Vector Search Reference

This page is the reference for pg_ripple's vector + SPARQL hybrid search.

Overview

pg_ripple integrates with pgvector to provide semantic similarity search over RDF entities. Entity embeddings are stored in _pg_ripple.embeddings with HNSW or IVFFlat indices. The pg:similar() SPARQL function queries the vector index and returns results ranked by cosine similarity.

Hybrid retrieval combines vector similarity ranking with SPARQL graph constraints, enabling queries like: "find entities semantically similar to X that also satisfy SHACL shape Y".

Status

SELECT feature_name, status FROM pg_ripple.feature_status()
WHERE feature_name LIKE '%vector%' OR feature_name LIKE '%embed%' OR feature_name LIKE '%kge%';

SQL Functions

SPARQL pg:similar() Function

Use the pg:similar() extension function inside SPARQL queries for inline vector search:

PREFIX pg: <http://pg_ripple.io/fn/>
SELECT ?entity ?score WHERE {
  ?entity pg:similar("machine learning", 10) ?score .
  ?entity a <http://example.org/Paper> .
}
ORDER BY DESC(?score)

Embedding Models

Embeddings are generated via the configured LLM embedding endpoint. Each entity-model pair is stored once in _pg_ripple.embeddings. The incremental embedding worker runs in the background and embeds new entities as they are inserted.

Knowledge Graph Embeddings (KGE)

Graph-structure embeddings (TransE, RotatE) are computed by src/kge.rs and stored alongside text embeddings. KGE embeddings capture structural relationship patterns and complement text-based semantic similarity.

Index Configuration

Related Pages

• Embedding Functions
• GraphRAG Reference
• Vector Index Trade-offs
• Feature Status Taxonomy