yaks.app

Querying: the filter grammar

← Documentation

One grammar reads an app's store, and everything that reads one uses it: query() and subscribe() on a page, GET ./api/query underneath them, graph_query in an agent's tools, a query in the app's own tools.json, and env.STORE.fetch from a worker. This page is that grammar with its examples — what selects rows, what a row comes back carrying, and the handful of things this store will not do.

The examples are a kitchen app: a vocab.json declaring

{ "$defs": {
    "recipe": { "properties": {
      "serves":  { "type": "number" },
      "minutes": { "type": "number" },
      "cuisine": { "type": "string" } } },
    "chore": { "properties": {
      "every": { "type": "number" } } },
    "reading": { "properties": {
      "pages":   { "type": "number" },
      "started": { "type": "string", "format": "date-time" },
      "done":    { "type": "boolean" } } } } }

beside the platform's own components — doc, task, comment, archived, created and the rest.

The shape of a filter

A filter is predicates joined by &. Each one is a dot, a component, a column, an operator, a value:

.recipe.minutes<=30
 ^      ^       ^ ^
 |      |       | value
 |      |       operator
 |      column of that component
 component

await query('.recipe.minutes<=30&.doc?')

GET ./api/query?.recipe.minutes<=30&.doc?

Every predicate must hold — & is an intersection, never an "or". (Any-of lives inside one predicate, below.) A predicate ends where the next & begins, so two of them run together is not a longer filter but a mistake, and the store tells you so:

.recipe!.doc!
→ presence filters end at !: .recipe! — join filters with &:
  .recipe!&.doc!

Your own components must be written qualified — .recipe.serves, never .serves — so a component you invent can never change what .title means in somebody else's store. The platform's columns do work bare (.title~=lemon is .doc.title~=lemon), but write the component anyway: it reads better, and it never becomes ambiguous.

An empty filter selects nothing. There is no "everything" — query('') and a bare query('limit=50') both return []. To list what you saved, name a component those rows have: .doc!, .recipe!.

Asking for a component

Four ways to name a component rather than one of its columns:

.recipe!      has a recipe
.recipe=      has no recipe
.recipe?      asks for the recipe without filtering on it
*             every component this store holds (the debugging form)

! and = select. ? selects nothing and screens nothing — it only asks for the component to come back with whatever the rest of the filter selected:

await query('.chore!&.completed=&.doc?')   // chores not yet done
await query('.recipe!&.doc?')              // recipes, with their titles

? is safe for a component this store never planted: asking for one is not claiming it exists, so .recipe!&.loan? returns the recipes and simply leaves the loan off. .loan! for a component that was never planted is refused instead, because an empty answer would lie about what is there.

* returns every component of every row it selected. Use it when you are looking rather than drawing:

await query('.doc.title~=drizzle&*')

It is part of the filter, so it works wherever a filter does: subscribe() keeps running the same query, and the updates it sends carry the same whole rows.

What an answer carries

A row carries only the components its filter named. This is the one rule to hold on to, because the mistake it prevents is a page that draws undefined.

A row is {kind, entity: {eid}, ...those components}. So this is the wrong page:

for (let r of await query('.recipe!')) {
  draw(r.doc.title)                  // TypeError: no doc on the row
}

→ [{ kind: 'recipe', entity: {eid: '940d…'},
     recipe: {serves: 4, minutes: 20, cuisine: 'american'} }]

and this is the right one:

for (let r of await query('.recipe!&.doc?')) {
  draw(r.doc.title, r.recipe.minutes)
}

→ [{ kind: 'recipe', entity: {eid: '940d…'},
     doc: {title: 'Pancakes', body: 'flour, milk, eggs'},
     recipe: {serves: 4, minutes: 20, cuisine: 'american'} }]

Anything a predicate mentions counts as naming it, whichever operator it uses: .doc.title~=cake carries the doc, .task.status=open carries the task, .created.at=today carries the stamp. Three things do not name one: an absence (.archived= asks for rows without one — there is nothing to carry), the * form, which asks for all of them, and a bare word, which searches the docs without naming anything to leave out — so a filter that is only words returns whole entities, like id=.

