# Citation

A **citation** links a span of generated text to the source it came from — with the quoted evidence
and, for PDFs, the exact region on the page. This is the object that makes output checkable.

## Shape

```json
{ "object": "citation", "index": 1,
  "quote": "We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs.",
  "source": { "type": "document", "id": "doc_8ba2f01c47",
              "title": "Attention is all you need.pdf", "url": "…" },
  "locations": [ { "page": 7, "bbox": { "x": 0.252, "y": 0.426, "width": 0.495, "height": 0.041 } } ] }
```

## The locator contract

- `page` is **1-indexed**.
- `bbox` values are fractions of page width and height, `0`–`1`, origin **top-left**.
- Multiply by your rendered page size to draw a highlight at any zoom level. No DPI assumptions.
- `locations` may hold more than one box when a quote spans columns or pages.

## `source.type`

| Value | Meaning | Has `quote`/`locations`? |
|---|---|---|
| `paper` | corpus paper | quote yes; locations only where we hold the PDF |
| `document` | your upload | yes |
| `model` | model-asserted, no retrieved source | no |

## Not to be confused with

**Formatted citation** — *a citation says where a claim came from; a formatted citation is what you
put in a bibliography.* Those come from citations. And `citation_count` on a Paper is
neither: it counts how often that paper has been cited by others.

## Related

answer · citations · citation-highlights · paper