Retrieval Architecture Spectrum
The Three Approaches
Retrieval architectures form a spectrum from similarity to structure:
Vector Search (Similarity-Based) Embeds queries and documents into the same space, retrieves by proximity. Finds what resembles the question.
Knowledge Graphs (Relationship-Based) Represents entities as nodes and connections as typed edges. Finds what relates to entities in the question, traversable via explicit links.
Ontology-Driven (Category-Based) Structures knowledge into hierarchical classes with inherited properties. Finds what belongs to the same categorical structure as the question.
What Each Optimizes
| Approach | Optimizes For | Struggles With |
|---|---|---|
| Vector Search | Directness, specific answers | Multi-hop reasoning, synthesis |
| Knowledge Graphs | Relationship traversal | Preserving categorical hierarchy |
| Ontology-Driven | Comprehensiveness, diversity | Directness, computational cost |
The Tradeoff Pattern
No single architecture wins across all query types.
Vector search excels at “find the passage that answers this” (local retrieval with direct answers. In comparative studies, it wins 90%+ on directness while losing badly on comprehensiveness.
Graph-based approaches enable multi-hop reasoning but cluster by co-occurrence rather than categorical belonging. They capture what appears together, not what belongs together.
Ontology-driven approaches preserve hierarchical structure but add computational overhead and sometimes bury specific answers inside broader categorical context.
When Each Applies
Use vector search when:
- Questions have single, locatable answers
- Speed matters more than completeness
- Domain has low relational complexity
Use knowledge graphs when:
- Answers require traversing explicit relationships
- Multi-hop reasoning is common
- Relationship types are well-defined
Use ontology-driven when:
- Questions require synthesizing across documents
- Categorical structure carries domain meaning
- Comprehensiveness outweighs response time
The Integration Question
The most capable systems likely combine approaches: vector search for initial candidate retrieval, graph traversal for relationship expansion, ontological structure for categorical coherence.
The challenge is knowing which to invoke for which query, a routing problem that most current implementations don’t address.
Related: 07-molecule—vectors-vs-graphs, 05-molecule—rag-architecture-taxonomy