EU law · 24 official languagesArticle-level JSON · typed citation graphMCP server · Claude, Cursor, VS CodeSDKs · Python, TypeScriptFree plan · 500 credits to start
Brussels · CET
Live·Corpus & Provenance
What we store, how we version it

The Corpus

Not a scraper.
A mirror with memory.

Every document we serve comes from our own corpus — content-hash-versioned and provenance-tagged. A EUR-Lex document we don't hold yet is fetched live, in the same shape.

document recordv4 · stable
{
  "celex": "32016R0679",
  "language": "en",
  "version": 4,
  "contentHash": "sha256:7a3f…b21e",
  "source": "delta",
  "firstSeenAt": "2024-03-12T08:14Z",
  "lastSeenAt": "2026-05-12T06:02Z",
  "fetchedAt": "2026-05-12T06:02Z",
  "contentChangedAt": "2025-11-18T11:47Z"
}
Every row is reproducible
Document records mirrored
—
 
Citation edges
—
typed & directional
New or updated, past 7 days
—
 
Last ingest
—
auto-refreshing
§ 01 — Live ingestion

Watch the corpus grow.

The newest full-text documents to land in the corpus — the same data the API serves.


Loading the feed…

Try it — semantic search, live against this corpus
Capped demo — top 3 hits, plain retrieval, rate-limited.The API adds filters, HyDE rewriting, full text and the citation graph.Get a free key
§ 02 — Notable developments

What's worth reading.

Recent cases and acts our screening flagged as worth a closer look.


§ 03 — Versioning

Content-hash versioned. Immutable history.

When the parsed text changes, the prior version is kept and the new one becomes current.


N° 01

Every version kept

The current text per (celex, language), plus one entry per distinct version, addressable by (celex, language, version).

N° 02

Hashes, not timestamps

A new version is recorded only when the SHA-256 of the parsed text changes — a corrigendum that flips one word counts.

N° 03

Immutable history

A new version never rewrites an old one. List and fetch past versions with /documents/:celex/versions.

§ 04 — Provenance

Where every row came from.

Every document carries a source tag explaining how it entered the corpus, plus three timestamps for first sight, last sight, and current-version fetch.


Source tags
TagMeaning
apiFetched live for an API call, then kept.
deltaPicked up from EUR-Lex as newly published.
backfillAdded from EUR-Lex by the historical backfill.
stub-backfillA metadata-only record later completed with full text.
citation-extractorMetadata recorded during citation extraction.
manualRefreshed on demand.
Timestamps
firstSeenAt
When the document first entered the corpus.
lastSeenAt
The most recent successful refetch, changed or not.
fetchedAt
The most recent fetch that returned the full content.
contentChangedAt
When the parsed text last changed — the version boundary. Pin against this for reproducible reads.
§ 05 — Citation Graph

A typed graph, not just hyperlinks.

Citations from EUR-Lex metadata are typed, directed edges. Walk them in either direction.


N° 01

Classified edges

Each citation carries a citationType: reference, amendment, repeal, corrigendum, implementation, legal-basis, or proposal.

N° 02

Both directions

/citations/cites/:celex walks outbound; /citations/cited-by/:celex walks inbound. Use /citations/network/:celex to fetch both with one call.

N° 03

No duplicates

Extraction is idempotent: each citation edge is stored once, however often a document is re-read.

§ 06 — Guarantees

What you can rely on.


  • N° 01

    The same (celex, language, version) tuple always returns the same parsed content. Hash-pinned reads are reproducible.

  • N° 02

    Citations are directional and typed, and their classification never changes silently.

  • N° 03

    Every response carries enough metadata (contentHash, fetchedAt, source) for an auditor to reconstruct exactly what we served and when.

  • N° 04

    We don't transform legal text — we structure it.

⁂ ⁂ ⁂

Build on a corpus
you can audit.

Hash-pinned reads, typed citations, full provenance. Free for your first 500 credits, then 150 every month.