A column of yours that nothing has written is on the row with the value null, not missing from it, so test the value and not in. The platform's own columns read back the same way, doc.title included.

entity and kind are on every row: entity.eid is the address to write back to — the only name an entity has here — and kind is what the entity is, the first component it has, one of your own included.

Three kinds of row are left out of a listing unless the filter names them: the platform's stamps (.created!, .updated!), the platform's own error rows (.exception!, .error!), and the person rows the store keeps so a byline has a name (.person!). What comes back is what your app saved.

subscribe(filter, cb) returns the same rows by the same rule, so a page swaps one for the other and nothing else changes.

The operators

.recipe.cuisine=thai          equals
.recipe.cuisine!=thai         not
.doc.title~=drizzle           contains, case-insensitive, literal
.recipe.minutes<30            less than
.recipe.minutes<=30           at most
.recipe.serves>4              more than
.recipe.serves>=4             at least
.recipe.cuisine=thai,british  any of
.recipe.minutes=20..35        a range, both ends included
.recipe.minutes=20...35       a range, the end excluded
.recipe.cuisine=              the column is empty or absent
.recipe.cuisine~=             the column is there at all

A number column compares numerically and every other column as text, which is why an ISO stamp compares correctly as text. A comparison a column's type cannot make — .recipe.serves>many — is refused rather than guessed at.

A list and a range are values, not extra syntax: .recipe.cuisine!=thai,indian is "neither", and .recipe.minutes!=20..35 is "outside that band".

!= also matches a row that lacks the column entirely — nothing there is not thai. When you mean "has a cuisine, and it is not thai", write both:

await query('.recipe!&.recipe.cuisine!=thai&.doc?')

~= is a plain substring test on the stored text, not full-text search: it finds drizzle inside Lemon drizzle, and it will not stem, rank, or match across a word boundary. Bare words do that (below).

Values, by type

The platform's own columns are typed, and a value that cannot be one is refused loudly rather than quietly matching nothing:

.task.status=finished
→ task.status is one of open, wip, done, cancelled — got 'finished'

.filed.priority=high
→ priority is a finite number, optionally P-prefixed (P2, p02, 1.5)
  — got 'high'
  • text — as typed. .doc.title~=cake
  • number — .recipe.serves>=4, decimals fine.
  • enum — task.status is open, wip, done or cancelled. It is read, never written: .task.status=done selects the entities that have a completed.
  • priority — a number, and P is optional: .filed.priority<=2 and .filed.priority<=P2 ask the same thing.
  • time — a stamp or a phrase; the next section is only about those.
  • url — text with a shape. Quote it (below), or the & in its query string starts a second predicate.
  • eid — a reference to another entity, by its eid: .comment.target=940d…, .filed.assignee=dc5e…, .created.by=<who.person>.

Your own columns work differently: they are stored as given and compared as text, with no parsing on either side. Two consequences worth knowing before you design a component:

  • A bool of yours lands in the store as 1 or 0, so filter it that way — .reading.done=1 finds the finished ones, .reading.done=true finds nothing.
  • A time of yours is kept verbatim and compared as text, so write ISO stamps (new Date().toISOString()) and compare with ISO: .reading.started>=2026-01-01 works, while .reading.started=today matches only a row whose value is the literal text today. Time phrases are for the platform's stamps.

Numbers of yours still compare as numbers — .recipe.serves>=4 — because both sides read as numbers.

Time phrases

A time phrase is a range, and the operator picks which edge of it you mean:

=   within it            .created.at=today
>=  from its start       .created.at>=2026-01-01
<=  until its end        .created.at<=yesterday
>   after it ends        .created.at>last-week
<   before it starts     .created.at<today

The phrases:

now                       today  yesterday  tomorrow
this|last|next minute|hour|day|week|month|year
5 minutes ago             2 hours ago   3 days ago   (also 5m, 2h, 3d)
in 2 days                 after 8h
9am   9:30pm   14:00   noon   midnight   9am tomorrow
2026-07-04                a whole day
2026-07-25T09:00          a minute; with seconds, a second

Whitespace separates predicates, so join a phrase with - or _, or quote it:

await query('.doc!&.created.at>=1-hour-ago')
await query('.doc!&.created.at>="1 hour ago"')

A phrase is read fresh every time the query runs, so a saved today still means today tomorrow. Weeks start Monday. Day boundaries are local time.

