Knowledge graphs — recap & real-world examples
Real-world examples & war stories
Google's "things, not strings" (2012). Google launched its Knowledge Graph to stop matching text and start knowing what text refers to — entities with identities, properties, and relationships, powering the info panels beside search results. It gave the field its name and its cleanest slogan. 🎯 Takeaway: the shift that mattered wasn't a database. It was treating identity as the product. The moment "jaguar" became three things instead of one string, whole product surfaces became possible.
The Panama Papers (2016). Handed 11.5 million leaked documents, the ICIJ loaded extracted entities — companies, officers, addresses — into a graph database and let hundreds of journalists traverse: from a politician to a shell company to a shared registered agent to the next shell. Connections no keyword search could surface fell out as paths. 🎯 Takeaway: traversal is a capability, not a report. When the question is "how is A connected to B?", the graph doesn't just answer faster — it answers questions you couldn't previously ask. (Note also what did the heavy lifting first: extraction and entity resolution across millions of messy documents.)
Amazon's Product Graph (2018→). Amazon set out to build an authoritative graph of products and everything about them — brands, attributes, compatibility, variants — feeding search, recommendations, and Alexa answers. The published lesson: the hard science was never storage. It was extraction from messy seller text and reconciling conflicting sources at catalog scale. 🎯 Takeaway: exactly the 80% claim — the pipeline is the product. Even with world-class infrastructure, the budget and the breakthroughs live in extraction and resolution.
LinkedIn's Economic Graph. LinkedIn frames its entire asset as one graph — members, companies, skills, jobs, schools — and its features as traversals of it. "People you may know" is link prediction. Recruiter search is multi-hop filtering. Economic research is aggregate graph analytics. 🎯 Takeaway: when the relationships are the product, the graph isn't infrastructure behind the product — it is the product. Every feature is a different query against the same compounding asset.
NASA's lessons-learned graph. NASA connected decades of siloed mission documents, incident reports, and engineering lessons into a knowledge graph, so an engineer designing a component can find what past programs learned — across programs, decades, and vocabularies that never matched. 🎯 Takeaway: the ontology work is organizational. The technical graph was the easy half. Agreeing what a "lesson," an "anomaly," and a "subsystem" mean across generations of programs was the actual project.
Microsoft GraphRAG (2024). Microsoft Research showed that vector-only RAG fails on global questions ("what are the main themes in this corpus?") and multi-hop questions. Building a graph over the corpus, detecting communities, and pre-summarizing them lets an LLM answer both, with sources. The pattern went from paper to product roadmaps in about a year. 🎯 Takeaway: the graph-LLM marriage is not vendor fog — it targets measured, structural failures of similarity-only retrieval. But adopt it the way the lesson says: after your evals show those failures, not before.
The enterprise graveyard (ongoing). Every data leader knows one: the two-year enterprise-ontology program with a 400-page model and no shipped feature, the pilot that wowed and then decayed unstewarded, the "graph initiative" renamed twice and quietly absorbed into a BI team. These don't make conference talks, which is why the pattern repeats. 🎯 Takeaway: the failures are portfolio failures — no killer query, no visible feature, no steady state — and the quiet death arrives without a postmortem. Sequencing and stewardship, not technology, decide survival.
Module recap
| Lesson | The one idea | The question it makes you ask |
|---|---|---|
| What is a knowledge graph? | Things, not strings — connections made explicit and queryable | Which valuable questions are multi-hop across our systems? |
| Ontologies & data modeling | The ontology is a product contract, negotiated not specified | What can this ontology never answer — and did we choose that? |
| Building the graph | The pipeline is the product; entity resolution is the boss fight | Who reviews the low-confidence matches every week? |
| Storage & querying | Query shapes pick the store; benchmark boring first | What's our deepest traversal, at what latency, in which feature? |
| Reasoning & analytics | Five families, one trust gradient — paths explain, predictions suggest | Is this answer traversed, inferred, or guessed — and does the UX say so? |
| Knowledge graphs & LLMs | The graph grounds the model; the model builds the graph — through review | Can every AI claim cite fact, source, and permission? |
| Governance, quality & trust | Trust is manufactured: provenance, metrics, permissions, owners | Can we show why the graph asserts this — and who may see it? |
| Knowledge graphs as a product | Fund features, sequence adjacent domains, let the asset compound | What ships in two quarters that a user will notice? |
The through-line: every lesson is an annotation on the module's own knowledge graph. An ontology gives meaning. A pipeline manufactures the asset. Storage serves it. Reasoning and LLMs cash it into features. Governance keeps it believable. The product loop funds the next domain. The organizations that win with knowledge graphs aren't the ones with the biggest node counts or the fanciest databases. They're the ones that picked killer queries, paid the unglamorous resolution-and-curation bill, showed their work on every fact, and let the asset compound one adjacent domain at a time.
Walk-away question: "For my knowledge graph: can I name the killer query, the first visible feature and its quarter, the steward behind each domain, the four quality numbers on the dashboard, the provenance chain behind any AI answer — and the moat in one sentence that doesn't mention a database?"
If yes — you're not buying a graph, you're building a compounding knowledge asset. If no, you now know exactly which lesson to reread.