AI

import { ai } from "platform:ai";

Routed through OpenRouter using the model configured in AI Settings, so guest code never holds a provider key.

ai.generate

const { text } = await ai.generate("Summarize this issue in one sentence.", {
  systemPrompt: "You write terse status updates.",
  temperature: 0.2,
});

AiGenerateOptions

Option	Default	Purpose
`model`	AI Settings	OpenRouter model id, e.g. `anthropic/claude-3.5-sonnet`
`systemPrompt`	none	Prepended system message
`temperature`	`0.7`	Sampling temperature, 0–2

Lower the temperature for anything you intend to parse. The default suits prose, not extraction.

ai.generateObject

Structured output against a schema. Use platform:zod to declare the shape.

import { ai } from "platform:ai";
import { z } from "platform:zod";

const { object } = await ai.generateObject(
  `Extract the severity and component from: ${issue.description}`,
  z.object({
    severity: z.enum(["low", "medium", "high", "critical"]),
    component: z.string(),
  }),
  { temperature: 0 },
);

object.severity; // typed

The schema constrains generation, so you get a conforming object rather than prose you have to parse.

ai.embed

Vector embeddings, for semantic search against pgvector.

const { embeddings } = await ai.embed("disk usage alert");
const [vector] = embeddings;

Accepts a string or an array; embeddings is always an array of vectors, so a single string still yields embeddings[0]. Defaults to openai/text-embedding-3-small, overridable with { model }.

Batch rather than looping:

const { embeddings } = await ai.embed(issues.map((i) => i.title));

ai.chat

Multi-turn conversation.

const { text } = await ai.chat([
  { role: "system", content: "You are a triage assistant." },
  { role: "user", content: "What should I do with PLAT-431?" },
]);

Roles are system, user, and assistant. Options are the same as generate minus systemPrompt — put the system message in the array instead.