VADD

is the name of the key that will hold the vector set data.

either a 32-bit floating point (FP32) blob of values or num floating point numbers as strings.

is the name of the element that is being added to the vector set.

implements random projection to reduce the dimensionality of the vector. The projection matrix is saved and reloaded along with the vector set. Please note that the REDUCE option must be passed immediately before the vector. For example,

VADD mykey REDUCE 50 VALUES ...

performs the operation partially using threads, in a check-and-set style. The neighbor candidates collection, which is slow, is performed in the background, while the command is executed in the main thread.

in the first VADD call for a given key, NOQUANT forces the vector to be created without int8 quantization, which is otherwise the default.

forces the vector to use binary quantization instead of int8. This is much faster and uses less memory, but impacts the recall quality.

forces the vector to use signed 8-bit quantization. This is the default, and the option only exists to make sure to check at insertion time that the vector set is of the same format.

plays a role in the effort made to find good candidates when connecting the new node to the existing Hierarchical Navigable Small World (HNSW) graph. The default is 200. Using a larger value may help in achieving a better recall. To improve the recall it is also possible to increase EF during VSIM searches.

associates attributes in the form of a JavaScript object to the newly created entry or updates the attributes (if they already exist). It is the same as calling the VSETATTR command separately.

is the maximum number of connections that each node of the graph will have with other nodes. The default is 16. More connections means more memory, but provides for more efficient graph exploration. Nodes at layer zero (every node exists at least at layer zero) have M * 2 connections, while the other layers only have M connections. For example, setting M to 64 will use at least 1024 bytes of memory for layer zero. That's M * 2 connections times 8 bytes (pointers), or 128 * 8 = 1024. For higher layers, consider the following:

• Each node appears in ~1.33 layers on average (empirical observation from HNSW papers), which works out to be 0.33 higher layers per node.
• Each of those higher layers has M = 64 connections.

So, the additional amount of memory is approximately 0.33 × 64 × 8 ≈ 169.6 bytes per node, bringing the total memory to ~1193 bytes.

If you don't have a recall quality problem, the default is acceptable, and uses a minimal amount of memory.