• [+] expand all

Phased Ranking

Ranking, in general, becomes more accurate by using complex expressions which use many features.

Read the query API guide to get an overview of how queries are executed in Vespa, before continuing this guide.

Vespa supports multiple rank phases:

  • Re-ranking of the top hits from the content nodes using a custom stateless searchers. See concrete example in reranking in searcher.
  • Re-ranking at each content node using two phase ranking configured by rank profiles with ranking expressions working on scalars, tensors or matching features
  • Retrieval and Ranking with query operators like WAND/Weak And which speeds up retrieval of top matches of a disjunction query (OR) or Nearest Neighbor Search
Ranking in 3 phases

Stateless re-ranking

It's possible to write stateless searchers which can re-rank hits using any custom scoring function or model. The searcher can also blend and re-order hits from multiple sources when using federation of content sources. See ReRankingSearcher for inspiration. The linked searcher re-ranks the top global results from all content nodes using a Transformer (BERT) model.

The searcher might request rank features calculated by the content nodes to be returned along with the hit fields using summary-features. The features returned can be configured in the rank-profile as summary-features.

The number of hits is limited by the query api hits parameter and maxHits setting. The hits available for container level re-ranking are the global top ranking hits after content nodes have retrieved and ranked the hits and global top ranking hits have been found by merging the responses from the content nodes.

Two-phase ranking content nodes

Rank expressions are configured in the rank-profile section of schema. The rank-profile supports having two phases of ranking; first-phase and second-phase:

By default, second-phase ranking (if specified) is run on the 100 best hits from the first-phase ranking per content node, after matching and before information is returned to the container. The number of hits to rerank can be configured as well in the rank-profile. Example:

schema myapp {
    …
    rank-profile title-freshness inherits default {
        first-phase {
            expression {
                nativeRank(title) + freshness(timestamp)
            }
        }
        second-phase {
            expression {
                xgboost("my-model.json")
            }
            rerank-count: 50
        }
    }
}

In this example, the first phase uses the text matching feature nativeRank scoped to the title field plus one of the built-in document ranking features named freshness over a timestamp field which stores the epoch time in second resolution. For each content node, the top 50 hits from the first phase function is re-ranked using a trained xgboost model.

Top-K Query Operators

If the first-phase ranking function can be approximated as a simple linear function, and the query mode is any, the Weak And/WAND implementations in Vespa allows avoiding fully evaluating all the documents matching the query with the first-phase function. Instead, only the top K hits using the internal operator ranking will be exposed to the first-phase ranking function.

Both wand implementations accepts a targeted number of hits (top N). Each content node will expose all hits that were evaluated to the first-phase ranking function, while skipped documents will not. The skipped set are the ones that matches the query, but which cannot outperform any of the already collected hits on the internal scoring heap. Due to not fully evaluating all documents matching the query, the total hit count becomes inaccurate when using weakAnd/Wand.

The nearest neighbor search operator is also a top k retrieval operator which does ranking (by distance) and retrieval in one process and exposes only a small subset of the matching documents to the first-phase ranking function.

Choosing phased ranking functions

A good ranking expression will for most applications consume too much CPU to be runnable on the entire result set within the latency budget/SLA. The application ranking function should hence in most cases be a second phase function. The task then becomes to find a first phase function, which correlates sufficiently well with the second phase function - this to ensure that relevance is not hurt too much by not evaluating the real ranking function on all the hits.

Rank phase statistics

Use match-features and summary-features to export detailed match- and rank-information per query. This requires post-processing and aggregation is an external system for analysis.

To evaluate how well the document corpus matches the queries, use mutable attributes to track how often each document survives each match- and ranking-phase. This is aggregated per document and makes it easy to analyse using the query and grouping APIs in Vespa - and no other processing/storage is required.

A mutable attribute is a number where an operation can be executed in 4 phases:

  1. on-match
  2. on-first-phase
  3. on-second-phase
  4. on-summary

The common use case is to increase the value by 1 for each execution. With this, it is easy to evaluate the document's performance to the queries, e.g. find the documents that appear in most queries, or the ones that never matched - run a query and order by the mutable attribute.

This example is based on the quickstart. It uses 4 attributes that each track how many times a document participates in any of the 4 phases. This is tracked only if using rank-profile rank_albums_track in the query:

schema music {

    document music {

        field artist type string {
            indexing: summary | index
        }

        field album type string {
            indexing: summary | index
        }

        field year type int {
            indexing: summary | attribute
        }

        field category_scores type tensor<float>(cat{}) {
            indexing: summary | attribute
        }

    }

    field match_count type long {
        indexing: attribute | summary
        attribute: mutable
    }
    field first_phase_count type long {
        indexing: attribute | summary
        attribute: mutable
    }
    field second_phase_count type long {
        indexing: attribute | summary
        attribute: mutable
    }
    field summary_count type long {
        indexing: attribute | summary
        attribute: mutable
    }

    fieldset default {
        fields: artist, album
    }

    rank-profile rank_albums inherits default {
        first-phase {
            expression: sum(query(user_profile) * attribute(category_scores))
        }
        second-phase {
            expression: attribute(year)
            rerank-count: 1
        }
    }

    rank-profile rank_albums_track inherits rank_albums {
        mutate {
            on-match        { match_count        += 1 }
            on-first-phase  { first_phase_count  += 1 }
            on-second-phase { second_phase_count += 1 }
            on-summary      { summary_count      += 1 }
        }
    }

    rank-profile rank_albums_reset_on_match inherits rank_albums {
        mutate {
            on-match        { match_count         = 0 }
        }
    }
    rank-profile rank_albums_reset_on_first_phase inherits rank_albums {
        mutate {
            on-match        { first_phase_count   = 0 }
        }
    }
    rank-profile rank_albums_reset_on_second_phase inherits rank_albums {
        mutate {
            on-match        { second_phase_count  = 0 }
        }
    }
    rank-profile rank_albums_reset_on_summary inherits rank_albums {
        mutate {
            on-match        { summary_count       = 0 }
        }
    }
}
$ vespa query \
  "select * from music where album contains 'head'" \
  "ranking=rank_albums_track"

Usage

The framework is flexible in use, the normal use case is:

  1. Reset the mutable attribute on all content nodes - use searchPath to make sure all nodes are reset by sending a query using a rank profile that resets the value. For each phase, run a query that matches all documents, and reset the attribute - e.g.:
    $ for phase in match first_phase second_phase summary; do \
          for node in {0..3}; do vespa query \
              "select * from music where true" \
              "ranking=rank_albums_reset_on_$phase" \
              "model.searchPath=$node/0"; \
          done \
      done
    
    Alternatively, run a query against a group and verify that coverage is 100%.
  2. Run query load, using the tracking rank-profile, like rank_albums_track above
  3. Run queries using sorting or grouping on the mutable attributes.

To initialize a mutable attribute with a different value than 0 when a document is PUT, use:

field match_count type long {
    indexing: 7 | to_long | attribute | summary   # Initialized to 7 for a new document. Default is 0.
    attribute: mutable
}

To dump values fast, from memory only (assuming the schema has an id field):

document-summary rank_phase_statistics {
    summary id type int {}
    summary match_count type long {}
    summary first_phase_count type long {}
    summary second_phase_count type long {}
    summary summary_count type long {}
}
$ vespa query \
  "select * from music where true" \
  "presentation.summary=rank_phase_statistics"