Knowledge graph for finance
What a knowledge graph for finance is, and what it holds
A knowledge graph for finance stores financial facts as connected nodes: each number with its concept, its period, the entity that reported it and the source it came from. Software and AI can read it the way an accountant reads a ledger or a filing, and every number opens to what's behind it.
RoboSystems is the open-source knowledge graph for finance: your general ledger and 8,000+ public companies' XBRL filings, read by your AI over MCP, on your infrastructure, with the model you choose.
14 days on the SEC repository, no card. Graphs start at $99 a month.
What a knowledge graph for finance holds
A spreadsheet or a warehouse table keeps the number and leaves the meaning to whoever reads the column header. A knowledge graph keeps the meaning with the number, as six kinds of thing:
- Facts
- A number with its unit and its precision: $215,938 million, reported to the million.
- Concepts
- What the number is: revenue, gross profit, an operating lease liability. Labelled in XBRL’s standard terms wherever XBRL has one.
- Periods
- A span for flows like revenue, a single date for balances like cash. Fiscal years that end in different months stay distinct.
- Entities
- The company that reported it, its subsidiaries, and every public filer, each one a node the facts point to.
- Sources
- Where the number came from: the journal entry and the event behind a ledger balance, the filing and its tagged fact behind a public one.
- Relationships
- Which lines sum to which totals, which breakdowns belong to which statement, which entity owns which. The arithmetic is checkable.
A live example: one concept, two companies' tags
We asked the SEC graph for revenue from NVIDIA's and AMD's annual reports. The two companies tag it differently and close their fiscal years in different months. The graph answers both under one concept, and names each company's own tag. Figures in millions of US dollars, as the tool returned them on November 21, 2026:
NVDA
us-gaap:Revenues
| Fiscal year ended Jan 25, 2026 | 215,938 |
| Fiscal year ended Jan 26, 2025 | 130,497 |
| Fiscal year ended Jan 28, 2024 | 60,922 |
AMD
us-gaap:RevenueFromContractWithCustomerExcludingAssessedTax
| Fiscal year ended Dec 27, 2025 | 34,639 |
| Fiscal year ended Dec 28, 2024 | 25,785 |
| Fiscal year ended Dec 30, 2023 | 22,680 |
The call behind it, over REST (the same tool works over MCP):
curl -X POST "https://api.robosystems.ai/extensions/roboledger/sec/operations/build-fact-grid" \
-H "X-API-Key: $ROBOSYSTEMS_API_KEY" \
-H "Content-Type: application/json" \
-d '{"canonical_concepts": ["revenue"], "entities": ["NVDA", "AMD"], "form": "10-K", "period_type": "annual"}'Ask for "revenue" with NVIDIA's ticker and the graph resolves the phrase to the tag NVIDIA actually reports, us-gaap:Revenues, with 35 facts behind it. Ask for NVIDIA's income statement and it resolves the latest 10-K, filed February 25, 2026: gross profit of $153,463 million and operating income of $130,387 million for the fiscal year ended January 25, 2026, each traced to its tagged fact. Open that filing →
What it does for a finance team
The numbers tie. An agent that reads a ledger app through an export drafts a variance explanation nobody can check. In a knowledge graph every number opens to the entries behind it, so the reviewer checks the trail instead of redoing the work.
The books stay where they're allowed to live. The graph is open source. Run it in your cloud, on your own server, or let us run it, and point it at the model you choose. Your data trains no one's model.
The filings sit beside the books. Your ledger and the public filings share the RoboSystems ontology: the graph model both use, labelled in XBRL's standard terms wherever XBRL has one. So your margins and a public company's line up without a mapping spreadsheet.
How it differs from a data warehouse or a semantic layer
A data warehouse stores tables, and the meaning of a column lives in the queries people write against it. A semantic layer names metrics over those tables, so two dashboards agree on what "revenue" means. A knowledge graph stores the meaning with each fact: its concept, period, entity and source travel with the number, and so does how it rolls up.
That matters most when the reader is an AI. A model handed a table guesses at what a column means; a model reading a graph is told. Some writers call the events and decisions around these facts a context graph. In finance that context is part of the knowledge graph: see context graph.
Three kinds of graph on one platform
Your ledger and the public filings share the RoboSystems ontology, labelled in XBRL's standard terms wherever XBRL has one, so your numbers and a public company's line up. Graphs you model yourself take your own schema instead: define it, load it from Parquet, query it with Cypher. All three run on LadybugDB.
- Ledger graphs: RoboLedger, the general ledger built on RoboSystems, with close, consolidation, schedules, bank feeds, reconciliations and forecasting, every number traced to the entry.
- The SEC repository: 8,000+ public companies' 10-K, 10-Q, 20-F and 40-F filings, updated every weekday. The SEC EDGAR API →
- Your own graphs: bring a schema, load Parquet, query with Cypher.
Open source, so the graph stays yours
Open source (Apache 2.0 and MIT). Run it in your cloud, on your own server, or let us run it. You keep the code and the data whatever happens to any vendor, including us.
Point its AI at the model you choose: Claude, ChatGPT or any MCP client, or, when you host it yourself, an open-weight model on your own hardware.
Built by an inactive CPA and ex-Workiva knowledge engineer. Our own companies' books are kept on it.