Overview

FTS3 is an SQLite virtual table module that allows users to perform full-text searches on a set of documents. The most common (and effective) way to describe full-text searches is "what Google, Yahoo and Altavista do with documents placed on the World Wide Web". Users input a term, or series of terms, perhaps connected by a binary operator or grouped together into a phrase, and the full-text query system finds the set of documents that best matches those terms considering the operators and groupings the user has specified. This document describes the deployment and usage of FTS3.

Portions of the original FTS3 code were contributed to the SQLite project by Scott Hess of Google. It is now developed and maintained as part of SQLite.

1. Introduction to FTS3

The FTS3 extension module allows users to create special tables with a built-in full-text index (hereafter "FTS3 tables"). The full-text index allows the user to efficiently query the database for all rows that contain one or more instances specified word (hereafter a "token", even if the table contains many large documents.

For example, if each of the 517430 documents in the "Enron E-Mail Dataset" is inserted into both the FTS3 table and the ordinary SQLite table created using the following SQL script:

CREATE VIRTUAL TABLE enrondata1 USING fts3(content TEXT);     /* FTS3 table */
CREATE TABLE enrondata2(content TEXT);                        /* Ordinary table */

Then either of the two queries below may be executed to find the number of documents in the database that contain the word "linux" (351). Using one desktop PC hardware configuration, the query on the FTS3 table returns in approximately 0.03 seconds, versus 22.5 for querying the ordinary table.

SELECT count(*) FROM enrondata1 WHERE content MATCH 'linux';  /* 0.03 seconds */
SELECT count(*) FROM enrondata2 WHERE content LIKE '%linux%'; /* 22.5 seconds */

Of course, the two queries above are not entirely equivalent. For example the LIKE query matches rows that contain terms such as "linuxophobe" or "EnterpriseLinux" (as it happens, the Enron E-Mail Dataset does not actually contain any such terms), whereas the MATCH query on the FTS3 table selects only those rows that contain "linux" as a discrete token. Both searches are case-insensitive. The FTS3 table consumes around 2006 MB on disk compared to just 1453 MB for the ordinary table. Using the same hardware configuration used to perform the SELECT queries above, the FTS3 table took just under 31 minutes to populate, versus 25 for the ordinary table.

1.1. Creating and Destroying FTS3 Tables

Like other virtual table types, new FTS3 tables are created using a CREATE VIRTUAL TABLE statement. The module name, which follows the USING keyword, is "fts3". The virtual table module arguments may be left empty, in which case an FTS3 table with a single user-defined column named "content" is created. Alternatively, the module arguments may be passed a list of comma separated column names.

If column names are explicitly provided for the FTS3 table as part of the CREATE VIRTUAL TABLE statement, then a datatype name may be optionally specified for each column. However, this is pure syntactic sugar, the supplied typenames are not used by FTS3 or the SQLite core for any purpose. The same applies to any constraints specified along with an FTS3 column name - they are parsed but not used or recorded by the system in any way.

-- Create an FTS3 table named "data" with one column - "content":
CREATE VIRTUAL TABLE data USING fts3();

-- Create an FTS3 table named "pages" with three columns:
CREATE VIRTUAL TABLE pages USING fts3(title, keywords, body);

-- Create an FTS3 table named "mail" with two columns. Datatypes
-- and column constraints are specified along with each column. These
-- are completely ignored by FTS3 and SQLite. 
CREATE VIRTUAL TABLE mail USING fts3(
  subject VARCHAR(256) NOT NULL,
  body TEXT CHECK(length(body)<10240)
);

As well as a list of columns, the module arguments passed to a CREATE VIRTUAL TABLE statement used to create an FTS3 table may be used to specify a tokenizer. This is done by specifying a string of the form "tokenize=<tokenizer name> <tokenizer args>" in place of a column name, where <tokenizer name> is the name of the tokenizer to use and <tokenizer args> is an optional list of whitespace separated qualifiers to pass to the tokenizer implementation. A tokenizer specification may be placed anywhere in the column list, but at most one tokenizer declaration is allowed for each CREATE VIRTUAL TABLE statement. The second and subsequent tokenizer declaration are interpreted as column names. See below for a detailed description of using (and, if necessary, implementing) a tokenizer.

