model

A model of any kind of AI, at its provider: one that writes, hears, speaks, makes vectors, or does any other work that a builder declares with types. It is always plain.

Example

import { model, number, object, secret, text } from 'tentjs'
export const backup = model('backup', { provider: 'openai', name: 'gpt-5-mini', key: secret('OPENAI_KEY') })
export const writer = model('writer', { provider: 'anthropic', name: 'claude-sonnet-5-5', key: secret('ANTHROPIC_KEY'), fallback: [backup] })

// functions/orders.ts, in an action, a route or a step of a flow
const summary = await writer.reply(`Summarise this order in one line: ${JSON.stringify(order)}`)
const { score, why } = await writer.fill({ score: number({ min: 0, max: 1 }), why: text() }, `Is this review angry? ${review}`)
for await (const part of writer.stream('Write a short welcome.')) show(part)

Options

Returns

The definition, with the six fixed calls. Each is called in an action, a route or a step of a flow, never in a query or a change (T0013), and answers a promise:

• reply(prompt, { images, reasoning }): the text of the answer. images are ids of files the caller may download; reasoning is a function that is handed the reasoning the model writes for itself, where its provider gives it, and null where the call began anew.
• stream(prompt, { images, reasoning }): the parts of the text as they arrive, read with for await. With shape, stream(prompt, { shape }) gives parts of { value, text }: each the whole value so far, with every field that has not come yet left out, and beside it the text so far. An object or a list still open is closed where it stands, a text that is cut is given as far as it has come, and a number or a word that may not be whole is left out until what follows it has come. The last part is the answer, checked as fill checks one: one that does not fit ends the stream with code.T7108, and one that was cut off with code.T7107, after the parts that were given, and neither goes to a model of fallback.
• fill(shape, prompt, { images, reasoning }): a value of that shape, one type or a record of types. An answer that does not fit is asked for once more with the error. A date of it is the text of the moment.
• embed(text): a vector; a list of texts gives a list of vectors.
• hear(file, { language, hint }): { text, language, seconds } of a recording.
• speak(text, { name, format, speed }): the id of a file that holds the sound. format is mp3 (the default), wav or opus.

A model that states input and output, or a call of a service, is also a function: await painter({ prompt }). Its input is checked before the call and its answer after it; a file of the input reaches the provider as its bytes, and one of the answer comes back as the id of a file of the caller.

On a failure of its server a call is tried again, and then handed to each model of fallback that can do it. Every try is counted and logged, and a try that its provider failed is also a line of tent logs --errors under the name of the model, with what the provider said: so the log says why a model failed where the next one answered. A call inside a step that is a long job is kept, so the step goes on with the same job after a restart.

Errors

• code.T7001: the model is asked for something it cannot do.
• code.T0002: what a model that is a function was handed does not fit its input.
• code.T7104: every model that could answer failed; the message names each.
• code.T7106: the model refused.
• code.T7107: the answer was cut off.
• code.T7108: the answer did not fit its shape or its output.
• code.T7109: the request holds more than the model takes.
• code.T7110: the provider did not accept the request, or refused the key.
• code.T7113: the model names a provider that neither ships nor was made.
• code.T0013: it was called in a query or a change.
• code.T0010: it was called with no engine: nothing stands in the runtime slot.

See also

adapter.Provider, topic.words