Vector indexing in Neptune Analytics

You can only create a vector search index for a Neptune Analytics graph at the time the graph is created. Neptune Analytics lets you create only one vector index for a graph, with a fixed dimension between 1 and 65,535 inclusive.

When you create a Neptune Analytics graph in the console, you specify the index dimension under Vector search settings near the end of the process.

Vector index transaction support

When using Neptune Analytics with a vector search index, it is important to understand that any updates performed on the vector index are not ACID compliant - specifically, any updates to the vector index are not atomic in nature. Atomicity in a database defines that when updates are performed, either all or none of them succeed. There are situations with the vector index where updating the embeddings may succeed, even when the remainder of the transaction fails:

To minimize the potential for this issue to occur, it is recommended that you either run a single query on a single vertex at a time, or if you are running concurrent queries, that the set of vertices being updated are distinct.

Loading vectors into a Neptune Analytics graph vector index

Note that the nodes in your graph must have at least one user property or label in order to associate them with embeddings. Also, Neptune Analytics does not support the special positive and negative infinity (INF, -INF) and not-a-number (NaN) floating-point values.

Neptune Analytics supports optional embeddings in the CSV file when the vector index is enabled. This means that not every node needs to be associated with an embedding.

Neptune Analytics does not currently support loading vectors from Neptune Database or a snapshot.

There are two ways you can load vectors associated with nodes in your graph:

Load the vectors from graph data files Amazon S3

When you're loading graph data from files in Amazon S3 using the console or the neptune.load{} openCypher integration, you can add a column to your CSV data with an embedding:vector header. This column should contain a list of integer or floating-point values separated by semicolons ( ; ) that forms a vector of the required dimension and is the embedding for the node in question.

For example, associating a 4-dimensional vector with nodes in your graph in the openCypher CSV format would look like this:

:ID, name:String, embedding:Vector, :LABEL
v1,"ABC",0.1;0.5;0.8;-1.32,person
v2,"DEF",8.1;-0.2;0.432;-1.02,person
v3,"GHI",12323343;24324;2433554;-4343434,person
v4,"JKL",121.12213;3223.212;265;-1.32,person

In the Gremlin CSV format, the same thing would look like this:

~id, name, embedding:vector, ~label
v1,"ABC",0.1;0.5;0.8;-1.32,person
v2,"DEF",8.1;-0.2;0.432;-1.02,person
v3,"GHI",12323343;24324;2433554;-4343434,person
v4,"JKL",121.12213;3223.212;265;-1.32,person

Using the vectors.upsert algorithm to load vectors for your graph

You can also use the vectors.upsert algorithm to insert or update embeddings in a Neptune Analytics graph that has a vector search index. For example, in openCypher you can call the algorithm like this:

CALL neptune.algo.vectors.upsert(
  "person933",
  [0.1, 0.2, 0.3, ..]
)
YIELD node, embedding, success
RETURN node, embedding, success

Another example is:

UNWIND [
  {id: "933", embedding: [1,2,3,4]},
  {id: "934", embedding: [-1,-2,-3,-4]}
] as entry
MATCH (n:person) WHERE id(n)=entry.id WITH n, entry.embedding as embedding
CALL neptune.algo.vectors.upsert(n, embedding)
YIELD success
RETURN n, embedding, success

Common errors you may encounter when loading embeddings

Vector-search algorithms in Neptune Analytics

Neptune Analytics supports a variety of vector-search algorithms that are listed in the VSS algorithms section.