• [+] expand all

Searching and ranking of multi-valued fields

This guide explores how to search and rank over structured multi-valued fields. The examples in this guide uses the weightedset field type. The generic map<key-type,value-type> field type does not currently support ranking and can only be used for matching and filtering.

Introduction

When building a search application we need to think about:

Matching

There is a lot of text matching options we should think about when designing and mapping our document model to a Vespa document schema:

  • For string fields we should think about using text style matching or database-style exact matching.
  • For string fields there are also several linguistic processing options like tokenization, normalization and language dependent stemming.
  • String fields which shares the same match and linguistic processing settings can be combined using fieldsets.

At query time, we can take the user query and translate it into a valid Vespa query request which implements our matching and retrieval strategy over the designed document schema.

Ranking

The documents which match the query and are retrieved by the query are scored using a ranking model. Once a document is retrieved by the query logic the document can be scored using the full flexibility of the Vespa ranking framework.

Exploration

In the following sections we explore matching and ranking over multi-valued string fields.

A minimal Vespa application

Assuming we have the following sample data document where we have a structured tag-like field where there is a weight associated with each element.

{
    "put": "id:photos:photo::0",
    "fields": {
        "title": "Mira in the sunset",
        "description": "A sunny afternoon with our dogs",
        "tags": {
            "no filter":1,
            "light": 3,
            "black and white": 3,
            "clear sky": 2,
            "sunset dogs": 4
        }
    }
}

Structured data like the tags, where we both want to match and rank is best represented using the weightedset field type. The Vespa weightedset field type can be used to represent:

How should we design our Vespa schema, and how should we match and search this data model for end-user free text queries?

  • We want to use text matching when searching the title and description
  • We also want to match the free form tags field as these tags might increase recall and the weight of the matched element(s) could influence ranking of documents matched. We can start with the following schema:
schema photo {

    stemming: none
  
    document photo {

        field title type string {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

        field description type string {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

        field interestingness type float {
            indexing: summary | attribute
        }

        field tags type weightedset<string> {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

    }

    fieldset default {
        fields: title, description, tags
    }

    rank-profile default {
        first-phase {
            expression: nativeRank
        }
    }
}

In the schema we disable stemming and also enable bm25 text ranking feature for all string fields.

Since all string fields shares the same match settings we can use a fieldset so that queries does not need to mention all three fields.

We also include a default rank profile (this is the implicit default rank profile) using the Vespa nativeRank text matching rank feature.

Along with the schema, we also need a services.xml file to make up a Vespa application package:

<?xml version="1.0" encoding="UTF-8"?>
<services version="1.0">

    <container id="default" version="1.0">
        <search />
        <document-api />
        <nodes>
            <node hostalias="node1"></node>
        </nodes>
    </container>

    <content id="photos" version="1.0">
        <redundancy>1</redundancy>
        <documents>
            <document type="photo" mode="index"/>
        </documents>
        <nodes>
            <node hostalias="node1" distribution-key="0" />
        </nodes>
    </content>

</services>

Starting Vespa

This example uses the vespa container image:

$ docker pull vespaengine/vespa
$ docker run --detach --name vespa --hostname vespa-container \
  --publish 8080:8080 --publish 19071:19071 \
  vespaengine/vespa

Install Vespa-cli using Homebrew:

$ brew install vespa-cli

Now we can deploy the application using vespa-cli:

$ vespa deploy --wait 300 my-app

Feeding to Vespa

Feed the sample document

$ vespa document -v doc.json 

Query our data

Assuming a free text query sunset photos featuring dogs, we translate the user query into a Vespa query request using YQL:

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=sunset photos featuring dogs' 'type=all'

The above query returns 0 hits, since the query requires that all query terms matches the document. By adding trace.level to the query request we can see how the query is parsed and executed against the content nodes:

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=sunset photos featuring dogs' 'type=all' 'trace.level=3'

In the trace we can see the query which is dispatched to the content nodes: query=[AND sunshot photos featuring dogs]

Using tracing is very useful when debugging why documents match or does not match.

