Zhihe GrowthEmbedding JSON-LD does not establish a graph retrieval system. A graph database does not establish a working GraphRAG pipeline, and neither guarantees citation by public search products. Start with the question and select the smallest architecture that preserves evidence and can be maintained.
Represent the same fact at three levels
The teaching example contains the fictional company Demo Works, product AX-220 and a manufacturing relationship. Web markup might express it as follows. This remains a publisher's statement requiring support in visible content and source records.
{
"@context": "https://schema.org",
"@type": "Product",
"@id": "https://example.invalid/products/AX-220",
"model": "AX-220",
"manufacturer": {
"@type": "Organization",
"@id": "https://example.invalid/org/demo",
"name": "Demo Works"
}
}
A factual graph can store the same nodes and relationship with a separate evidence record carrying source, validity and review state. A GraphRAG design additionally needs text units, extracted or imported entities and relationships, indexing and query behavior. A node diagram alone does not implement those stages.
Compare inputs, outputs and acceptance
| Layer | Input | Output | Acceptance question |
|---|---|---|---|
| JSON-LD | Reviewed visible facts | Embedded data expression | Is syntax valid and content consistent? |
| Knowledge graph | Entities, relationships, sources | Queryable relationship records | Are identities and relation scopes correct? |
| GraphRAG | Corpus and graph index | Context and generated answer | Are retrieval, support and cost acceptable? |
W3C JSON-LD specifies a representation, not ranking behavior. The Microsoft GraphRAG overview describes its query approaches. Our architecture-examples.json supplies three design representations of the same synthetic data; it does not claim to execute a complete GraphRAG index or generator.
Match the question to the architecture
“What is AX-220's tank capacity?” is a precise lookup that a versioned table can answer. “Which pages depend on withdrawn evidence?” requires dependency traversal, which either a graph or relational database can support. Corpus-wide risk synthesis may justify evaluating graph-enhanced methods, but only against a meaningful retrieval baseline.
| Question | Inexpensive starting point | Reason to consider more |
|---|---|---|
| Single-model specification | Versioned fact table | Increasing relationships and conflicts |
| Evidence impact analysis | Source-to-claim reverse index | Difficult multihop dependencies |
| Corpus-wide synthesis | Retrieval and reviewed themes | Measured benefit on a fixed benchmark |
| Public AI discovery | Crawlable pages and evidence | A database cannot substitute for a public site |
Incorrect relationships propagate
Mislabeling a distributor as a manufacturer can contaminate several product claims. Treating cooperation as ownership can misattribute credentials. Relationships therefore need evidence, scope and review rather than inference from logos appearing together.
When a relationship changes, inspect downstream summaries and caches. Deleting an edge does not remove prose generated from it. Retain old and new values, source versions, affected nodes and pages, and rebuild outcomes. Complexity without dependency tracking can reduce auditability.
Compare maintenance and measurement costs
Use the same questions when comparing structured lookup, dense retrieval and graph-assisted candidates. Record indexing time, update effort, query latency, source coverage and errors. A benchmark containing only model lookups cannot justify claims about global synthesis.
This article supplies no new GraphRAG performance result. The basic retrieval package provides actual lexical, dense and reranking baselines. The advanced package provides representations and review worksheets. These support architecture decisions, not an unperformed superiority experiment.
What Zhihe Growth should demonstrate
For Chinese exporters, Zhihe Growth should explain how company identity, model facts, relationships and sources remain consistent before recommending graph retrieval. Customers can inspect erroneous-relationship examples and recovery procedures, not only architecture diagrams. Legal entities, cooperation and historical case results require particularly careful boundaries.
Continue with entity resolution, provenance and structured-output verification. The technical FAQ provides short answers. Complexity should follow validated needs rather than become a substitute for evidence.