The Operators

Loka extends SPARQL 1.1 with two operators that integrate vector similarity search directly into graph query patterns:

VECTOR_SIMILAR

Filters results to those with similarity above a threshold:

# Find documents similar to a query vector that discuss AI entities
SELECT ?doc ?entity WHERE {
  ?entity rdf:type :Person .
  ?doc :mentions ?entity .
  VECTOR_SIMILAR(?doc :hasEmbedding "0.23 -0.11 ..."^^loka:f32vec, 0.85)
}

The key: VECTOR_SIMILAR participates in the same query plan as regular triple patterns. The planner decides whether to run the vector search first or the graph pattern first based on cost.

VECTOR_SCORE

Returns the similarity score for use in ORDER BY and projections:

SELECT ?doc ?score WHERE {
  ?doc :hasEmbedding ?emb .
  BIND(VECTOR_SCORE(?doc :hasEmbedding "0.23 -0.11 ..."^^loka:f32vec) AS ?score)
}
ORDER BY DESC(?score)
LIMIT 10

Query Hints

Both operators accept optional hints for fine-tuning HNSW search behavior:

# Explicit ef_search for higher recall
VECTOR_SIMILAR(?doc :hasEmbedding "..."^^loka:f32vec, 0.85, ef:=200)

# Explicit k limit
VECTOR_SIMILAR(?doc :hasEmbedding "..."^^loka:f32vec, 0.80, k:=50)

How It Integrates with Graph Patterns

The query planner treats VECTOR_SIMILAR as another pattern with a cost estimate. The planning rules:

Situation	Strategy
Subject unbound before VECTOR_SIMILAR	Run vector search first (returns top-k candidates), then evaluate graph patterns over candidates
Subject bound before VECTOR_SIMILAR	Run graph pattern first (binds subjects), then filter by vector similarity

This means a query like "find papers similar to X that are written by authors from Tokyo" will automatically choose the most efficient execution order based on the data distribution.

Why Not a Separate API?

A separate vector search API (like Qdrant's REST API) requires the application to:

1. Call the vector API to get candidate IDs
2. Parse the JSON response
3. Feed those IDs into a SPARQL query
4. Execute the SPARQL query

With VECTOR_SIMILAR in SPARQL, it is one query, one round trip, one result set. The planner handles the interleaving internally.