value, and emits its output. If getValue emits null, onNull is called with no arguments, and the new Transform emits its output. If onNull is omitted, the new Transform emits null when getValue emits null.

import { match } from 'prague';

const greet = match(
    fullName,
    m => `Nice to meet you, ${m.value}.`,
    () => `I don't know you.`,
);

greet("Kevin").subscribe(console.log);     // Value{ value: "Nice to meet you, Kevin Leung." }
greet("Yomi").subscribe(console.log);      // Value{ value: "I don't know you." }

matchIf

matchIf is a special case of match for the common case of testing a "truthy" predicate.

import { matchIf } from 'prague';

const greet = matchIf(
    (t: string) => t === "Bill",
    () => `I greet you, my creator!`,
    () => `Meh.`,
);

greet("Bill").subscribe(console.log); // Value{ value: "I greet you, my creator!" }
greet("Yomi").subscribe(console.log); // Value{ value: "Meh." }

tap

tap returns a Transform that executes a function but ignores its output, returning the original input. This is a great way to debug:

pipe(
    (t: string) => t === "Bill" ? "Bill Barnes" : null,
    tap(console.log),
    t => t.repeat(2),
).("Bill")
.subscribe();
// Value{ value: "Bill Barnes" }

This is common enough that Prague provides a helper called log which is equivalent to tap(console.log).

Action and run

Imagine we're creating a chatbot that can respond to several phrases:

const bot = from((t: string) => {
    if (t === "current time")
        console.log(`The time is ${new Date().toLocaleTimeString()}`);
    else if (t === "I'm hungry")
        console.log(`You shoud eat some protein.`);
    else if (t === "Wassup")
        console.log(`WAAAASSSUUUUUUP!`);
});

bot("Wassup").subscribe(); // WAAAASSSUUUUUUP

This works, but it isn't the Prague way. Rather than executing code immediately, we prefer to return Actions:

const bot = from((t: string) => {
    if (t === "current time")
        return () => console.log(`The time is ${new Date().toLocaleTimeString()}`);
    else if (t === "I'm hungry")
        return () => console.log(`You shoud eat some protein.`);
    else if (t === "Wassup")
        return () => console.log(`WAAAASSSUUUUUUP!`);
})

Now we can use tap to call the action:

pipe(
    bot,
    tap(m => {
        if (m instanceof Action)
            return m.action();
    }),
)("Wassup").subscribe(); // WAAAASSSUUUUUUP

This is common enough that Prague provides a helper called run:

pipe(
    bot,
    run,
)("Wassup").subscribe(); // WAAAASSSUUUUUUP

Obviously actions can do much more than console.log. This approach of waiting to executing side effects until you're done is a classic functional programming pattern, and makes for much more declarative code.

Scoring: best, sorted, and top

Something we have not touched on is that every Result has a score, a floating point numeric value between 0 and 1, inclusive. By default this score is 1, but you can specify a different score when creating any Result:

new Value("Bill", .5); // Value{ value: "Bill", score: .5 }

Scores are useful when the situation is ambiguous. Say our chatbot asks the user for their name. The user's response might be their name, or they might be ignoring your question and giving a command. How can you know for sure? Certain responses are more likely than others to mean "I am telling you my name". One strategy is to assign a score to each outcome, and choose the highest-scoring outcome. That's where scoring comes in.

In this example we'll first score two different potential responses to a request for a name, then we'll choose the highest scoring one. If there is one, we'll create an action with that score. Finally we'll put that against a differently scored action.

import { best } from 'prague';

const bot = best(
    match(
        best(
            pipe(
                (t: string) => /My name is (.*)/i.exec(t),
                matches => matches.value[1], // gets converted to a Value of score 1
            ),
            t => new Value(t, .5),
        ),
        m => new Action(() => console.log(`Nice to meet you, ${m.value}`), m.score)
    ),
    matchIf(
        t => t === "current time",
        () => new Action(() => console.log(`The time is ${new Date().toLocaleTimeString()}`), .9),
    ),
);

const test = (a: string) => pipe(
    bot,
    run
)(a).subscribe();

test("Bill"); // Nice to meet you, Bill
test("My name is Bill"); // Nice to meet you, Bill
test("current time"); // The time is 6:50:15 AM
test("My name is current time") // // Nice to meet you, Current Time

So far, so good. But consider this case:

const transforms = [
    () => new Value("hi", .75),
    () => new Value("hello", .75),
    () => new Value("aloha", .70),
    () => new Value("wassup", .65),
];

best(
    ...transforms
)().subscribe(console.log) // Value{ value: "hi", score: .75 }

Calling best can be unsatisfactory when there is a tie at the top. Things get even more challenging if you want to program in some wiggle room, say 5%, so that "aloha" becomes a third valid result.

It turns out that best is a special case of a helper called sorted, which returns a Transform which calls each supplied Transform with the supplied arguments. If none emit, neither does it. If one returns a Result, it returns that. If two or more return a Result, it returns a Multiple, which is a Result containing an array of all the Results.

const sortme = sorted(
    ...transforms
);

sortme().subscribe(console.log); // Multiple{ results:[ /* all the results */ ] }

We can narrow down this result using a helper called top.

To retrieve just the high scoring result(s):

pipe(
    sortme,
    top(),
)().subscribe(console.log); // Multiple{ results:[ Value{ value: "hi", score: .75 }, Value{ value: "hello", score: .75 }, ] }

To include "aloha" we can add a tolerance of 5%:

pipe(
    sortme,
    top({
        tolerance: .05,
    }),
)().subscribe(console.log); // Multiple{ results:[ Value{ value: "hi", score: .75 }, Value{ value: "hello", score: .75 }, Value{ value: "aloha", score: .70 }, ] }

We can set a tolerance of 1 (include all the results) but set the maximum results to 3. This will have the same effect as the above:

pipe(
    sortme,
    top({
        maxResults: 3,
        tolerance: 1,
    }),
)()
.subscribe(console.log); // Multiple{ results:[ Value{ value: "hi", score: .75 }, Value{ value: "hello", score: .75 }, Value{ value: "aloha", score: .70 }, ] }

Increasing tolerance includes more items in the "high score". It defaults to 0 and has a maximum value of 1.

Decreasing maxResults limits of the number of "high score" results retrieved. It defaults to Number.POSITIVE_INFINITY and has a minimum value of 1.

In fact, best is just a special case of piping the results of sorted into top:

const best = (...transforms) => pipe(
    sorted(...transforms),
    top({
        maxResults: 1,
    }),
);

top is just one way to narrow down multiple results. There are others. You may apply multiple heuristics. You may even ask for human intervention. For instance, in a chatbot you may wish to ask the user to do the disambiguation ("Are you asking the time, or telling me your name?"). Of course their reply to that may also be ambiguous...

ActionReference and ActionReferences

tk

Observables and Promises

Observables are a powerful and flexible approach to writing asynchronous code, but you don't have to go all the way down that rabbit hole to use Prague. All you need to knoe is that an Observable emits zero or more values, and then either throws an error or completes. Prague Transforms never emit more than one value, which will always be a Result.

Calling a Transform

fullName("Bill")
    .subscribe(
        result => {
            // handle result here
        },
        err => {
            // handle error here
        },
    )

Using Promises instead

If you think this looks similar to writing resolve/reject handlers for a Promise, you're right. In fact, you can easily convert an Observable to a Promise as follows:

fullName("Bill")
    .toPromise() // returns a Promise<Value<string> | null>
    .then(
        result  => {
            // handle result here
        },
        err => {
            // handle error here
        },
    )

Reference

tk

Samples

tk