Vector database vs graph database

A vector database stores embeddings — points in a high-dimensional space — and answers nearest-neighbour queries: given this vector, what's close? Closeness approximates semantic similarity, which makes it the right index for unstructured text, images and audio.

A graph database stores entities and typed relationships and answers traversal queries: from this node, what's connected, through what, how far? The relationships are explicit data, so the answers are exact, repeatable and explainable rather than ranked by similarity.

Semantic recall at scale. "Find passages about deployment rollbacks" works even when no document contains those words, because similarity lives in meaning-space. Fuzzy matching, cross-lingual recall, deduplication, clustering — anywhere "reads like" is the right relation, vectors are the right index, and none of that is what a graph does.

Structural recall. "Which services depend on the one that failed, and who owns them" has an exact answer that similarity can't compute — it requires following DEPENDS_ON and OWNS edges hop by hop. Multi-hop questions, variable-depth reachability, provenance chains, supersedence: these are traversals, deterministic and citable. The same query returns the same path, and the path is the explanation.

The retrieval stack that holds up in production uses each for its half. Vectors find entry points: embed the question, pull the semantically-relevant entities or passages. The graph expands and explains: from those entry points, traverse to the facts, dependencies and sources that similarity can't see. An agent's memory follows the same split — similarity for "have I seen something like this," traversal for "what do I actually know about it."

If your corpus is unstructured text and your questions are topical, start with a vector database alone — it may be all you need. If your questions span relationships — dependencies, ownership, identity, agent memory — you need a graph. If both, integrate them: the id of a vector hit becomes the entry node of a traversal.

CognoDB is deliberately the graph half. It ships Cypher traversals over Bolt, full-text BM25 keyword search, and an MCP server for agents — and no vector index. Pair it with the vector store you already run; the two compose over a shared entity id.