Dense Vector Search

Solr’s Dense Vector Search adds support for indexing and searching dense numerical vectors.

Deep learning can be used to produce a vector representation of both the query and the documents in a corpus of information.

These neural network-based techniques are usually referred to as neural search, an industry derivation from the academic field of Neural information Retrieval.

Important Concepts

Dense Vector Representation

A traditional tokenized inverted index can be considered to model text as a "sparse" vector, in which each term in the corpus corresponds to one vector dimension. In such a model, the number of dimensions is generally quite high (corresponding to the term dictionary cardinality), and the vector for any given document contains mostly zeros (hence it is sparse, as only a handful of terms that exist in the overall index will be present in any given document).

Dense vector representation contrasts with term-based sparse vector representation in that it distills approximate semantic meaning into a fixed (and limited) number of dimensions.

The number of dimensions in this approach is generally much lower than the sparse case, and the vector for any given document is dense, as most of its dimensions are populated by non-zero values.

In contrast to the sparse approach (for which tokenizers are used to generate sparse vectors directly from text input) the task of generating vectors must be handled in application logic external to Apache Solr.

There may be cases where it makes sense to directly search data that natively exists as a vector (e.g., scientific data); but in a text search context, it is likely that users will leverage deep learning models such as BERT to encode textual information as dense vectors, supplying the resulting vectors to Apache Solr explicitly at index and query time.

For additional information you can refer to this blog post.

Dense Retrieval

Given a dense vector v that models the information need, the easiest approach for providing dense vector retrieval would be to calculate the distance (euclidean, dot product, etc.) between v and each vector d that represents a document in the corpus of information.

This approach is quite expensive, so many approximate strategies are currently under active research.

The strategy implemented in Apache Lucene and used by Apache Solr is based on Navigable Small-world graph.

It provides efficient approximate nearest neighbor search for high dimensional vectors.

See Approximate nearest neighbor algorithm based on navigable small world graphs [2014] and Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs [2018] for details.

Index Time

This is the Apache Solr field type designed to support dense vector search:

DenseVectorField

The dense vector field gives the possibility of indexing and searching dense vectors of float elements.

For example:

[1.0, 2.5, 3.7, 4.1]

Here’s how DenseVectorField should be configured in the schema:

<fieldType name="knn_vector" class="solr.DenseVectorField" vectorDimension="4" similarityFunction="cosine"/>
<field name="vector" type="knn_vector" indexed="true" stored="true"/>
vectorDimension

Required

Default: none

The dimension of the dense vector to pass in.

Accepted values: Any integer.

similarityFunction

Optional

Default: euclidean

Vector similarity function; used in search to return top K most similar vectors to a target vector.

Accepted values: euclidean, dot_product or cosine.

this similarity is intended as an optimized way to perform cosine similarity. In order to use it, all vectors must be of unit length, including both document and query vectors. Using dot product with vectors that are not unit length can result in errors or poor search results.
the preferred way to perform cosine similarity is to normalize all vectors to unit length, and instead use DOT_PRODUCT. You should only use this function if you need to preserve the original vectors and cannot normalize them in advance.

To use the following advanced parameters that customise the codec format and the hyperparameter of the HNSW algorithm, make sure the Schema Codec Factory, is in use.

Here’s how DenseVectorField can be configured with the advanced hyperparameters:

<fieldType name="knn_vector" class="solr.DenseVectorField" vectorDimension="4" similarityFunction="cosine" knnAlgorithm="hnsw" hnswMaxConnections="10" hnswBeamWidth="40"/>
<field name="vector" type="knn_vector" indexed="true" stored="true"/>
knnAlgorithm

Optional

Default: hnsw

(advanced) Specifies the underlying knn algorithm to use

Accepted values: hnsw.

Please note that the knnAlgorithm accepted values may change in future releases.

vectorEncoding

Optional

Default: FLOAT32

(advanced) Specifies the underlying encoding of the dense vector elements. This affects memory/disk impact for both the indexed and stored fields (if enabled)

Accepted values: FLOAT32, BYTE.

hnswMaxConnections

Optional

Default: 16

(advanced) This parameter is specific for the hnsw knn algorithm:

Controls how many of the nearest neighbor candidates are connected to the new node.

It has the same meaning as M from the 2018 paper.

Accepted values: Any integer.

hnswBeamWidth

Optional

Default: 100

(advanced) This parameter is specific for the hnsw knn algorithm:

It is the number of nearest neighbor candidates to track while searching the graph for each newly inserted node.

It has the same meaning as efConstruction from the 2018 paper.

Accepted values: Any integer.

DenseVectorField supports the attributes: indexed, stored.

currently multivalue is not supported

Here’s how a DenseVectorField should be indexed:

JSON

[{ "id": "1",
"vector": [1.0, 2.5, 3.7, 4.1]
},
{ "id": "2",
"vector": [1.5, 5.5, 6.7, 65.1]
}
]

XML

<add>
<doc>
<field name="id">1</field>
<field name="vector">1.0</field>
<field name="vector">2.5</field>
<field name="vector">3.7</field>
<field name="vector">4.1</field>
</doc>
<doc>
<field name="id">2</field>
<field name="vector">1.5</field>
<field name="vector">5.5</field>
<field name="vector">6.7</field>
<field name="vector">65.1</field>
</doc>
</add>

SolrJ

final SolrClient client = getSolrClient();

final SolrInputDocument d1 = new SolrInputDocument();
d1.setField("id", "1");
d1.setField("vector", Arrays.asList(1.0f, 2.5f, 3.7f, 4.1f));


final SolrInputDocument d2 = new SolrInputDocument();
d2.setField("id", "2");
d2.setField("vector", Arrays.asList(1.5f, 5.5f, 6.7f, 65.1f));

client.add(Arrays.asList(d1, d2));

Query Time

This is the Apache Solr query approach designed to support dense vector search:

knn Query Parser

The knn k-nearest neighbors query parser allows to find the k-nearest documents to the target vector according to indexed dense vectors in the given field. The set of documents can be Pre-Filtered to reduce the number of vector distance calculations that must be computed, and ensure the best topK are returned.