-- Create an FTS3 table named "papers" with two columns that uses
-- the tokenizer "porter".
CREATE VIRTUAL TABLE papers USING fts3(author, document, tokenize=porter);

-- Create an FTS3 table with a single column - "content" - that uses
-- the "simple" tokenizer.
CREATE VIRTUAL TABLE data USING fts3(tokenize=simple);

-- Create an FTS3 table with two columns that uses the "icu" tokenizer.
-- The qualifier "en_AU" is passed to the tokenizer implementation
CREATE VIRTUAL TABLE names USING fts3(a, b, tokenize=icu en_AU);

FTS3 tables may be dropped from the database using an ordinary DROP TABLE statement. For example:

-- Create, then immediately drop, an FTS3 table.
CREATE VIRTUAL TABLE data USING fts3();
DROP TABLE data;

1.2. Populating FTS3 Tables

FTS3 tables are populated using INSERT, UPDATE and DELETE statements in the same way as ordinary SQLite tables are.

As well as the columns named by the user (or the "content" column if no module arguments where specified as part of the CREATE VIRTUAL TABLE statement), each FTS3 table has a "rowid" column. The rowid of an FTS3 table behaves in the same way as the rowid column of an ordinary SQLite table, except that the values stored in the rowid column of an FTS3 table remain unchanged if the database is rebuilt using the VACUUM command. For FTS3 tables, "docid" is allowed as an alias along with the usual "rowid", "oid" and "_oid_" identifiers. Attempting to insert or update a row with a docid value that already exists in the table is an error, just as it would be with an ordinary SQLite table.

There is one other subtle difference between "docid" and the normal SQLite aliases for the rowid column. Normally, if an INSERT or UPDATE statement assigns discrete values to two or more aliases of the rowid column, SQLite writes the rightmost of such values specified in the INSERT or UPDATE statement to the database. However, assigning a non-NULL value to both the "docid" and one or more of the SQLite rowid aliases when inserting or updating an FTS3 table is considered an error. See below for an example.

-- Create an FTS3 table
CREATE VIRTUAL TABLE pages USING fts3(title, body);

-- Insert a row with a specific docid value.
INSERT INTO pages(docid, title, body) VALUES(53, 'Home Page', 'SQLite is a software...');

-- Insert a row and allow FTS3 to assign a docid value using the same algorithm as
-- SQLite uses for ordinary tables. In this case the new docid will be 54,
-- one greater than the largest docid currently present in the table.
INSERT INTO pages(title, body) VALUES('Download', 'All SQLite source code...');

-- Change the title of the row just inserted.
UPDATE pages SET title = 'Download SQLite' WHERE rowid = 54;

-- Delete the entire table contents.
DELETE FROM pages;

-- The following is an error. It is not possible to assign non-NULL values to both
-- the rowid and docid columns of an FTS3 table.
INSERT INTO pages(rowid, docid, title, body) VALUES(1, 2, 'A title', 'A document body');

To support full-text queries, FTS3 maintains an inverted index that maps from each unique term or word that appears in the dataset to the locations in which it appears within the table contents. For the curious, a complete description of the data structure used to store this index within the database file is described below. A feature of this data structure is that at any time the database may contain not one index b-tree, but several different b-trees that are incrementally merged as rows are inserted, updated and deleted. This technique improves performance when writing to an FTS3 table, but causes some overhead for full-text queries that use the index. Executing an SQL statement of the form "INSERT INTO <fts3-table>(<fts3-table>) VALUES('optimize')" causes FTS3 to merge all existing index b-trees into a single large b-tree containing the entire index. This can be an expensive operation, but may speed up future queries.

For example, to optimize the full-text index for an FTS3 table named "docs":

-- Optimize the internal structure of FTS3 table "docs".
INSERT INTO docs(docs) VALUES('optimize');

The statement above may appear syntacticly incorrect to some. Refer to the section describing the simple fts3 queries for an explanation.

There is another, deprecated, method for invoking the optimize operation using a SELECT statement. New code should use statements similar to the INSERT above to optimize FTS3 structures.

1.3. Simple FTS3 Queries

As for all other SQLite tables, virtual or otherwise, data is retrieved from FTS3 tables using a SELECT statement.

FTS3 tables can be queried efficiently using SELECT statements of two different forms:

• Query by rowid. If the WHERE clause of the SELECT statement contains a sub-clause of the form "rowid = ?", where ? is an SQL expression, FTS3 is able to retrieve the requested row directly using the equivalent of an SQLite INTEGER PRIMARY KEY index.

• Full-text query. If the WHERE clause of the SELECT statement contains a sub-clause of the form "<column> MATCH ?", FTS3 is able to use the built-in full-text index to restrict the search to those documents that match the full-text query string specified as the right-hand operand of the MATCH clause.

