Open content kit · 23 July 2026

Build with a year of BMJ content.

Three public, read-only R2 datasets for rapid prototyping: article metadata, open-access full text, and Cloudflare-native BGE vectors.

32,894article records

87journals requested

222full-text articles

384vector dimensions

Choose a dataset

Every dataset has a machine-readable catalog and a paginated listing API. No read credentials are required.

R2 · JSON

Titles, abstracts & authors

BMJ journal article metadata from 21 July 2025 through 21 July 2026, preserving the source journal directory structure.

Articles32,894

Fieldstitle · abstract · authorship

Accesspublic read

Open catalog Journal index Browse prefixes

R2 · Open access

Article full text

Full-text JSON for three open-access titles: BMJ Diabetes Research & Care, BMJ Medicine and BMJ Oncology.

Articles222

Journals3

Accesspublic read

Open catalog Export manifest Browse prefixes

R2 · Workers AI

Full-corpus BGE vectors

A normalized float32 matrix aligned to 32,814 usable metadata records and the embedding space used by Cloudflare Workers AI.

Model@cf/baai/bge-small-en-v1.5

Shape32,814 × 384

Matrix50.4 MB · float32 LE

Open catalog Storage guide Vector sidecar

Dataset documentation

R2 object keys describe where a file lives. JSON property keys describe the data inside that file. Both contracts are documented below.

Buckets and object keys

R2 is a flat object store. The slash-separated “folders” shown here are key prefixes retained from the source export.

Bucket	Objects	Object-key patterns	Contents
metadata	32,984	`<journal-id>/<YYYYMMDD>_<uid>.json` `_state/<journal-id>.json`	32,894 article files, 87 journal state files, export and journal manifests, and a generated catalog.
fulltext	227	`<journal-id>/<YYYYMMDD>_<uid>.json` `_state/<journal-id>.json`	222 full-text articles from `bmjdrc`, `bmjmed` and `bmjonc`, plus state, manifest and catalog files.
vectors	4	`title_abstract.cf-baai-bge-small-en-v1-5.{f32,json}`	A binary float32 matrix, its ordered-ID sidecar, the vector storage guide and a generated catalog.

Public read API

All four hosts are served by one read-only Worker. No read token is required, writes are rejected, CORS is open, and large objects support byte ranges.

GET /Dataset description, catalog URL and useful entry points.

GET /catalog.jsonEvery source key with byte size, SHA-256 and public URL.

GET /__list?prefix=bmjopen/&limit=100Live paginated listing. Follow next_url until truncated is false.

GET /<object-key>Fetch an exact object. HEAD, conditional requests and Range are supported.

Article JSON

Article files use a deliberately compact normalized schema. The metadata and full-text exports share the same identity, title, abstract and author fields.

Metadata article

Found under a journal prefix in the metadata bucket.

uidstringStable UUID used in the filename and in the ordered vector sidecar IDs.
titlestringArticle title as plain text.
abstractstringNormalized abstract text.
authorsarrayOrdered author objects.
authors[].namestringDisplay name for an author.
authors[].affiliationsstring[]Zero or more affiliation strings for that author.

{
  "uid": "6908057e-2698-5153-bf0f-134f410dd4d3",
  "title": "…",
  "abstract": "…",
  "authors": [
    { "name": "…", "affiliations": ["…"] }
  ]
}

Full-text article

Uses every metadata-article key above and adds one field.

full_textstringNormalized full article text for the open-access article.

Index and manifest JSON

These files support discovery, reproducibility and integrity checking. Most prototypes should begin with catalog.json or journals.json.

`catalog.json`

Key	Type	Meaning
`schema_version`	number	Catalog schema version.
`dataset`, `title`	string	Machine and display names.
`access`	string	Currently `public-read`.
`public_base_url`	string	Dataset host used to build object URLs.
`generated_at`	string	ISO 8601 catalog generation time.
`source_object_count`	number	Number of source objects, excluding the catalog itself.
`source_total_bytes`	number	Total size of source objects.
`files[]`	array	Objects with `key`, `bytes`, `sha256` and `url`.

`journals.json`

Key	Type	Meaning
`retrieved_at`	string	ISO 8601 source retrieval time.
`journals[]`	array	Journal records.
`journals[].id`	string	Journal identifier used as the R2 key prefix.
`journals[].title`	string	Journal display title.
`journals[].issn_l`	string	Linking ISSN.
`journals[].document_count`	number	Source-system document count, not the 12-month export count.

`manifest.json` and `_state/*.json`

The export manifest records schema_version, status, created_at, completed_at, date_range, requested fields, journals, results, batching configuration and batch_history. Per-journal state files record status, timestamps, articles_written, completed/pending windows, split counts, source signatures and DOI-to-file mappings in article_files. Full-text state also includes article_ranks.

Vector snapshot

The binary matrix and JSON sidecar are inseparable. Row order is defined by the sidecar, and the BGE space must not be mixed with AllenAI SPECTER vectors.

Sidecar key	Type	Meaning
`model`	string	`@cf/baai/bge-small-en-v1.5`, the Workers AI checkpoint.
`dims`	number	384 float32 values per vector.
`count`	number	32,814 matrix rows.
`normalized`	boolean	`true`; each row is L2-normalized.
`ids[]`	string[]	Ordered article UUIDs. `ids[N]` identifies matrix row `N`.

vector N =
  floats[N * dims : (N + 1) * dims]

binary format =
  little-endian, row-major float32

Start in seconds

Fetch catalogs from a browser, notebook, Worker or command line. Cross-origin GET and range requests are enabled.

# Find BMJ Open objects
curl 'https://metadata.bmj-cloudflare-hackathon.com/__list?prefix=bmjopen/&limit=100'

# Download the vector contract and matrix
curl -O https://vectors.bmj-cloudflare-hackathon.com/title_abstract.cf-baai-bge-small-en-v1-5.json
curl -O https://vectors.bmj-cloudflare-hackathon.com/title_abstract.cf-baai-bge-small-en-v1-5.f32