Terraphim uses graph embeddings instead of neural vector embeddings. Terms and concepts are represented as nodes in a role-specific knowledge graph, with relationships encoded as edges. Matching a query against a graph is deterministic, auditable, and fast.
See the full design note for how RoleGraph composes Aho-Corasick matching with PageRank ordering to produce ranked results in sub-millisecond time, without any floating-point vector math.
Why graph embeddings
- Deterministic: same query, same graph, same result. No stochastic retrieval.
- Explainable: every matched node traces back to a term in the thesaurus.
- Efficient: 15-20 MB RAM for a typical graph, no GPU required.
- Domain-adapted: each role has its own vocabulary and relationships.
This is a fundamentally different model from dense vector retrieval. It is what makes Terraphim run on a laptop, a Raspberry Pi, or inside a browser extension.