The score for a retrieved document is the approximate distance to the target vector(defined by the similarityFunction configured at indexing time).

It takes the following parameters:

f

Required

Default: none

The DenseVectorField to search in.

topK

Optional

Default: 10

How many k-nearest results to return.

preFilter

Optional

Default: Depends on usage, see below.

Specifies an explicit list of Pre-Filter query strings to use.

includeTags

Optional

Default: none

Indicates that only fq filters with the specified tag should be considered for implicit Pre-Filtering. Must not be combined with preFilter.

excludeTags

Optional

Default: none

Indicates that fq filters with the specified tag should be excluded from consideration for implicit Pre-Filtering. Must not be combined with preFilter.

Here’s how to run a simple KNN search:

?q={!knn f=vector topK=10}[1.0, 2.0, 3.0, 4.0]

The search results retrieved are the k=10 nearest documents to the vector in input [1.0, 2.0, 3.0, 4.0], ranked by the similarityFunction configured at indexing time.

Explicit KNN Pre-Filtering

The knn query parser’s preFilter parameter can be specified to reduce the number of candidate documents evaluated for the k-nearest distance calculation:

?q={!knn f=vector topK=10 preFilter=inStock:true}[1.0, 2.0, 3.0, 4.0]

In the above example, only documents matching the Pre-Filter inStock:true will be candidates for consideration when evaluating the k-nearest search against the specified vector.

The preFilter parameter may be blank (ex: preFilter="") to indicate that no Pre-Filtering should be performed; or it may be multi-valued — either through repetition, or via duplicated Parameter References.

These two examples are equivalent:

?q={!knn f=vector topK=10 preFilter=category:AAA preFilter=inStock:true}[1.0, 2.0, 3.0, 4.0]
?q={!knn f=vector topK=10 preFilter=$knnPreFilter}[1.0, 2.0, 3.0, 4.0]
&knnPreFilter=category:AAA
&knnPreFilter=inStock:true

Implicit KNN Pre-Filtering

While the preFilter parameter may be explicitly specified on any usage of the knn query parser, the default Pre-Filtering behavior (when no preFilter parameter is specified) will vary based on how the knn query parser is used:

  • When used as the main q param: fq filters in the request (that are not Solr Post Filters) will be combined to form an implicit KNN Pre-Filter.

    • This default behavior optimizes the number of vector distance calculations considered, eliminating documents that would eventually be excluded by an fq filter anyway.

    • includeTags and excludeTags may be used to limit the set of fq filters used in the Pre-Filter.

  • When used as an fq param, or as a subquery clause in a larger query: No implicit Pre-Filter is used.

    • includeTags and excludeTags must not be used in these situations.

The example request below shows two usages of the knn query parser that will get no implicit Pre-Filtering from any of the fq parameters, because neither usage is as the main q param:

?q=(color_str:red OR {!knn f=color_vector topK=10 v="[1.0, 2.0, 3.0, 4.0]"})
&fq={!knn f=title_vector topK=10}[9.0, 8.0, 7.0, 6.0]
&fq=inStock:true

However, the next example shows a basic request where all fq parameters will be used as implicit Pre-Filters on the main knn query:

?q={!knn f=vector topK=10}[1.0, 2.0, 3.0, 4.0]
&fq=category:AAA
&fq=inStock:true

If we modify the above request to add tags to the fq parameters, we can specify an includeTags option on the knn parser to limit which fq filters are used for Pre-Filtering:

?q={!knn f=vector topK=10 includeTags=for_knn}[1.0, 2.0, 3.0, 4.0]
&fq=category:AAA
&fq={!tag=for_knn}inStock:true

In this example, only the inStock:true filter will be used for KNN Pre-Filtering to find the the topK=10 documents, and the category:AAA filter will be applied independently; possibly resulting in less then 10 total matches.

Some use cases where includeTags and/or excludeTags may be more useful then an explicit preFilter parameters:

  • You have some fq parameters that are re-used on many requests (even when you don’t use the knn parser) that you wish to be used as KNN Pre-Filters when you do use the knn query parser.

  • You typically want all fq params to be used as KNN Pre-Filters, but when users "drill down" on Facets, you want the fq parameters you add to be excluded from the KNN Pre-Filtering so that the result set gets smaller; instead of just computing a new topK set.

Usage as Re-Ranking Query

The knn query parser can be used to rerank first pass query results:

&q=id:(3 4 9 2)&rq={!rerank reRankQuery=$rqq reRankDocs=4 reRankWeight=1}&rqq={!knn f=vector topK=10}[1.0, 2.0, 3.0, 4.0]

When using knn in re-ranking pay attention to the topK parameter.

The second pass score(deriving from knn) is calculated only if the document d from the first pass is within the k-nearest neighbors(in the whole index) of the target vector to search.

This means the second pass knn is executed on the whole index anyway, which is a current limitation.

The final ranked list of results will have the first pass score(main query q) added to the second pass score(the approximated similarityFunction distance to the target vector to search) multiplied by a multiplicative factor(reRankWeight).

Details about using the ReRank Query Parser can be found in the Query Re-Ranking section.

Additional Resources