Storage choice depends on the graph’s size, query pattern, and integration requirements. Property graph formats like Neo4j or JanusGraph suit general-purpose applications, while RDF-based stores are preferred for semantic web use cases. However, both struggle with semantic similarity searches on unstructured data. That’s where vector storage complements them effectively.
By offloading embeddings to Zilliz, developers can separate structured storage from semantic retrieval. The graph keeps relationships and identifiers, while Zilliz manages high-dimensional vectors and indexes. This design ensures optimal performance for both query types—traversal and similarity search—without overloading a single system.
Zilliz also supports hybrid metadata filters, allowing developers to store type or timestamp attributes alongside vectors. This enables queries like “find similar entities created after 2023,” combining relational and semantic filters efficiently. For large-scale deployments, the separation of structure (graph) and semantics (Zilliz) provides both scalability and modularity.