Semantic search plugin

Placement

The JAR file generated by gradle has to be placed under {SOLR_HOME}/{collection}/lib to be loaded by Solr.

Update Request Processor

About

A custom Solr UpdateRequestProcessor to get the embeddings of a specified field and store them in a valid knn_vector field.

URP Configuration

The URP is configured in the collection's solrconfig as follows:

<updateRequestProcessorChain name="get-embeddings">
    <processor class="com.rondhuit.solr.update.dense.EmbeddingsProcessorFactory">
        <str name="hostName">host.docker.internal</str>
        <int name="portNo">8080</int>
        <str name="modelName">sonoisa</str>
        <str name="sourceField">title</str>
        <str name="targetField">title_vector</str>
    </processor>

    <processor class="solr.LogUpdateProcessorFactory"/>
    <processor class="solr.RunUpdateProcessorFactory"/>
</updateRequestProcessorChain>

• hostName and portNo can be omitted. Their default values are localhost and 8080 respectively.
• modelName has to match a model deployed in TorchServe.
• sourceField is the text field for which we want to calculate embeddings (to be vectorized).
• targetField is a field of type knn_vector to hold the calculated embeddings.

Search component

About

A custom Solr SearchComponent to get the embeddings of a query and perform a Dense vector search.

SC Configuration

The Search component is configured in the collection's solrconfig.xml as follows:

• Component definition:

    <searchComponent name="dense-search"
                     class="com.rondhuit.solr.handler.component.DenseVectorSearchComponent">
      <str name="hostName">host.docker.internal</str>
      <int name="portNo">8080</int>
      <str name="modelName">sonoisa</str>
    </searchComponent>
  

  • hostName and portNo can be omitted. Their default values are localhost and 8080 respectively.
  • modelName has to match a model deployed in TorchServe.

• Search Handler definition:

  DenseVectorSearchComponent should be placed before QueryComponent. To do that, we define dense-search in the first-components of the SearchHandler.

    <requestHandler name="/select" class="solr.SearchHandler">
      <lst name="defaults">
        <str name="echoParams">explicit</str>
        <int name="rows">10</int>
      </lst>
      <arr name="first-components">
        <str>dense-search</str>
      </arr>
    </requestHandler>
  

Query parameters

Query parameters to perform a KS Dense Vector Search:

Standard query request examples

http://localhost:8983/solr/covid/select?q=covid%20incubation%20period&_ks.dense=true&_ks.dense.field=answer_embeddings

http://localhost:8983/solr/covid/select?q=covid%20incubation%20period&_ks.dense=true&_ks.dense.field=answer_embeddings&fl=title%2C%20answer

Json Request API examples

curl -X POST "http://localhost:8983/solr/covid/select" -d '
{
  "query" : "covid incubation time",
  "fields": "title,answer",
  "params": {
    "_ks.dense": "true",
    "_ks.dense.model": "sonoisa",
    "_ks.dense.field": "answer_embeddings",
    "_ks.dense.k": 10
  }
}'