JSON schema
The PaperJson, MetadataJson and Outline shapes, a trimmed example for 1706.03762, and the schema 1 compatibility promise.
/1706.03762.json returns a PaperJson. /1706.03762/metadata.json returns a MetadataJson. /1706.03762/outline.json returns an Outline. All carry schema: 1.
PaperJson
| Field | Type | Notes |
|---|---|---|
schema |
1 |
Major schema version |
id, version, latestVersion, idv |
"1706.03762", 7, 7, "1706.03762v7" |
|
urls.arxiv |
{ abs, pdf, html } |
arXiv URLs for this version |
urls.arxivv |
{ json, md, txt, outline } |
arXivv URLs for this version |
metadata |
object | From arXiv: title, authors, abstract, categories, primaryCategory, license?, doi?, journalRef?, comments?, reportNo?, versions[] |
build |
object | source, builder, builtAt, quality: { score, stats } |
tokens |
{ approx, method: "chars/4" } |
Approximate |
content |
Document |
The document tree |
inferred |
Inferred |
See inferred metadata |
{
"schema": 1,
"id": "1706.03762",
"version": 7,
"latestVersion": 7,
"idv": "1706.03762v7",
"urls": {
"arxiv": {
"abs": "https://arxiv.org/abs/1706.03762v7",
"pdf": "https://arxiv.org/pdf/1706.03762v7",
"html": "https://arxiv.org/html/1706.03762v7"
},
"arxivv": {
"json": "https://arxivv.org/1706.03762v7.json",
"md": "https://arxivv.org/1706.03762v7.md",
"txt": "https://arxivv.org/1706.03762v7.txt",
"outline": "https://arxivv.org/1706.03762v7/outline.json"
}
},
"metadata": {
"title": "Attention Is All You Need",
"authors": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin",
"abstract": "The dominant sequence transduction models are based on ...",
"categories": ["cs.CL", "cs.LG"],
"primaryCategory": "cs.CL",
"versions": [{ "version": 1, "submittedAt": "2017-06-12T17:57:34Z" }]
},
"build": {
"source": "arxiv-html",
"builder": "0.1.0",
"builtAt": "2026-10-14T09:00:00Z",
"quality": {
"score": 0.97,
"stats": {
"errors": 1,
"sections": 7,
"paragraphs": 120,
"mathNodes": 180,
"mathWithoutLatex": 0,
"cites": 60,
"unresolvedCites": 1,
"figures": 5,
"figuresWithoutImages": 0
}
}
},
"tokens": { "approx": 10240, "method": "chars/4" },
"content": {
"title": [{ "type": "text", "text": "Attention Is All You Need" }],
"body": [
{
"type": "section",
"id": "S1",
"level": 1,
"number": "1",
"title": [{ "type": "text", "text": "Introduction" }],
"children": [
{
"type": "paragraph",
"id": "S1.p1",
"content": [{ "type": "text", "text": "Recurrent neural networks ..." }]
}
]
}
]
},
"inferred": { "status": "pending", "queuedAt": "2026-10-14T09:00:01Z" }
}
Fields and arrays are trimmed in this example.
MetadataJson
/1706.03762/metadata.json is the cheap way to look a paper up. It never builds the paper and ignores Prefer.
| Field | Type | Notes |
|---|---|---|
schema, id, version, latestVersion, idv, urls, metadata |
As in PaperJson |
|
build? |
object | As in PaperJson. Present only once the tree has been built |
inferred? |
Inferred |
Present only once the tree has been built; pending until enrichment finishes |
A paper that has never been built returns only the arXiv metadata and URLs. Request any format to build it.
Outline
| Field | Type | Notes |
|---|---|---|
schema, id, version, idv |
As in PaperJson |
|
title, authors[], primaryCategory, submittedAt |
authors split from the arXiv author string on , and and |
|
inferred |
{ status, tldr?, contributions?, paperType? } |
status always; the rest once ready |
tokens |
{ approx, method: "chars/4" } |
Whole paper |
abstract |
string | |
sections[] |
{ id, alias, number?, title, tokens, children } |
Body and appendices, nested; tokens from each section’s Markdown |
figures[], tables[] |
{ id, number?, caption } |
Caption as plain text; subfigures listed too |
equations, references |
number | Display equations and bibliography entries |
{
"schema": 1,
"id": "1706.03762",
"version": 7,
"idv": "1706.03762v7",
"title": "Attention Is All You Need",
"authors": ["Ashish Vaswani", "Noam Shazeer"],
"primaryCategory": "cs.CL",
"submittedAt": "2017-06-12T17:57:34Z",
"inferred": {
"status": "ready",
"tldr": "Introduces the Transformer ...",
"paperType": "method"
},
"tokens": { "approx": 10240, "method": "chars/4" },
"abstract": "The dominant sequence transduction models ...",
"sections": [
{
"id": "S3",
"alias": "S3",
"number": "3",
"title": "Model Architecture",
"tokens": 3100,
"children": [
{
"id": "S3.SS2",
"alias": "S3.2",
"number": "3.2",
"title": "Attention",
"tokens": 1200,
"children": []
}
]
}
],
"figures": [{ "id": "S3.F1", "number": "1", "caption": "The Transformer - model architecture." }],
"tables": [
{ "id": "S6.T2", "number": "2", "caption": "The Transformer achieves better BLEU scores ..." }
],
"equations": 3,
"references": 40
}
Other shapes
| Endpoint | Shape |
|---|---|
/1706.03762/references.json |
{ schema, idv, references: Reference[] } |
/1706.03762/figures.json |
{ schema, idv, figures: FigureJson[], tables: TableJson[] } |
/1706.03762/figures/F1.json |
FigureJson |
/1706.03762/assets/<path> |
The image file itself (see figure assets) |
| Any error | Problem (see errors) |
| Type | Fields |
|---|---|
Reference |
id, number?, text, content, arxivId?, doi?, url? |
FigureJson |
id, number?, caption (plain text), images: [{ url, alt? }], referencedBy |
TableJson |
id, number?, caption (plain text), rows: Cell[][], referencedBy |
Problem |
type, title, status, detail?, instance?, code |
images[].urlis absolute:https://arxivv.org/1706.03762v7/assets/Figures/ModalNet-21.png. It returns404until the asset is mirrored.- A figure’s
imagesare only its own direct images. Subfigures are listed as figures of their own, with their ownimages. referencedByholds the ids of paragraphs whosereftargets the node (or a node inside it).
Compatibility
The schema field and the X-Arxivv-Schema header carry the major version.
| Change | Within schema 1? |
|---|---|
| New optional field | Yes |
| New node type | Yes |
| Rename or removal | No: new major |
| Structural change | No: new major |
- Ignore fields and node types you do not recognise.
- A new major ships alongside the old one.
- The old major stays available for a 6-month deprecation period.
- MCP tools and the API reference follow the same version.