The platform's stamps

Every row the store keeps is stamped with who saved it and when. The stamps sit in components of their own, so they are asked for by name and left out of a listing that does not name them:

.created!        who saved it and when
.created.at      when                (time)
.created.by      who                 (a reference)
.updated.at      when it last changed
.updated.by      who changed it

await query('.recipe!&.created!&.doc?')

→ [{ kind: 'recipe', entity: {…},
     doc: {title: 'Lemon drizzle', body: '…'},
     created: {by: {eid: 'dc5e…', name: 'Maya'},
               at: '2026-09-03T12:59:27.876Z', via: null},
     recipe: {serves: 8} }]

A reference to somebody this store has met reads back as {eid, name}, so one query draws a list with its bylines. Anything else stays the bare eid.

at and by are shared by several components, so write out which one you mean: .created.at, .updated.by, .completed.at, .archived.at. Some lines that come up:

.doc!&.created.by=<who.person>&.doc?      what this person saved
.doc!&.updated.at>=today&.doc?            what changed today
.doc!&.created.by!&.doc?                  the rows that have a byline
.doc!&.created.at=last-week&.created!     last week's, with the stamp

Quoting

Quotes hold a value together against both separators — whitespace and &:

await query('.doc.title~="Lemon drizzle"')
await query('.web.url="https://x.test/p?a=1&b=2"&.web?')

Unquoted, the & in that URL would end the predicate and start a second one. Quoting is how you avoid that; percent-encoding a value with an & in it does not survive the trip, because the HTTP endpoint decodes each part of the query string before the grammar reads it. Quote it.

A value with a space but no & survives on its own — .doc.title~=Lemon drizzle is one predicate — but only until something else is added to the filter. Quoting always works, so quote.

Walking a reference

A dotted path walks a reference column and tests a column on the far side. Every hop but the last must be a reference:

.comment.target.doc.title~=Pancakes    comments on the pancake recipe
.filed.project.doc.title~=Kitchen       tasks in the project called Kitchen
.assignee.title~=maya                  bare: filed.assignee → doc.title

Only the platform's columns can be references — a vocab.json column is text, number, bool, time or url — so paths walk comment.target, filed.project, filed.assignee, attachment.blob, and the by of each stamp. Point at another entity from your own component by keeping its eid in a text column; it holds the address, but a path will not walk it.

The plural form walks the other way — the entities pointing back at this one:

.comments!                    has at least one comment
.comments=                    has none
.comments>=2                  has two or more (any comparison works)
.comments.doc.body~=butter    has some comment mentioning butter
.comments!.doc.body~=butter   has none mentioning butter
.comments!.doc.body!=butter   every comment mentions it (De Morgan)

The name is the component's plural — .comments are the entities whose comment.target names this row. When a component has two reference columns the name picks which one: .tasks_project are the projects that have tasks, .tasks_assignee the people who do; .attachments are the blobs a file row points at.

// the recipes nobody has said anything about
await query('.recipe!&.comments=&.doc?')

// whoever has been given something to do
await query('.tasks_assignee!&.doc?')

And .refs is the union of all of them — everything pointing at one entity, by whichever column:

await query(`.refs=${eid}&.doc?`)   // what mentions this
await query('.refs=&.doc?')         // what mentions nothing

Counting and tallying

Three directives return a value instead of rows. They sit beside the filters that select what they reduce, and they read from the index — a page that wants a number asks for the number, never for the rows to count.

await query('.recipe!&.count!')
→ {count: 3}

await query('.recipe!&.tally=recipe.cuisine')
→ {tally: {american: 1, british: 1, thai: 1}}

await query('.recipe!&.distinct=recipe.cuisine')
→ {distinct: ['american', 'british', 'thai']}

.count! counts what the rest of the filter selects. .tally= counts each value of one column; .distinct= returns those values themselves, sorted. Both name one column, never a path, and your own columns must be written with their component — .tally=recipe.cuisine, not .tally=cuisine.

