Concept
Graph retrieval
Representing relationships explicitly instead of hoping a similarity score finds them.
Vector search answers 'what looks like this'. A graph answers 'what is connected to this, and through what'. Most questions that matter are the second kind, and the two retrieval modes fail differently, which is why combining them is more useful than choosing.
The practical work is in the edges: what counts as a relationship, how it is typed, and what happens when two edges contradict. A graph that cannot represent contradiction will average it away.
BM25 for lexical precision, cosine for semantic reach, and a graph for structure — each covering the others' blind spot.
Two consequences follow. Judgment is usually implicit — unrepeatable, uninspectable, unteachable — and making a decision an explicit graph turns it into something you can audit, hand down and improve. And a system that forgets between sessions cannot accumulate anything: durable, structured memory is what turns repeated information into compounding knowledge rather than a series of first conversations.
Implementation
Where this runs
A concept that never left a document is a hypothesis. These are the systems that implement it.