> Parel documentation. Index of every page: https://parel.si/docs/llms.txt. This page in HTML: https://parel.si/docs/tools/extract_structured

# extract_structured

> One web page as the records you describe: a table of names and titles, prices or any fields, from a sentence or a schema. ₹1.13 at most per page, settles lower; no records is ₹0.

## What it does

One web page read into the records you describe, from `url` plus a plain-English `instruction` ("every person's name and job title"), a JSON `schema` for one record, or both. `data` is the list of records, one object per entry found, in the page's order, so a team page, a price list or a directory comes back as a table in one call. `example` (a sample of the shape you want) can stand in for a schema.

Use it when you want fields rather than prose: [`read_page`](https://parel.si/docs/tools/read_page.md) gives the page as Markdown, [`extract_page`](https://parel.si/docs/tools/extract_page.md) recognises an article or product on its own, and for a list that spans several pages [`map_site`](https://parel.si/docs/tools/map_site.md) lists the page URLs first and this tool then runs once per page. In the live check it returned 40 people with titles from the first page of a 98-person team list in one call; the other pages need their own calls.

India: no locale knob; the page is read as served.

## Inputs

Generated from the contract captured from staging on 2026-11-05.

| Field | Type | Required | Description |
|---|---|---|---|
| `url` | string | one of | e.g. "https://news.ycombinator.com" |
| `schema` | object | one of | JSON schema each returned record must match |
| `example` | object | no | a sample of the shape you want (structure, not values) |
| `instruction` | string | one of | e.g. "the top 3 stories on the front page" |

One of these sets must be complete (anything else is `invalid_input`, ₹0): `url` + `instruction`; `url` + `schema`.

## What drives the cost

The size of the page: the source bills a base for fetching it plus the model's work, by blocks of text. The price shown is the price of a large page and the charge settles at the page's real cost; the 40-person team page settled at about a quarter of it. A page that yields no records, including one that blocks automated readers, is `class: "ok_free"`, ₹0. Usually two to ten seconds.

## Price today

| | |
|---|---|
| Price today | **₹1.13** per page (`cost_paise: 113`) at ₹93.93 per US dollar on 17 November 2026 |
| Hold before the call | ₹1.70 (`hold_paise: 170`); when the call settles the real price is charged and the unused part is released |
| ₹0 when | it fails, or the provider answers but there is nothing in it (an empty list, no email, no page): `class: "ok_free"`, ₹0 |
| Timeout | 30 s, then `upstream_timeout` (₹0, retryable) |
| Over MCP | `run_tool` with `tool: "extract_structured"` |
| Over REST | `POST https://api.parel.si/v1/run` with `"tool": "extract_structured"` |
| Confirmation | not needed (the estimate is under the ₹200 threshold) |

## Request

```bash
curl https://api.parel.si/v1/run \
  -H "Authorization: Bearer $PAREL_KEY" \
  -H "Content-Type: application/json" \
  -d '{"tool":"extract_structured","input":{"url":"https://americanindustrial.com/team","schema":{"type":"object","required":["name"],"properties":{"name":{"type":"string"},"title":{"type":"string"}}},"instruction":"every person listed on the page: their full name and their job title"}}'
```

## What comes back

`data`, the list of records. A real answer recorded during the live check of 2026-11-04 (americanindustrial.com/team, 40 records in 4.3 s), trimmed to four:

```json
[
  {
    "name": "Justin Antony",
    "title": "Operations"
  },
  {
    "name": "Drew Barot",
    "title": "Credit Opportunities,Transactions"
  },
  {
    "name": "Eric Baroyan",
    "title": "Firm Leadership,Transactions"
  },
  {
    "name": "John Becker",
    "title": "Managing Partner Emeritus,Operations"
  },
  "… 36 more item(s) truncated"
]
```

## Notes

- Per page. The price shown is the maximum for a large page; the charge settles at the page's real cost, usually about a quarter of it. No records is free.
- schema describes one record (type object with properties); data[] is the list of records. instruction alone works for simple asks; both together are most reliable.
- One URL per call: a list that spans several pages needs one call per page (map_site finds them).
- Usually two to ten seconds. A page that blocks automated readers comes back with no records, free.

## Siblings

- [`map_site`](https://parel.si/docs/tools/map_site.md) to find every page of a list first.
- [`extract_page`](https://parel.si/docs/tools/extract_page.md) when the page is an article or a product and you want its usual fields.
- [`read_page`](https://parel.si/docs/tools/read_page.md) for the page as Markdown.
