Industry Knowledge Graph (GraphRAG)
Fact layer plus science layer, a four-layer causal chain, traceable reasoning and a model-agnostic API.
Graph retrieval-augmented generation (GraphRAG) solves the problem of answers that sound right but cannot explain why. We split domain knowledge into two assets: a fact layer that answers "what is in this product" and a science layer that answers "through which mechanism does this element affect that endpoint".
Two data assets
The fact layer is built from regulatory filings and other authoritative sources; the science layer from literature, patents and supplier clinical data.
Four-layer causal chain
Element → target → mechanism → measurable endpoint. Each chain carries an evidence grade (strong / moderate / weak), in-vitro / in-vivo / clinical provenance, effective concentration ranges and citations.
Evidence-chain-constrained generation
Generation must reason along the chain; hallucination is suppressed at the architecture level and every statement traces to a source.
Model-agnostic API
Applications call the graph through one interface. Swap the model without losing capability; no vendor lock-in.
In production: skincare vertical
| Dimension | Product database (fact layer) | Knowledge graph (science layer) |
|---|---|---|
| Question answered | What ingredients are in product X | Through which target / mechanism does ingredient A produce which effect |
| Sources | NMPA regulatory filings | Literature / patents / supplier clinical data |
| Scale | 740K+ skincare products | 811 de-duplicated active ingredients × four-layer causal chain |
| Product-level coverage | — | 72% (active-ingredient basis) |
| Evidence grading | — | strong / moderate / weak; in-vitro / in-vivo / clinical; concentration ranges; citations |
These are vertical-industry figures, not a cross-industry scale. The same four-layer methodology extends to adjacent regulated categories such as health and food.
- A listed cosmetics company: R&D moved from experience-driven to data-driven on AI-Ready Data
Start with one concrete scenario, not a product list.
Tell us which scenario is still not in production: where the data lives, who uses it, and who signs off. One workshop gives you a verifiable path forward.