@qzsy/zod-to-arrow

Generate arrow format prewarm prompts from JSON Schema (typically produced by zod-to-json-schema). Teaches the LLM how to write scalar fields, nested objects, string arrays, and object arrays (ITEM -> START ... END).

Install

npm install @qzsy/zod-to-arrow @qzsy/arrow-parser

Quick start

import { schemaToArrowPrompt } from "@qzsy/zod-to-arrow";

const schema = {
  type: "object",
  properties: {
    title: { type: "string" },
    strengths: { type: "array", items: { type: "string" } },
    dimensions: {
      type: "array",
      items: {
        type: "object",
        properties: {
          description: { type: "string" },
          chartType: { type: "string" },
        },
      },
    },
  },
  required: ["title", "strengths"],
};

const systemPrompt = schemaToArrowPrompt(schema, "课堂评价结果");

Generated prompt includes:

• Rules for DATA -> START / DATA -> END
• String array: STRENGTHS -> START + plain lines + STRENGTHS -> END
• Object array: DIMENSIONS -> START + DIMENSION -> START ... DIMENSION -> END
• Example structure and explicit bad cases (no JSON arrays)

API

| Export | Description | |--------|-------------| | schemaToArrowPrompt(schema, purpose) | Full system prompt for LLM arrow output | | schemaToArrowSchemaBlock(schema) | Documentation / secondary-agent SCHEMA block | | JsonSchema | Re-exported type from @qzsy/arrow-parser |

Full pipeline

import { schemaToArrowPrompt } from "@qzsy/zod-to-arrow";
import { parseToJson } from "@qzsy/arrow-parser";
import { z } from "zod";

const EvalResult = z.object({
  strengths: z.array(z.string()),
  suggestions: z.array(z.string()),
});

const schema = zodToJsonSchema(EvalResult);
const prompt = schemaToArrowPrompt(schema, "评价结果");

// 1. Send prompt + context to LLM → arrow text
// 2. Parse:
const data = parseToJson(llmOutput, schema);
// 3. Validate:
const result = EvalResult.safeParse(data);

Version

0.2.0 — aligned with Go SchemaToArrowPrompt; array/object examples; bad-case hints.

See also: @qzsy/arrow-parser · 箭头格式规范.md