Since the sample document does not contain the term featuring or photos, the query fails to retrieve the example document. Relax the query matching to instead of requiring that all terms match, to use any. See model.type query api reference for supported query types:

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=sunset photos featuring dogs' 'type=any'

Changing the type to any, recalls the sample document as we no longer require that all query terms must match. With type it also possible to require that individual query terms match by using +:

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=+sunset photos featuring +dogs' 'type=any'

In this example sunset and dogs must be matched. Note that we have disabled stemming so querying for dogs won’t recall documents with dog. This is one of the reasons we disabled stemming, to demonstrate that stemming has impact on recall. Requiring dog will cause the query to not recall our sample document.

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=+sunset photos featuring +dog' 'type=any'

Now, let us explore how Vespa matches the multi-valued tags field of type weightedset. Notice that we change back to type=all. In this example we also use the default-index query parameter to limit matching to the tags field.

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=clear sky' 'type=all' 'default-index=tags'

The query matches the document which is no surprise since a tag contains the exact content clear sky. Let us search for just clear instead:

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=clear' 'type=all' 'default-index=tags'

Also matches the document, this demonstrates that matching is partial, it does not require to match the set element exactly. clear matches clear sky and sky will match clear sky.

But what about black sky:

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=black sky' 'type=all' 'default-index=tags'

Also matches the document. This is an example of cross-element matching. With weightedset using indexing:index with match:text multi term queries match across elements.

This might be a good decision, as we increase recall, however in some cases we want to differentiate an exact match from a partial match during ranking, so that exact matches are ranked higher than partial matches.

Ranking

We have now explored querying and matching, now it’s time to focus on how to rank the documents matched. You might not have noticed, but in the above examples, each of the queries produced a relevance score per hit, this score was in our previous examples calculated using the default rank profile which in our case used nativeRank. We can start by analyzing other rank features by asking Vespa to produce them for us. We use match-features to return rank features with the retrieved documents. We explicitly mention which ranking features we want to have calculated and returned. Notice that we don’t change the actual scoring, we still use nativeRank as the scoring function.

schema photo {

    stemming: none

    document photo {

        field title type string {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

        field description type string {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

        field interestingness type float {
            indexing: summary | attribute
        }

        field tags type weightedset<string> {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

    }

    fieldset default {
        fields: title, description, tags
    }

    rank-profile default {
        first-phase {
            expression: nativeRank
        }

        match-features {
            bm25(title)
            bm25(description)
            bm25(tags)

            nativeRank
            nativeRank(title)
            nativeRank(description)

            elementSimilarity(tags)
            elementCompleteness(tags).elementWeight
            elementCompleteness(tags).fieldCompleteness
            elementCompleteness(tags).queryCompleteness
            elementCompleteness(tags).completeness
        }
    }
}

Re-deploy with the changed rank profile:

$ vespa deploy --wait 300 my-app

Now we will see a list of features in the response:

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=clear sky' 'type=any'

The output includes matchfeatures where we can see the various scores for the features:

Especially look at the elementCompleteness and elementSimilarity rank features which are example of features for indexed multivalued string fields.

We can also notice that elementCompleteness(tags).fieldCompleteness is 1.0 which means that the tag was matched exactly and the "elementCompleteness(tags).elementWeight outputs the weight of the best matched element.

The elementSimilarity(tags) ranking feature is very flexible and even allow us to override the calculation and output new features.

In this example we defined two new ranking feature

  • elementSimilarity(tags).sumWeight which uses the sum of matching elements using field completeness x weight.
  • elementSimilarity(tags).maxWeight which uses the max over the matching elements using field completeness x weight.
schema photo {

    stemming: none

    document photo {

        field title type string {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

        field description type string {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

        field interestingness type float {
            indexing: summary | attribute
        }

        field tags type weightedset<string> {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