The answer is an object, not an array, so a page has to branch on which it asked for. Ask for one aggregate at a time: with two in one filter, only the first is used.

let { count } = await query('.chore!&.completed=&.count!')
badge.textContent = `${count} to do`

Windows

A window bounds the answer without changing what matches.

limit=20            the newest 20 matches

A plain listing is oldest first, in the order the rows were written. A windowed listing is the newest that many — still in that same oldest-first order among themselves. So limit is the front page of a feed:

let page = await query('.doc!&limit=20')

Both forms work — limit=20 and .limit=20 are the same window — and a bad bound is refused rather than dropped:

.limit=abc
→ .limit takes a whole number: .limit=200

Remember that a window without a filter is nothing: limit=50 alone returns [], because an empty filter selects nothing.

One entity, whole

await query(`id=${eid}`)

id= addresses instead of selecting, and an address names no component to leave out — so it returns the whole bundle, stamps and all. Several at once is a comma list, or a repeat:

await query(`id=${a},${b}`)

Name a component beside it and you are back to the ordinary rule: id=<eid>&.doc? returns that entity's doc alone. Anything the store minted an eid for is addressable this way, including a blob you uploaded.

Words

A bare word is a full-text term over every doc — title and body — and the answer comes back ranked, best first:

await query('lemon')
await search('lemon')                  // the same search
await query('lemon&.recipe!&.doc?')    // ranked, and only recipes

→ [{ kind: 'recipe', entity: {…},
     doc: {title: 'Lemon drizzle', body: '3 lemons, 200g butter'},
     recipe: {minutes: 50, serves: 8},
     rank: {title: 'Lemon drizzle', snip: '3 \x01lemons\x02, 200g butter',
            score: 2.0000017} }]

A word names no component to leave out, so — like id= above — a filter that is only words returns the whole entity, the app's own components included. That is what lets a page draw cards straight from a search. Name a component beside the word and you are back to the ordinary rule: lemon&.recipe! returns recipes with no titles, lemon&.recipe!&.doc? returns both.

rank is added to the row for a text query and is never stored. Its snip marks the hit with \x01 and \x02 so a page can wrap them in whatever it likes — never HTML from the store.

Terms match whole tokens: lemon does not find lemons. A trailing * prefix-matches, and quotes make a phrase:

await query('lemo*')
await query('"coconut milk"')

Words and filters mix freely in one filter, which is why a search box can hand its whole string to query() and a saved filter is a valid search.

What this store will not do

Two pieces of the wider platform grammar are refused in an app's store, by name, rather than quietly doing something else:

.doc!&work=build       → work lanes are not served by this store
.doc!&.order=similar   → semantic ranking is not served by this store

There is no vector search in an app's store, so .near=<eid> alone changes nothing about the answer — do not reach for it. Ranking is what a text term gives you.

.kind= knows the platform's own components only (.kind=task, .kind=comment); a component of your own is not a kind to it, and .kind=recipe is refused. Ask for the component instead — .recipe! — which is what you meant.

Everything else the grammar has, this store serves.

A dozen filters that answer something

.recipe!&.doc?
  every recipe, with its title and body

.recipe.minutes<=30&.doc?
  what you can cook in half an hour

.recipe.cuisine=thai,vietnamese&.doc?
  two cuisines at once

.doc.title~=lemon&.recipe!
  recipes whose title contains lemon — a substring, not a search, and the
  title predicate already asked for the doc

.chore!&.completed=&.doc?
  the chores still to do

.chore!&.completed.at=today&.doc?
  the chores done today

.chore!&.count!
  how many chores there are, without listing one

.task.status=open&.tally=filed.domain
  how much open work each part of the house has

.reading!&.reading.done=1&.reading.started>=2026-01-01&.doc?
  books finished since the new year (your own bool is 1, your own time
  is ISO text)

.recipe!&.comments=&.doc?
  recipes nobody has commented on

.doc!&.created.by=<who.person>&.created!&limit=20
  the newest twenty things this person saved, with their stamps

id=<eid>
  one entity, whole, however many components it has

The whole guide is at The guide.