robosystems/sec-xbrl-knowledge-graphs

SEC XBRL Knowledge Graph — one embedded LadybugDB file

The XBRL filings made with the SEC since January 2024 — annual and quarterly reports (10-K, 10-Q), foreign-filer annuals (20-F, 40-F), proxy statements (DEF 14A) and registration statements (S-1) — as a single queryable property graph: 8,525 filers, 76,816 filings, 84.1 million facts on 294.5 million nodes, in one 130.3 GiB LadybugDB file that ships as a 39.3 GiB zstd archive, written by LadybugDB 0.18.1 — open it with that release (see Engine version below).

It is the same file that backs the hosted sec graph on robosystems.ai, built from EDGAR by the open-source RoboSystems pipeline. Download it, drop it into a local RoboSystems checkout, and you have the whole corpus under Cypher, the API, and MCP on your own machine — no server, no pipeline run.

This is not a Parquet dataset: it is a database file. The dataset viewer is disabled on purpose.

What is in the file

Snapshot	2026-08-31 (filings through 2026-08-28) — see the latest commit title for the current snapshot
Coverage	filings dated 2024-01-02 → 2026-08-28; prior-period comparatives inside those filings reach back to 2016
Forms	10-Q 45,241 · 10-K 16,394 · DEF 14A 9,962 · 20-F 3,006 · S-1 1,792 · 40-F 419 (and one S-3)
Filers	8,525 entities (6,875 with a ticker), with CIK, SIC code and description, legal name, fiscal year end
Facts	84.1 million XBRL facts (73.0 million numeric), each linked to its element, entity, period, unit, dimensions, and the presentation/calculation structure it appears in
Graph	14 node types · 22 relationship types · 294.5 million nodes
Engine	LadybugDB 0.18.1 (storage format v42) — the release that wrote the file and the one to open it with; see Engine version below
Sizes	sec.lbug.zst 39.3 GiB (42,247,675,436 bytes) → sec.lbug 130.3 GiB (139,865,260,032 bytes)
Cadence	monthly; the fixed path sec.lbug.zst is overwritten each snapshot, so main is always the latest and this URL never changes

Node types: Entity, Report, Fact, FactSet, Element, Label, Reference, Period, Unit, Dimension, Structure, Association, Classification, Taxonomy. The full schema — every property and relationship — is one call away once the file is open (CALL show_tables()), and is documented in the Querying the Analytical Graph wiki page.

Download and open

You need ~170 GiB free while decompressing (39.3 + 130.3; the archive is deleted afterwards), a machine with 16 GB RAM or more, and uv + just. Budget real time: this is 130 GiB on disk, not a laptop-minutes toy.

git clone https://github.com/RoboFinSystems/robosystems.git
cd robosystems
just sec-dump          # downloads sec.lbug.zst (resumable), verifies it, decompresses to data/lbug-dbs/sec.lbug

just sec-dump checks the archive's checksum and recorded size as it streams, refuses to overwrite an existing sec.lbug without --force, and warns if the engine that wrote the file differs from the one the checkout pins.

Without just — the archive is plain zstd:

hf download robosystems/sec-xbrl-knowledge-graphs sec.lbug.zst --repo-type dataset --local-dir .
zstd -d sec.lbug.zst -o data/lbug-dbs/sec.lbug      # or: python -m zstandard, 7-Zip, any zstd

Engine version

The file was written by LadybugDB 0.18.1 and must be opened with a release that reads that storage format (v42). An older release refuses the file; a newer one may open it, but a read-write open converts the storage format in place — a one-way change — so pin the version unless that is what you want. A RoboSystems checkout pins it for you (ladybug==0.18.1 in pyproject.toml); on your own:

pip install ladybug==0.18.1

The version that wrote each snapshot is in that snapshot's commit title (sec.lbug.zst: snapshot YYYY-MM-DD (ladybug X.Y.Z)), and it moves when the pipeline upgrades its engine — check it before opening a new snapshot with an existing install. just sec-dump compares the two and warns on a mismatch.

