SPARQL & Vector Extensions

Standard SPARQL 1.1 Support

SutraDB implements the core of SPARQL 1.1 Query Language:

Vector Extensions

SutraDB adds two new operators to SPARQL for approximate nearest neighbor (ANN) search over vector embeddings. These operators integrate directly into graph patterns — you can traverse the graph, jump into vector space, and come back in a single query.

VECTOR_SIMILAR

Finds entities whose vector embeddings are similar to a query vector, above a cosine similarity threshold.

-- Basic usage: find entities similar to a query vector
SELECT ?entity WHERE {
  VECTOR_SIMILAR(?entity :hasEmbedding
    "0.23 -0.11 0.87 ..."^^sutra:f32vec, 0.85)
}

-- With ef_search hint (higher = better recall, slower)
VECTOR_SIMILAR(?entity :hasEmbedding
    "..."^^sutra:f32vec, 0.85, ef:=200)

-- With top-K mode (return K nearest, ignore threshold)
VECTOR_SIMILAR(?entity :hasEmbedding
    "..."^^sutra:f32vec, 0.0, k:=10)

VECTOR_SCORE

Returns the similarity score for ranking in ORDER BY clauses:

SELECT ?paper WHERE {
  ?paper a :Paper .
  VECTOR_SIMILAR(?paper :hasEmbedding "..."^^sutra:f32vec, 0.7)
} ORDER BY DESC(VECTOR_SCORE(?paper :hasEmbedding "..."^^sutra:f32vec))
LIMIT 10

Wormhole Queries

"Wormhole" queries traverse between graph space and vector space in a single query. This is what makes SutraDB unique — no other triplestore can express these patterns.

Graph → Vector → Graph

Start with a graph pattern, jump into vector space to find similar entities, then traverse back into the graph:

-- Find papers by a specific author, then find papers with similar
-- embeddings, then get their titles
SELECT ?similar_paper ?title WHERE {
  -- Graph: find papers by Vaswani
  ?original ex:author ex:Vaswani .

  -- Vector: find papers with similar embeddings
  VECTOR_SIMILAR(?similar_paper ex:hasEmbedding
    "..."^^sutra:f32vec, 0.8)

  -- Graph: get the title of the similar paper
  ?similar_paper ex:title ?title .
}

Vector → Graph

Start in vector space, then use graph patterns to filter or enrich:

-- Find entities similar to a query, then check what type they are
SELECT ?entity ?type WHERE {
  VECTOR_SIMILAR(?entity ex:hasEmbedding "..."^^sutra:f32vec, 0.9)
  ?entity a ?type .
}

How Vectors Live in the Graph

Vectors are not a separate system — they are primitive objects in the graph, like strings or integers. An entity connects to its vector via an ordinary triple:

-- A vector is the object of a triple
:paper42 :hasEmbedding "0.23 -0.11 0.87 ..."^^sutra:f32vec .

-- Multiple entities can share the same vector (disambiguation)
:bank_financial :hasEmbedding "0.5 0.3 ..."^^sutra:f32vec .
:bank_river     :hasEmbedding "0.5 0.3 ..."^^sutra:f32vec .

When you run VECTOR_SIMILAR, it searches the HNSW index for matching vector objects, then follows the triples backward to find which entities connect to those vectors. This means:

• A vector never exists without at least one triple pointing to it
• Triples always outnumber or equal vectors in a database
• Vector search inherently does entity resolution — if two entities share a vector, both are returned

Query Planner

The query planner automatically decides execution order based on pattern selectivity:

SPARQL Protocol

SutraDB implements the SPARQL 1.1 Protocol over HTTP: