jvector-examples

JVector Examples

JVector comes with the following sample programs to try:

SiftSmall

A simple benchmark for the sift dataset located in the siftsmall directory in the project root.

mvn compile exec:exec@sift

Bench

Performs grid search across the GraphIndexBuilder parameter space to find the best tradeoffs between recall and throughput.

This benchmark requires datasets from https://github.com/erikbern/ann-benchmarks to be downloaded to hdf5 and fvec directories hdf5 or fvec under the project root depending on the dataset format.

You can use plot_output.py to graph the pareto-optimal points found by Bench.

mvn compile exec:exec@bench

Some sample KNN datasets for testing based on ada-002 embeddings generated on wikipedia data are available in ivec/fvec format for testing at:

aws s3 ls s3://astra-vector/wikipedia_squad/ --no-sign-request 
                           PRE 100k/
                           PRE 1M/
                           PRE 4M/

Bench automatically downloads the 100k dataset to the ./fvec directory .

To run SiftSmall/Bench without the JVM vector module available, you can use the following invocations:

mvn -Pjdk11 compile exec:exec@bench

mvn -Pjdk11 compile exec:exec@sift

IPCService

A simple service for adding / querying vectors over a unix socket.

Install socat using homebrew on mac or apt/rpm on linux

Mac:

brew install socat

Linux:

apt-get install socat

Start the service with:

mvn compile exec:exec@ipcserve

Now you can interact with the service

socat - unix-client:/tmp/jvector.sock

CREATE 3 DOT_PRODUCT 2 20
OK
WRITE [0.1,0.15,0.3]
OK
WRITE [0.2,0.83,0.05]
OK
WRITE [0.5,0.5,0.5]
OK
OPTIMIZE
OK
SEARCH 20 3 [0.15,0.1,0.1]
RESULT [2,1,0]

Commands

All commands are completed with \n.

No spaces are allowed inside vector brackets.

• CREATE {dimensions} {similarity-function} {M} {searchDepthConstruction}

  • Creates a new index for this session

• WRITE [N,N,N] ... [N,N,N]

  • Add one or more vectors to the index

• OPTIMIZE

  • Call when indexing is complete

• MEMORY

  • Get the in memory size of index

• SEARCH {overquerySearch} {top-k} [N,N,N] ... [N,N,N]

  • Search index for the top-k closest vectors (ordinals of indexed values returned per query)

• BULKLOAD {localpath}

  • Bulk loads a local file in numpy format Rows x Columns