SEC 10-K filings graph
25 annual reports (FY2021-FY2025) from 5 mega-cap companies - Alphabet, NVIDIA, Microsoft, Amazon, Apple - as one knowledge graph of the relationships they declare about each other: who competes with whom, which suppliers several of them depend on, who they have acquired, who regulates them, and how their risk factors shift year over year. Built from official SEC filings and queried in natural language through a working question-answering demo.
WHAT'S IN THE GRAPH
(:Filing)-[:HAS_SUBSIDIARY]->(:Company)
(:Filing)-[:HAS_RISK_FACTOR]->(:RiskTopic)
(:Company)-[:COMPETES_WITH]->(:Company)
(:Company)-[:DEPENDS_ON]->(:Company)
(:Company)-[:ACQUIRED]->(:Company)
(:Company)-[:REGULATED_BY]->(:Regulator)
The graph is built in layers. L1-L3 come straight from each filing's structured data with zero LLM involved: financials, segments and hedging from XBRL, subsidiaries and jurisdictions from the Exhibit 21 list, and risk-factor headings tracked as topics across five years. L4 is the layer that actually carries the story - competitors, dependencies, acquisitions and regulators - extracted from prose with an LLM and resolved down to one node per real-world company, since the same supplier or rival gets named differently across filings and years. L5 is derived on top with plain Cypher: year-over-year deltas, risks that appeared or disappeared, and competitor reciprocity.
WHERE THE DATA COMES FROM
- SEC 10-K filings - the annual report each public company is legally required to file, covering business description, risk factors, legal proceedings and management's discussion, item by item. sec.gov
- XBRL financial data - the machine-readable numbers behind the prose: revenue, segments, hedging positions. sec.gov
- Exhibit 21 - each filing's own list of subsidiaries and their jurisdictions.
EVERY FACT CARRIES ITS ADDRESS
A 10-K has a fixed table of contents - Item 1 (business), Item 1A (risk factors), Item 3 (legal proceedings), Item 7 (management's discussion). Every fact in the graph keeps the section it came from, so an answer can point back to "Alphabet, FY2024, Item 1A" instead of asking you to take it on faith.
QUESTIONS IT CAN ANSWER
- Competitor reciprocity - which of the competitors NVIDIA names also name NVIDIA back, and which name it one-way.
- Shared dependencies - which named partners more than one mega-cap admits depending on (in the filings it is Alphabet and OpenAI, not the supplier you would guess).
- Risk propagation over time - when AI regulation entered each company's risk factors, and in what order.
- Acquisitions vs. revenue - whose acquisitions since 2021 fed which revenue segment.
HOW QUESTIONS ARE ANSWERED
Questions are answered with GraphRAG: the query is grounded against the graph, the relevant subgraph is retrieved, and the answer is written from those facts with the filings and sections it relied on. Answer generation runs on OpenAI GPT-5.5, with GPT-4.1-mini handling the lighter routing and extraction steps.
WORKING DEMO
