Lesson 05

Capstone: a structured data extractor

Turn free-form text (emails, invoices, messages) into reliable JSON, with tests. The portfolio piece of this course.

What we are building

A function extract(text) that takes free-form text and returns a JS object with structured data. For example:

"Hi, I'm Ana García, my ID is 12345678A and I live at 12 Mayor St, Madrid.
My phone is +34 600 123 456. I want to book for May 15th at 9pm."

…should return:

{
  "name": "Ana García",
  "id": "12345678A",
  "address": "12 Mayor St, Madrid",
  "phone": "+34 600 123 456",
  "bookingDate": "2026-05-15",
  "bookingTime": "21:00"
}

Step 1 — Define the schema

Before the prompt, define the schema in code. That is what makes your system reliable.

import { z } from "zod";

const BookingSchema = z.object({
  name: z.string(),
  id: z.string().regex(/^\d{8}[A-Z]$/),
  address: z.string(),
  phone: z.string(),
  bookingDate: z.string().regex(/^\d{4}-\d{2}-\d{2}$/),
  bookingTime: z.string().regex(/^\d{2}:\d{2}$/),
});

type Booking = z.infer<typeof BookingSchema>;

If the model goes off-schema, we catch it in parse() and retry.

Step 2 — The prompt

Apply everything you have learned:

const SYSTEM = `You are a data extractor. You receive a message and ALWAYS return valid JSON with this shape:

{
  "name": string,
  "id": string (8 digits + 1 uppercase letter),
  "address": string,
  "phone": string (with international prefix),
  "bookingDate": "YYYY-MM-DD",
  "bookingTime": "HH:MM"
}

Rules:
- If a field is not in the message, return it as null.
- Do NOT include text outside the JSON.
- Do NOT use markdown blocks (no \`\`\`).`;

const EXAMPLES = `
Message: "I'm Luis Pérez (ID 87654321B). Book for June 1st at 2pm."
{ "name": "Luis Pérez", "id": "87654321B", "address": null, "phone": null, "bookingDate": "2026-06-01", "bookingTime": "14:00" }
`;

Step 3 — The function

async function extract(text: string): Promise<Booking> {
  const res = await client.messages.create({
    model: "claude-haiku-4-5",
    max_tokens: 500,
    system: SYSTEM,
    messages: [
      { role: "user", content: `${EXAMPLES}\n\nMessage: "${text}"` },
      { role: "assistant", content: "{" },
    ],
  });

  const raw = "{" + res.content[0].text;
  const json = JSON.parse(raw);
  return BookingSchema.parse(json);
}

Two key details:

1. Assistant prefill with { — forces the model to start with JSON, no intro text.
2. BookingSchema.parse() — if the model goes off-schema, it throws an exception you can catch and retry.

Step 4 — Tests

No tests, no production. Tests turn magic into engineering.

import { describe, it, expect } from "vitest";

describe("extract", () => {
  it("extracts all fields when present", async () => {
    const r = await extract(
      "I'm Ana García (12345678A), 12 Mayor St Madrid, +34 600 123 456. Booking 15/05 9pm."
    );
    expect(r.name).toBe("Ana García");
    expect(r.id).toBe("12345678A");
    expect(r.bookingDate).toBe("2026-05-15");
  });

  it("returns null for missing fields", async () => {
    const r = await extract("I'm Luis. I want to book on June 1st.");
    expect(r.name).toBe("Luis");
    expect(r.id).toBeNull();
  });

  it("normalizes time to HH:MM", async () => {
    const r = await extract("Book for 9 in the evening.");
    expect(r.bookingTime).toBe("21:00");
  });
});

Step 5 — Deploy

Drop your function into a serverless endpoint (Cloudflare Workers, Vercel, Deno Deploy) and you have a real extraction API.

Estimated cost: under €0.001 per extraction with claude-haiku-4-5.

What you have learned

• Anatomy of a prompt: role, context, instruction, examples, format.
• Zero-shot vs few-shot, and when each pays off.
• Chain-of-thought for reasoning.
• JSON prefill for reliable formatting.
• Validation with zod for production.

Next step

Want more? The next course is “Build an AI chatbot”, where we ship an assistant with conversational memory and tools. See you there.