If neither of the two query strategies enumerated above can be used, all queries on FTS3 tables are implemented using a linear scan of the entire table. If the table contains large amounts of data, this may be an impractically approach (the first example on this page shows that a linear scan of 1.5 GB of data takes around 30 seconds using a modern PC).

-- The examples in this block assume the following FTS3 table:
CREATE VIRTUAL TABLE mail USING fts3(subject, body);

SELECT * FROM mail WHERE rowid = 15;                -- Fast. Rowid lookup.
SELECT * FROM mail WHERE body MATCH 'sqlite';       -- Fast. Full-text query.
SELECT * FROM mail WHERE mail MATCH 'search';       -- Fast. Full-text query.
SELECT * FROM mail WHERE rowid BETWEEN 15 AND 20;   -- Slow. Linear scan.
SELECT * FROM mail WHERE subject = 'database';      -- Slow. Linear scan.
SELECT * FROM mail WHERE subject MATCH 'database';  -- Fast. Full-text query.

In all of the full-text queries above, the right-hand operand of the MATCH operator is a string consisting of a single term. In this case, the MATCH expression evaluates to true for all documents that contain one or more instances of the specified word ("sqlite", "search" or "database", depending on which example you look at). Specifying a single term as the right-hand operand of the MATCH operator results in the simplest (and most common) type of full-text query possible. However more complicated queries are possible, including phrase searches, term-prefix searches and searches for documents containing combinations of terms occuring within a defined proximity of each other. The various ways in which the full-text index may be queried are described below.

Normally, full-text queries are case-insensitive. However, this is is dependent on the specific tokenizer used by the FTS3 table being queried. Refer to the section on tokenizers for details.

The paragraph above notes that a MATCH operator with a simple term as the right-hand operand evaluates to true for all documents that contain the specified term. In this context, the "document" may refer to either the data stored in a single column of a row of an FTS3 table, or to the contents of all columns in a single row, depending on the identifier used as the left-hand operand to the MATCH operator. If the identifier specified as the left-hand operand of the MATCH operator is an FTS3 table column name, then the document that the search term must be contained in is the value stored in the specified column. However, if the identifier is the name of the FTS3 table itself, then the MATCH operator evaluates to true for each row of the FTS3 table for which any column contains the search term. The following example demonstrates this:

-- Example schema
CREATE VIRTUAL TABLE mail USING fts3(subject, body);

-- Example table population
INSERT INTO mail(docid, subject, body) VALUES(1, 'software feedback', 'found it too slow');
INSERT INTO mail(docid, subject, body) VALUES(2, 'software feedback', 'no feedback');
INSERT INTO mail(docid, subject, body) VALUES(3, 'slow lunch order',  'was a software problem');

-- Example queries
SELECT * FROM mail WHERE subject MATCH 'software';    -- Selects rows 1 and 2
SELECT * FROM mail WHERE body    MATCH 'feedback';    -- Selects row 2
SELECT * FROM mail WHERE mail    MATCH 'software';    -- Selects rows 1, 2 and 3
SELECT * FROM mail WHERE mail    MATCH 'slow';        -- Selects rows 1 and 3

At first glance, the final two full-text queries in the example above seem to be syntacticly incorrect, as there is a table name ("mail") used as an SQL expression. The reason this is acceptable is that each FTS3 table actually has a HIDDEN column with the same name as the table itself (in this case, "mail"). The value stored in this column is not meaningful to the application, but can be used as the left-hand operand to a MATCH operator. This special column may also be passed as an argument to the FTS3 auxiliary functions.

The following example illustrates the above. The expressions "docs", "docs.docs" and "main.docs.docs" all refer to column "docs". However, the expression "main.docs" does not refer to any column. It could be used to refer to a table, but a table name is not allowed in the context in which it is used below.

-- Example schema
CREATE VIRTUAL TABLE docs USING fts3(content);

-- Example queries
SELECT * FROM docs WHERE docs MATCH 'sqlite';              -- OK.
SELECT * FROM docs WHERE docs.docs MATCH 'sqlite';         -- OK.
SELECT * FROM docs WHERE main.docs.docs MATCH 'sqlite';    -- OK.
SELECT * FROM docs WHERE main.docs MATCH 'sqlite';         -- Error.

