Perform a substring search

In addition to full token matching, Spanner search indexes support substring searches. This page describes how to perform a substring search as part of a full-text search in Spanner.

Overview

Substring searches have the following characteristics:

  • Case insensitive, discards most punctuation, and normalizes whitespace.
  • No Chinese, Japanese, Korean (CJK) segmentation, since partial CJK queries often segment incorrectly.
  • For multiple search terms, the result must contain a substring from each term. For example, 'happ momen' matches "happy moment", because both substrings are found in the text. It doesn't match "happy day".

Examples

Stored text Substring query Match
Bridge over Troubled Water ridg roub Yes
Bridge over Troubled Water ridg , roub Yes
Bridge over Troubled Water over brid Yes
Bridge over Troubled Water ate bridge Yes
Bridge over Troubled Water Bridge bridge bridge Yes
Bridge over Troubled Water bri trou ter Yes
Bridge over Troubled Water bri dge Yes
Bridge over Troubled Water troubledwater No
Bridge over Troubled Water trubled No

For a substring search, use the TOKENIZE_SUBSTRING function in the TOKENLIST column definition, as shown in the following DDL example:

CREATE TABLE Albums (
  AlbumId STRING(MAX) NOT NULL,
  AlbumTitle STRING(MAX),
  AlbumTitle_Tokens TOKENLIST AS (TOKENIZE_SUBSTRING(AlbumTitle)) HIDDEN
) PRIMARY KEY(AlbumId);

In the SQL query, use the SEARCH_SUBSTRING function in the WHERE clause. For example, the following query matches an album with title "happy" from the table created in the previous example:

SELECT Album
FROM Albums
WHERE SEARCH_SUBSTRING(AlbumTitle_Tokens, 'happ');

TOKENIZE_SUBSTRING generates n-grams for each token and stores these n-grams in the search index. The minimum and maximum length of n-grams to generate are configured through optional arguments.

Substring search indexes can use 10-30x more storage as full-text indexes over the same data, because the tokenization produces a lot more tokens. This is especially true if as the difference between ngram_size_min and ngram_size_max grows. Substring queries also use more resources to execute.

Like TOKENIZE_FULLTEXT, you can configure TOKENIZE_SUBSTRING to use specific types of content.

In addition to the basic substring search, SEARCH_SUBSTRING supports the relative search mode. A relative search refines substring search results.

To enable the relative search mode, set the TOKENIZE_SUBSTRING relative_search_types parameter to TRUE.

When relative search is enabled in tokenization, SEARCH_SUBSTRING can perform queries with the following relative search types:

  • value_prefix: matches contiguous substrings and the match has to start at the beginning of the value. This is conceptually similar to the STARTS_WITH function for case and whitespace normalized strings.

    Examples

    Stored text Substring query Match
    Bridge over Troubled Water bridge over Yes
    Bridge over Troubled Water bridge , over Yes
    Bridge over Troubled Water ridge over No
    Bridge over Troubled Water troubled water No

  • value_suffix: matches contiguous substrings and the match has to match at the end of the value. This is conceptually similar to the ENDS_WITH function for case and whitespace normalized strings.

    Examples

    Stored text Substring query. Match
    Bridge over Troubled Water troubled water Yes
    Bridge over Troubled Water troubled ; water Yes
    Bridge over Troubled Water roubled water Yes
    Bridge over Troubled Water troubled wate No
    Bridge over Troubled Water trouble water No
    Bridge over Troubled Water bridge over No

  • word_prefix: like value_prefix, but the string has to match at a term boundary (rather than a value boundary).

    Examples

    Stored text Substring query Match
    Bridge over Troubled Water over trouble Yes
    Bridge over Troubled Water Over , trouble Yes
    Bridge over Troubled Water troub water No
    Bridge over Troubled Water over water No
    Bridge over Troubled Water ove troubled No
    Bridge over Troubled Water ver troubled Yes

  • word_suffix: like value_suffix, but the string has to match at the end of a term boundary.

    Examples

    Stored text Substring query Match
    Bridge over Troubled Water ver troubled Yes
    Bridge over Troubled Water over trouble No
    Bridge over Troubled Water over water No
    Bridge over Troubled Water ove troubled No

What's next