Query it

Directly, no services — the ladybug Python package (pip install ladybug==0.18.1) opens the file in-process:

just lbug-query sec "MATCH (e:Entity {ticker: 'NVDA'})-[:ENTITY_HAS_REPORT]->(r:Report) RETURN r.form, r.filing_date ORDER BY r.filing_date DESC"

import ladybug as lbug

db = lbug.Database("data/lbug-dbs/sec.lbug")
conn = lbug.Connection(db)
result = conn.execute("MATCH (e:Entity) RETURN count(e)")
while result.has_next():
    print(result.get_next())

Through the API and MCP — start the stack, mint a local key, and subscribe your local user to the sec repository; the graph is then served at http://localhost:8000 exactly as the hosted one is:

just start                 # API, Graph API, PostgreSQL, Valkey, Dagster
just demo-user             # writes an API key to .local/config.json
just demo-sec-subscribe    # grants the local user access to the sec graph (no data loading)

curl -X POST "http://localhost:8000/v1/graphs/sec/query/cypher" \
  -H "X-API-Key: $(jq -r .api_key .local/config.json)" \
  -H "Content-Type: application/json" \
  -d '{"query": "MATCH (e:Entity) RETURN count(e) AS filers"}'

MCP clients (Claude Desktop, Cursor, any MCP host) connect to http://localhost:8000/v1/graphs/sec/mcp with the same key — see AI Operators and MCP. The MCP server carries the schema, example queries, a concept resolver, statement builders, and fact-grid pivots.

Example queries

Consolidated annual revenue for one filer, robust across the different revenue elements filers use:

MATCH (f:Fact {has_dimensions: false})-[:FACT_HAS_ELEMENT]->(e:Element),
      (f)-[:FACT_HAS_ENTITY]->(ent:Entity {ticker: 'NVDA'}),
      (f)-[:FACT_HAS_PERIOD]->(p:Period {duration_type: 'annual'})
WHERE e.canonical_concept = 'revenue' AND f.numeric_value IS NOT NULL
RETURN ent.ticker, e.qname, p.end_date, f.numeric_value AS revenue
ORDER BY p.end_date DESC LIMIT 10

A whole income statement via the presentation structure (anchor on the entity first — thousands of filings share each statement type):

MATCH (ent:Entity {ticker: 'NVDA'})<-[:FACT_HAS_ENTITY]-(f:Fact {has_dimensions: false})-[:FACT_HAS_ELEMENT]->(e:Element),
      (f)-[:FACT_HAS_PERIOD]->(p:Period {duration_type: 'annual'}),
      (fs:FactSet)-[:FACT_SET_CONTAINS_FACT]->(f),
      (s:Structure {canonical_type: 'income_statement'})-[:STRUCTURE_HAS_FACT_SET]->(fs)
WHERE f.numeric_value IS NOT NULL
RETURN DISTINCT e.qname, f.numeric_value AS value, p.end_date
ORDER BY p.end_date DESC LIMIT 40

Revenue by segment (dimensional breakdown):

MATCH (f:Fact {has_dimensions: true})-[:FACT_HAS_ELEMENT]->(e:Element),
      (f)-[:FACT_HAS_DIMENSION]->(d:Dimension)
WHERE e.qname = 'us-gaap:Revenues' AND f.numeric_value IS NOT NULL
RETURN d.axis_uri, d.member_uri, f.numeric_value LIMIT 10

Two things every query should do: filter Fact.has_dimensions = false for consolidated totals (true when you want segment breakdowns), and anchor on an entity, report, element, or date before reaching Structure — an unanchored structure scan touches every filing.

Versioning

• sec.lbug.zst is a fixed path. Each monthly snapshot overwrites it in a new commit whose title carries the snapshot date and engine version (sec.lbug.zst: snapshot YYYY-MM-DD (ladybug X.Y.Z)). main is always the latest.
• The engine version can move between snapshots. The commit title carries it; when it changes, install the matching ladybug before opening the new file (a RoboSystems checkout gets the new pin with the release that shipped it).
• Superseded snapshots are removed to keep the dataset one snapshot deep; pin a commit with --revision only while it is still listed.
• The hosted graph at robosystems.ai is rebuilt nightly; this file is that graph as of the snapshot date.

Provenance and licence

This dataset — the compiled graph, its schema, and the cross-filer normalisation — is released under CC BY 4.0: use it for anything, including commercially, with attribution (citation below). The source data is SEC EDGAR XBRL, public domain. The pipeline that parses, normalises, classifies, and materialises it is RoboSystems, Apache-2.0. Facts are as filed: restatements appear as later filings, not as corrections to earlier ones. Element names follow the filer's taxonomy (us-gaap, ifrs-full, extensions); canonical_concept and Structure.canonical_type are RoboSystems' cross-filer normalisation on top.

RoboSystems. SEC XBRL Knowledge Graph (LadybugDB). Hugging Face, snapshot 2026-08-31.
https://huggingface.co/datasets/robosystems/sec-xbrl-knowledge-graphs

Issues and questions: github.com/RoboFinSystems/robosystems/issues.