1.4. Summary

From the users point of view, FTS3 tables are similar to ordinary SQLite tables in many ways. Data may be added to, modified within and removed from FTS3 tables using the INSERT, UPDATE and DELETE commands just as it may be with ordinary tables. Similarly, the SELECT command may be used to query data. The following list summarizes the differences between FTS3 and ordinary tables:

1. As with all virtual table types, it is not possible to create indices or triggers attached to FTS3 tables. Nor is it possible to use the ALTER TABLE command to add extra columns to FTS3 tables (although it is possible to use ALTER TABLE to rename an FTS3 table).

2. Data-types specified as part of the "CREATE VIRTUAL TABLE" statement used to create an FTS3 table are ignored completely. Instead of the normal rules for applying type affinity to inserted values, all values inserted into FTS3 table columns (except the special rowid column) are converted to type TEXT before being stored.

3. FTS3 tables permit the special alias "docid" to be used to refer to the rowid column supported by all virtual tables.

4. The FTS3 MATCH operator is supported for queries based on the built-in full-text index.

5. The FTS3 auxiliary functions, snippet() and offsets(), are available to support full-text queries.

6. Each FTS3 table has a HIDDEN column with the same name as the table itself. The value contained in each row for the special column is only useful when used on the left-hand side of a MATCH operator, or when specified as an argument to one of the FTS3 auxiliary functions.

2. Compiling and Enabling FTS3

Although FTS3 is distributed as part of the SQLite source code, it is not enabled by default. To build SQLite with FTS3 functionality enabled, define the preprocessor macro SQLITE_ENABLE_FTS3 when compiling. New applications should also define the SQLITE_ENABLE_FTS3_PARENTHESIS macro to enable the enhanced query syntax (see below). Usually, this is done by adding the following two switches to the compiler command line:

-DSQLITE_ENABLE_FTS3
-DSQLITE_ENABLE_FTS3_PARENTHESIS

If using the amalgamation autoconf based build system, setting the CPPFLAGS environment variable while running the 'configure' script is an easy way to set these macros. For example, the following command:

CPPFLAGS="-DSQLITE_ENABLE_FTS3 -DSQLITE_ENABLE_FTS3_PARENTHESIS" ./configure <configure options>

where <configure options> are those options normally passed to the configure script, if any.

Because FTS3 is a virtual table, it is incompatible with the SQLITE_OMIT_VIRTUALTABLE option.

If an SQLite build does not include FTS3, then any attempt to prepare an SQL statement to create an FTS3 table or to drop or access an existing FTS3 table in any way will fail. The error message returned will be similar to "no such module: fts3".

If the C version of the ICU library is available, then FTS3 may also be compiled with the SQLITE_ENABLE_ICU pre-processor macro defined. Compiling with this macro enables an FTS3 tokenizer that uses the ICU library to split a document into terms (words) using the conventions for a specified language and locale.

-DSQLITE_ENABLE_ICU

3. Full-text Index Queries

The most useful thing about FTS3 tables is the queries that may be performed using the built-in full-text index. Full-text queries are performed by specifying a clause of the form "<column> MATCH <full-text query expression>" to the WHERE clause of a SELECT statement that reads data from an FTS3 table. Simple FTS3 queries that return all documents that contain a given term are described above. In that discussion the right-hand operand of the MATCH operator was assumed to be a string consisting of a single term. This section describes the more complex query types supported by FTS3 tables, and how they may be utilized by specifying a more complex query expression as the right-hand operand of a MATCH operator.

FTS3 tables support three basic query types:

• Token or token prefix queries. An FTS3 table may be queried for all documents that contain a specified term (the simple case described above), or for all documents that contain a term with a specified prefix. As we have seen, the query expression for a specific term is simply the term itself. The query expression used to search for a term prefix is the prefix itself with a '*' character appended to it. For example:

-- Virtual table declaration
CREATE VIRTUAL TABLE docs USING fts3(title, body);

-- Query for all documents containing the term "linux":
SELECT * FROM docs WHERE docs MATCH 'linux';

-- Query for all documents containing a term with the prefix "lin". This will match
-- all documents that contain "linux", but also those that contain terms "linear",
--"linker", "linguistic" and so on.
SELECT * FROM docs WHERE docs MATCH 'lin*';

• Normally, a token or token prefix query is matched against the FTS3 table column specified as the right-hand side of the MATCH operator. Or, if the special column with the same name as the FTS3 table itself is specified, against all columns. This may be overridden by specifying a column-name followed by a ":" character before a basic term query. There may be space between the ":" and the term to query for, but not between the column-name and the ":" character. For example:

