Prerequisites

• Install CUDA (12.6 or 12.8)
• Install CMake 3.30+
• apt install ninja-build libnccl-dev httpie curl
• Install JDK 22
• Install Maven 3.9.6+

Building cuVS 25.02

• Clone https://github.com/rapidsai/cuvs
• Checkout branch branch-25.02
• ./build.sh libcuvs
• ./build.sh java

(This should install the Java artifacts to the local Maven repository)

Building Lucene with cuVS

• Clone https://github.com/searchscale/lucene
• Checkout branch cuvs-integration-10x
• ./gradlew mavenToLocal

(This will build Lucene 10.2.0-SNAPSHOT and install artifacts to local Maven repository)

Building Solr with Lucene and cuVS

• Clone https://github.com/searchscale/solr
• Checkout branch ishan/cuvs-integration
• ./gradlew assemble distTar

(This will build Solr 10.0.0-SNAPSHOT and place artifacts in ./solr/packaging/build/distributions/solr-10.0.0-SNAPSHOT.tgz)

Building benchmarking project

• Clone https://github.com/searchscale/cuvs-bench
• Checkout branch noble/cuvs-panama
• mvn compile assembly:single

(This will create the artifact ./target/solr-cuvs-benchmarks-1.0-SNAPSHOT-jar-with-dependencies.jar)

Preparing Dataset

• Create a new work folder, say /home/ishan/workingarea.

• Copy the following into the workingarea:

  cp code/solr/solr/packaging/build/distributions/solr-10.0.0-SNAPSHOT.tgz workingarea/.

  cp -r code/solr/solr/example/cuvsexample workingarea/.

  cp -r code/cuvs-bench/upload_all.sh workingarea/.

  cp -r code/cuvs-bench/target/*jar workingarea/.

  cp -r code/cuvs-bench/start-solr10.sh workingarea/.

• Download the dataset:

  wget https://accounts.searchscale.com/datasets/wikipedia/wiki_dump_5Mx2048D.csv.gz

• Create batches of Javabin payloads

  mkdir batches

  java -cp solr-cuvs-benchmarks-1.0-SNAPSHOT-jar-with-dependencies.jar com.searchscale.benchmarks.Indexer data_file=wiki_dump_5Mx2048D.csv.gz output_file=batches/wiki batch_size=50000 docs_count=200000 legacy=true

(This will create 4 batches of 50k each, total of 200k documents, in batches folder)

Running Solr & Uploading batches

chmod +x *.sh
./start-solr10.sh
./upload_all.sh

Visit http://localhost:8983/ in browser, click "test" on left pane under "Collections", and issue a query "*:*". It should show numFound as 200k

Changing hyper parameters for HNSW and cuVS

To use HNSW, define the knn_vector field type as follows in cuvsexample/conf/schema.xml:

    <!--fieldType name="knn_vector" class="solr.DenseVectorField" vectorDimension="2048" knnAlgorithm="cuvs" similarityFunction="cosine" /-->
    <fieldType name="knn_vector" class="solr.DenseVectorField" vectorDimension="2048" knnAlgorithm="hnsw" hnswMaxConnections="256" hnswBeamWidth="512" similarityFunction="cosine" />

To use cuVS, define the knn_vector field type as follows in cuvsexample/conf/schema.xml:

    <fieldType name="knn_vector" class="solr.DenseVectorField" vectorDimension="2048" knnAlgorithm="cuvs" similarityFunction="cosine" />
    <!--fieldType name="knn_vector" class="solr.DenseVectorField" vectorDimension="2048" knnAlgorithm="hnsw" hnswMaxConnections="256" hnswBeamWidth="512" similarityFunction="cosine" /-->

Hyper parameters for HNSW is modified as above. Hyper parameters for cuVS is modified as follows in cuvsexample/conf/solrconfig.xml:

    <codecFactory name="CodecFactory" class="org.apache.solr.core.CuvsCodecFactory">
        <int name="cuvsWriterThreads">8</int> 
        <int name="graphDegree">32</int> 
        <int name="intGraphDegree">32</int> 
    </codecFactory>

knn Search

https://solr.apache.org/guide/solr/latest/query-guide/dense-vector-search.html#query-time