The assistant inside the app

These directives put the model to work inside the app rather than beside it. They all use the one model configured once in the assistant’s settings: a model on this machine (Ollama), or an OpenAI-compatible endpoint with its key.

Where the data goes is a property of the endpoint, not of the directive. With a local model nothing leaves the computer. With a remote provider, whatever the directive puts in the prompt is sent to it. That is one sentence and it holds for all fifteen; there is no directive more careful than another, there is a setting.

If no model is configured, each of them draws one line saying so and the app keeps working. A document written for a model must not become a page of broken widgets on a browser that has none.

The rule that makes them safe

These are directives you will meet inside documents other people wrote, so the right question is not what they can do but what they cannot. The answer is one sentence:

The model never produces anything that is executed.

It produces a word from a closed list the document wrote, or an object whose keys the document declared, or a query plan in which every field name was checked against the ones the collection actually has — and that plan is then run here, on rows this device already holds. Everything else is refused before it can become a row.

It is the same argument as ::python and ::sql: the data travels in a channel of its own and is never pasted into code.

Reading, and asking

::ai-summary{data="expenses"}
Summarise the expenses: how much, on what, and what changed.
::/ai-summary

::ai-chat{data="workouts"}
You are the coach. Answer briefly.
::/ai-chat

::ai-query{data="expenses" into="answer" placeholder="How much did I spend on transport in June?"}

::ai-summary writes a summary and rewrites it whenever the rows change. The body — or prompt= — says what to look at. Only a sample of the rows travels, never the whole collection. It also takes rag=.

::ai-chat answers from the collections in data=, and says when the answer is not in them instead of inventing it. The body is the persona (or persona=), placeholder is what the empty box says. The conversation stays on this device.

::ai-query is the interesting one. The question becomes a plan — filters, a group-by, one aggregation — and the widget shows the plan as well as the answer, because a reader who cannot see what was asked cannot tell a right answer from a plausible one. The single number appears in the widget; the breakdown lands in into=, ready for a ::table or a ::chart-bar.

Filling a form

::form{path="expenses" id="f1"}
::input{field="item" legend="Item"}
::input{field="amount" type="number" legend="Amount"}
::input{field="category" legend="Category"}

::ai-assist{form="f1" placeholder="Pizza with Mario, Thursday, 24 euro"}
::ai-field{form="f1" field="category" values="food,home,transport"}
::ai-suggest{form="f1" path="expenses" fields="item,amount:number"}
::save{label="Save"}
::/form

None of these writes a row. They fill the form’s draft, and the person saves by pressing the form’s own button. It is the same distinction the agent below makes, and it is deliberate: a model that writes rows on its own is a model whose whole output has to be re-checked.

fields= is written name:type — text (the default), number, date, boolean — and a value that is not of that type is left out rather than stored as prose. ::ai-field chooses from values= and from nothing else: an answer outside the list is dropped, never added to it.

::ai-extract without source draws a box of its own; with source="#dictated" it reads a reactive key. ::ai-translate[note]{to="it"} works on a key when the brackets name one, on a field when form/field say so.

::ai-vision needs a model that reads images: Ollama with a multimodal model, or OpenAI.

Working through a whole collection

::ai-classify{path="expenses" field="category" values="food,home,transport" label="Classify"}

::ai-rule{data="expenses" when="amount above 100" do="set review to 'to check'"}

::ai-pipeline{data="tickets" fields="office,urgency,summary"}
Classify the ticket by office and urgency, and write a one-line summary.
::/ai-pipeline

::ai-classify looks only at rows where that field is still empty, unless you write overwrite. Attributes: path, field, values, overwrite, label.

::ai-pipeline looks only at rows where the first declared field is empty — that is what “the new rows” means here — and does at most 25 at a time. Attributes: data, fields, label.

::ai-rule is the one that stops needing the model. It is compiled once into a checked plan — a field, a comparison, a value, a field to write — kept in IndexedDB; from then on it runs on every data change with no request at all. Deterministic, and idempotent because it only touches rows whose value would actually change. Attributes: data, when, do, label.

The agent

::ai-agent{data="bookings,rooms" tools="query,insert"}
You are the bookings assistant: check availability with a query before proposing
an insertion.
::/ai-agent

Two tools and no others. query reads the collections named in data=; insert proposes a row, which appears with a confirm button and is written only when somebody presses it. Attributes: data, tools, placeholder; the body is the persona.

Every call leaves a line in the visible log, because an agent whose steps are invisible is one nobody can check.

Search by meaning

::ai-search{rag="documents.attachment,documents.notes" placeholder="Search the documents"}

::ai-chat{data="documents" rag="documents.attachment,documents.notes"}
Answer citing the documents in square brackets.
::/ai-chat

rag= names the fields as collection.field. Their text — including the content of a ::file attachment, when it is text this browser can read on its own — is cut into passages, embedded, and kept in IndexedDB. The index is rebuilt only when the text it was built from has actually changed, and it never leaves the device except as a request to whatever endpoint is configured. Attributes: rag, placeholder, into.

The embedding model is Qwen3-Embedding-0.6B (ollama pull qwen3-embedding:0.6b, about 600 MB), and it is deliberately not a setting: the choice has two honest answers and the endpoint decides between them. Where the endpoint is OpenAI’s own, text-embedding-3-small is used instead, because that is the one host that will not serve Qwen whatever you ask it for.

PDFs and scans are not indexed. Extracting those needs a parser and an OCR. Such a file is indexed by its name, and a search that quietly indexed the name while looking as though it had read the file would be worse than not offering it: the citation could point at nothing else.