-- Query the database for documents for which the term "linux" appears in
-- the document title, and the term "problems" appears in either the title
-- or body of the document.
SELECT * FROM docs WHERE docs MATCH 'title:linux problems';

-- Query the database for documents for which the term "linux" appears in
-- the document title, and the term "driver" appears in the body of the document
-- ("driver" may also appear in the title, but this alone will not satisfy the.
-- query criteria).
SELECT * FROM docs WHERE body MATCH 'title:linux driver';

• Phrase queries. A phrase query is a query that retrieves all documents that contain a nominated set of terms or term prefixes in a specified order with no intervening tokens. Phrase queries are specified by enclosing a space separated sequence of terms or term prefixes in double quotes ("). For example:

-- Query for all documents that contain the phrase "linux applications".
SELECT * FROM docs WHERE docs MATCH '"linux applications"';

-- Query for all documents that contain a phrase that matches "lin* app*". As well as
-- "linux applications", this will match common phrases such as "linoleum appliances"
-- or "link apprentice".
SELECT * FROM docs WHERE docs MATCH '"lin* app*"';

• NEAR queries. A NEAR query is a query that returns documents that contain a two or more nominated terms or phrases within a specified proximity of each other (by default with 10 or less intervening terms). A NEAR query is specified by putting the keyword "NEAR" between two phrase, term or term prefix queries. To specify a proximity other than the default, an operator of the form "NEAR/<N>" may be used, where <N> is the maximum number of intervening terms allowed. For example:

-- Virtual table declaration.
CREATE VIRTUAL TABLE docs USING fts3();

-- Virtual table data.
INSERT INTO docs VALUES('SQLite is an ACID compliant embedded relational database management system');

-- Search for a document that contains the terms "sqlite" and "database" with
-- not more than 10 intervening terms. This matches the only document in
-- table docs (since there are only six terms between "SQLite" and "database" 
-- in the document).
SELECT * FROM docs WHERE docs MATCH 'sqlite NEAR database';

-- Search for a document that contains the terms "sqlite" and "database" with
-- not more than 6 intervening terms. This also matches the only document in
-- table docs. Note that the order in which the terms appear in the document
-- does not have to be the same as the order in which they appear in the query.
SELECT * FROM docs WHERE docs MATCH 'database NEAR/6 sqlite';

-- Search for a document that contains the terms "sqlite" and "database" with
-- not more than 5 intervening terms. This query matches no documents.
SELECT * FROM docs WHERE docs MATCH 'database NEAR/5 sqlite';

-- Search for a document that contains the phrase "ACID compliant" and the term
-- "database" with not more than 2 terms separating the two. This matches the
-- document stored in table docs.
SELECT * FROM docs WHERE docs MATCH 'database NEAR/2 "ACID compliant"';

-- Search for a document that contains the phrase "ACID compliant" and the term
-- "sqlite" with not more than 2 terms separating the two. This also matches
-- the only document stored in table docs.
SELECT * FROM docs WHERE docs MATCH '"ACID compliant" NEAR/2 sqlite';

• More than one NEAR operator may appear in a single query. In this case each pair of terms or phrases separated by a NEAR operator must appear within the specified proximity of each other in the document. Using the same table and data as in the block of examples above:

-- The following query selects documents that contains an instance of the term 
 -- "sqlite" separated by two or fewer terms from an instance of the term "acid",
 -- which is in turn separated by two or fewer terms from an instance of the term
 -- "relational". As it happens, the only document in table docs satisfies this criteria.
 SELECT * FROM docs WHERE docs MATCH 'sqlite NEAR/2 acid NEAR/2 relational';

 -- This query matches no documents. There is an instance of the term "sqlite" with
 -- sufficient proximity to an instance of "acid" but it is not sufficiently close
 -- to an instance of the term "relational".
 SELECT * FROM docs WHERE docs MATCH 'acid NEAR/2 sqlite NEAR/2 relational';

Phrase and NEAR queries may not span multiple columns within a row.

The three basic query types described above may be used to query the full-text index for the set of documents that match the specified criteria. Using the FTS3 query expression language it is possible to perform various set operations on the results of basic queries. There are currently three supported operations:

• The AND operator determines the intersection of two sets of documents.
• The OR operator calculates the union of two sets of documents.
• The NOT operator (or, if using the standard syntax, a unary "-" operator) may be used to compute the relative complement of one set of documents with respect to another.

