On the Gap Between Retrieval and Understanding
Embeddings are a big step towards representing knowledge in a format that can undergo significant computation.
A vector search returns chunks ordered by geometric proximity in a high-dimensional space. What it cannot tell you is whether those chunks, taken together, constitute a coherent argument, an incidental co-occurrence, or an outright contradiction. The knowledge graph exists to supply what the vector store cannot. Where the vector store answers "what is near?", the graph answers "what is related, and how?"
The Neo4j layer in this system captures not just that two concepts appear in the same corpus, but that they belong to the same subdomain, share a measurable intersection strength, and can be reached by a traceable Cypher path. That distinction matters enormously the moment you want to generate a question that requires synthesis across domains rather than recall within one.