G28 — Embedding-similarity and nearest-neighbor graph · Graph systems

169 words. Researched design, reviewed 2026-09-12: a library view or proposed design with its sources stated; not a deployed system, a live audit or a provider requirement. Source: Graph/G28-embedding-similarity-and-nearest-neighbor-graph.md. Graph systems, shelf 5 of 8; library release 2026.09.12-g40.

Status: NEW RESEARCHED DESIGN; not deployed
Reviewed: 2026-09-12
Nature: published algorithmic method
Priority: P2
Accountability: R10, R21

Question

Which vectors are approximately close under a specified representation and metric?

Nodes

Vector; Text/entity record; Model version; Neighbor relation

Relationship sketch

vector → source record
vector → approximate neighbor

Evidence inputs

Actual embedding model/version, dimension, similarity metric, indexed corpus, measured retrieval results.

Implementation

Keep an ANN index such as HNSW conceptually separate from the semantic fact graph. Record model changes and compare candidate retrieval against labeled relevant passages.

Acceptance checks

Embedding/version metadata is recorded
Approximate retrieval is evaluated on task examples
Similarity does not auto-merge entity identities

Do not infer

Nearest-neighbor proximity is not a factual predicate or proof that one statement supports another.

Publication boundary

Private search infrastructure unless explicitly published.

Source basis

S26, S24. See Research/sources.json. Sources support the primitives or named technology. The organization, labels, actions, and acceptance contracts in this catalog are this library's design, not claimed provider requirements.