Vector Search

TigerGraph functions as a multi-modal database, allowing you to store and index high-dimensional vectors alongside your graph data. This is a critical feature for AI applications like Retrieval-Augmented Generation (RAG) and recommendation systems.

1. Schema Configuration

Vector attributes are added to existing vertex types via a schema change job.

gsqlterminal
Copy
CREATE GLOBAL SCHEMA_CHANGE JOB add_vector {
  ALTER VERTEX User ADD VECTOR ATTRIBUTE embedding(
    DIMENSION=1536, 
    METRIC="COSINE", 
    INDEXTYPE="HNSW"
  );
}
RUN GLOBAL SCHEMA_CHANGE JOB add_vector

Key Parameters

• DIMENSION: Vector length (e.g., 1536 for OpenAI embeddings).
• METRIC: Similarity measure ("COSINE", "L2", or "IP").
• INDEXTYPE: The algorithm used for indexing ("HNSW" is the default).

2. Loading Vectors

You can load vectors in bulk using a standard loading job with the SPLIT function.

gsqlterminal
Copy
LOAD file_object TO VECTOR ATTRIBUTE embedding ON VERTEX User 
VALUES ($0, SPLIT($1, ",")) USING SEPARATOR="|";

3. Querying with vectorSearch()

The vectorSearch function is used within a GSQL query to find the top-K similar vertices.

gsqlterminal
Copy
CREATE QUERY find_similar_users (LIST<FLOAT> query_vec) SYNTAX v3 {
  MapAccum<Vertex, Float> @@distances;
  
  // Find top 10 users similar to the input vector
  result = vectorSearch({User.embedding}, query_vec, 10, {distance_map: @@distances});
  
  PRINT result;
}

4. Hybrid Graph+Vector Search

Because vector search is just another function in GSQL, you can combine it with graph traversals:

1. Perform a Vector Search to find a starting set of vertices.
2. Traverse the Graph from those vertices to find related entities or calculate PageRank.

[!IMPORTANT] Indexing Lag: The HNSW index is updated incrementally. New vectors might not be searchable until the indexing process (rebuild) is complete. Monitor status via GET /restpp/vector/status.