Rank search results

This page describes how to rank search results for full-text searches in Spanner.

Spanner supports computing a topicality score, which provides a building block for creating sophisticated ranking functions. These scores calculate the relevance of a result to a query, based on the query term frequency and other customizable options.

The following example shows a ranked search:

SELECT AlbumId
FROM Albums
WHERE SEARCH(AlbumTitle_Tokens, @p)
ORDER BY SCORE(AlbumTitle_Tokens, @p) DESC

Score query terms with the SCORE function

The SCORE function computes a score for each query term and then combines the scores. The per-term score is roughly based on term frequency–inverse document frequency (TF/IDF). The score is one component of the final ordering for a record. The query combines it with other signals, such as the freshness modulating the topicality score.

In the current implementation, the IDF part of TF/IDF is only available when enhance_query=>true is used. It calculates the relative frequency of words based on the full web corpus used by Google Search, rather than a specific search index. If rquery enhancement isn't enabled, the scoring only uses the term frequency (TF) component (that is, the IDF term is set to 1).

Usually, the enhance_query option is the same across both the SEARCH and SCORE functions.

The SCORE function returns values that serve as relevance scores that Spanner uses to establish a sort order. They have no standalone meaning. The higher the score, the better it matches the query.

Score multiple columns

Spanner uses the SCORE function to score each field individually. The query then combines these individual scores together. A common way of doing this is to sum up the individual scores and then boost them according to user-provided field weights (which are provided using SQL query parameters).

For example, the following query combines the output of two SCORE functions:

SELECT AlbumId
FROM Albums
WHERE SEARCH(Title_Tokens, @p1) AND SEARCH(Studio_Tokens, @p2)
ORDER BY SCORE(Title_Tokens, @p1) * @titleweight + SCORE(Studio_Tokens, @p2) * @studioweight
LIMIT 25

The following example adds two boost parameters:

• Freshness (FreshnessBoost) increases the score with (1 + @freshnessweight * GREATEST(0, 30 - DaysOld) / 30)
• Popularity(PopularityBoost) increases the score by multiplying it by factor (1 + IF(HasGrammy, @grammyweight, 0).

For readability, the query uses the WITH operator.

SELECT AlbumId
FROM Albums
WHERE SEARCH(Title_Tokens, @p1) AND SEARCH(Studio_Tokens, @p2)
ORDER BY WITH(
  TitleScore AS SCORE(Title_Tokens, @p1) * @titleweight,
  StudioScore AS SCORE(Studio_Tokens, @p2) * @studioweight,
  DaysOld AS (UNIX_MICROS(CURRENT_TIMESTAMP()) - ReleaseTimestamp) / 8.64e+10,
  FreshnessBoost AS (1 + @freshnessweight * GREATEST(0, 30 - DaysOld) / 30),
  PopularityBoost AS (1 + IF(HasGrammy, @grammyweight, 0)),
  (TitleScore + StudioScore) * FreshnessBoost * PopularityBoost)
LIMIT 25

Boost query order matches

You can apply a multiplicative boost to the topicality score for values that contain the query terms in the same order that they appear in the query. There are two versions of this boost: partial match and exact match. A partial match boost is applied when:

1. The TOKENLIST contains all the original terms in the query.
2. The tokens are adjacent to one another, and in the same order as they appear in the query.

There are certain special rules for conjunctions, negations, and phrases:

• A query with a negation can't receive a partial match boost.
• A query with a conjunction receives a boost if part of the conjunction appears in the appropriate locations.
• A query with a phrase receives a boost if the phrase appears in the TOKENLIST, and the term to the left of the phrase in the query appears to the left of the phrase in the TOKENLIST, and the same applies to the term to the right of the phrase.

Spanner applies an exact match boost when all of the previous rules are true. The first and last tokens in the query are the first and last tokens in the document.

Example document: Bridge Over Troubled Water

Limit retrieval depth

Search indexes often contain millions of documents. For queries where the predicates have low selectivity, it's impractical to rank all the results. Scoring queries usually have two limits:

1. Retrieval depth limit: the maximum number of rows to score.
2. Result set size limit: the maximum number of rows that the query should return (typically the page size).

Queries can limit retrieval depth with SQL subqueries:

SELECT *
FROM (
  SELECT AlbumId
  FROM Albums
  WHERE SEARCH(Title_Tokens, @p1)
  ORDER BY ReleaseTimestamp DESC
  LIMIT @retrieval_limit
)
ORDER BY SCORE(Title_Tokens, @p1)
LIMIT @page_size

This works particularly well if Spanner uses the most important ranking signal to sort the index.

What's next

• Learn about full-text search queries.
• Learn how to perform a substring search.
• Learn how to paginate search results.
• Learn how to mix full-text and non-text queries.
• Learn how to search multiple columns.