    }

    fieldset default {
        fields: title, description, tags
    }

    rank-profile default {
        rank-properties {
            elementSimilarity(tags).output.sumWeight: "sum(f*w)"
            elementSimilarity(tags).output.maxWeight: "max(f*w)"
        }
   
        first-phase {
            expression: nativeRank
        }

        match-features {
            bm25(title)
            bm25(description)
            bm25(tags)

            nativeRank
            nativeRank(title)
            nativeRank(description)

            elementSimilarity(tags)
            elementSimilarity(tags).sumWeight
            elementSimilarity(tags).maxWeight

            elementCompleteness(tags).elementWeight
            elementCompleteness(tags).fieldCompleteness
            elementCompleteness(tags).queryCompleteness
            elementCompleteness(tags).completeness
        }
    }
}

Re-deploy with the changed rank profile:

$ vespa deploy --wait 300 my-app

Now we will see a list of features in the response:

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=clear sky' 'type=any'

Each hit returned contains a matchfeatures field where we can see the various scores for the features:

Now, we can include these features in a ranking expression used in first-phase to actually change the ranking. The actual best scoring function is data dependent. A trained function using machine learning is by far the easiest way. The bag of words bm25 ranking feature is not normalized, so combining it in a linear function is challenging, as the score range of the feature is unbound. To overcome this, and allow easy exploration without changing the rank profile, make the parameters in the function overridable on a per-query basis by:

first-phase {
    expression {
        query(titleWeight)*bm25(title) +
        query(descriptionWeight)*bm25(description) +
        query(tagWeight)*elementSimilarity(tags).maxWeight
    }
}

See using query variables.

schema photo {

    stemming: none

    document photo {

        field title type string {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

        field description type string {
            indexing: summary | index
            match:text
            index: enable-bm25
        }

        field interestingness type float {
            indexing: summary | attribute
        }

        field tags type weightedset<string> {
            indexing: summary | index
            match:text
            index: enable-bm25
        }
    }

    fieldset default {
        fields: title, description, tags
    }

    rank-profile tunable inherits default {
        inputs {
            query(titleWeight): 2
            query(descriptionWeight): 1
            query(tagWeight): 2
        }

        rank-properties {
            elementSimilarity(tags).output.sumWeight: "sum(f*w)"
            elementSimilarity(tags).output.maxWeight: "max(f*w)"
        }
   
        first-phase {
            expression {
                query(titleWeight)*bm25(title) + query(descriptionWeight)*bm25(description) +
                query(tagWeight)*elementSimilarity(tags).maxWeight
            }
        }

        match-features {
            bm25(title)
            bm25(description)
            bm25(tags)
            elementSimilarity(tags).maxWeight
            firstPhase
        }
    }
}

Re-deploy

$ vespa deploy --wait 300 my-app

Run a query with the new rank profile

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=clear sky' 'type=any' 'ranking=tunable'

With the function above, since ‘clear sky’ does not match any of the title or description fields, the bm25 features becomes zero.

Our elementSimilarity(tags).maxWeight feature is 2.0 and the first phase expression becomes 4 which is reflected in the hit relevance score.

Change the query(tagWeight) with the query request and observe that the relevance becomes 6.0

$ vespa query 'yql=select * from photos where userQuery()' \
  'query=clear sky' 'type=any' 'ranking=tunable' \
  'input.query(tagWeight)=3'

Similar, we could also include a document-only signal to our ranking function by:

inputs {
    query(titleWeight): 2
    query(descriptionWeight): 1
    query(tagWeight): 2
    query(staticWeight): 1
}
rank-properties {
    elementSimilarity(tags).output.sumWeight: "sum(f*w)"
    elementSimilarity(tags).output.maxWeight: "max(f*w)"
}
first-phase {
    expression {
        query(titleWeight)*bm25(title) + query(descriptionWeight)*bm25(description) +
        query(tagWeight)*elementSimilarity(tags).maxWeight + query(staticWeight)*attribute(interestingness)
    }
}

That concludes our matching and ranking experiments. To shut down the container run:

$ docker rm -f vespa

Conclusion

In this guide we have looked at how one can build a query retrieval strategy and how to change ranking when searching multi-valued fields using the weightedset field type.