The FTS3 module may be compiled to use one of two slightly different versions of the full-text query syntax, the "standard" query syntax and the "enhanced" query syntax. The basic term, term-prefix, phrase and NEAR queries described above are the same in both versions of the syntax. The way in which set operations are specified is slightly different. The following two sub-sections describe the part of the two query syntaxes that pertains to set operations. Refer to the description of how to compile fts3 for compilation notes.

3.1. Set Operations Using The Enhanced Query Syntax

The enhanced query syntax supports the AND, OR and NOT binary set operators. Each of the two operands to an operator may be a basic FTS3 query, or the result of another AND, OR or NOT set operation. Operators must be entered using capital letters. Otherwise, they are interpreted as basic term queries instead of set operators.

The AND operator may be implicitly specified. If two basic queries appear with no operator separating them in an FTS3 query string, the results are the same as if the two basic queries were separated by an AND operator. For example, the query expression "implicit operator" is a more succinct version of "implicit AND operator".

-- Virtual table declaration
CREATE VIRTUAL TABLE docs USING fts3();

-- Virtual table data
INSERT INTO docs(docid, content) VALUES(1, 'a database is a software system');
INSERT INTO docs(docid, content) VALUES(2, 'sqlite is a software system');
INSERT INTO docs(docid, content) VALUES(3, 'sqlite is a database');

-- Return the set of documents that contain the term "sqlite", and the
-- term "database". This query will return the document with docid 3 only.
SELECT * FROM docs WHERE docs MATCH 'sqlite AND database';

-- Again, return the set of documents that contain both "sqlite" and
-- "database". This time, use an implicit AND operator. Again, document
-- 3 is the only document matched by this query. 
SELECT * FROM docs WHERE docs MATCH 'database sqlite';

-- Query for the set of documents that contains either "sqlite" or "database".
-- All three documents in the database are matched by this query.
SELECT * FROM docs WHERE docs MATCH 'sqlite OR database';

-- Query for all documents that contain the term "database", but do not contain
-- the term "sqlite". Document 1 is the only document that matches this criteria.
SELECT * FROM docs WHERE docs MATCH 'database NOT sqlite';

-- The following query matches no documents. Because "and" is in lowercase letters,
-- it is interpreted as a basic term query instead of an operator. Operators must
-- be specified using capital letters. In practice, this query will match any documents
-- that contain each of the three terms "database", "and" and "sqlite" at least once.
-- No documents in the example data above match this criteria.
SELECT * FROM docs WHERE docs MATCH 'database and sqlite';

The examples above all use basic full-text term queries as both operands of the set operations demonstrated. Phrase and NEAR queries may also be used, as may the results of other set operations. When more than one set operation is present in an FTS3 query, the precedence of operators is as follows:

When using the enhanced query syntax, parenthesis may be used to override the default precedence of the various operators. For example:

-- Return the docid values associated with all documents that contain the
-- two terms "sqlite" and "database", and/or contain the term "library".
SELECT docid FROM docs WHERE docs MATCH 'sqlite AND database OR library';

-- This query is equivalent to the above.
SELECT docid FROM docs WHERE docs MATCH 'sqlite AND database'
  UNION
SELECT docid FROM docs WHERE docs MATCH 'library';

-- Query for the set of documents that contains the term "linux", and at least
-- one of the phrases "sqlite database" and "sqlite library".
SELECT docid FROM docs WHERE docs MATCH '("sqlite database" OR "sqlite library") AND linux';

-- This query is equivalent to the above.
SELECT docid FROM docs WHERE docs MATCH 'linux'
  INTERSECT
SELECT docid FROM (
  SELECT docid FROM docs WHERE docs MATCH '"sqlite library"'
    UNION
  SELECT docid FROM docs WHERE docs MATCH '"sqlite database"'
);

3.2. Set Operations Using The Standard Query Syntax

FTS3 query set operations using the standard query syntax are similar, but not identical, to set operations with the enhanced query syntax. There are four differences, as follows:

1. Only the implicit version of the AND operator is supported. Specifying the string "AND" as part of an standard query syntax query is interpreted as a term query for the set of documents containing the term "and".

2. Parenthesis are not supported.

3. The NOT operator is not supported. Instead of the NOT operator, the standard query syntax supports a unary "-" operator that may be applied to basic term and term-prefix queries (but not to phrase or NEAR queries). A term or term-prefix that has a unary "-" operator attached to it may not appear as an operand to an OR operator. An FTS3 query may not consist entirely of terms or term-prefix queries with unary "-" operators attached to them.

-- Search for the set of documents that contain the term "sqlite" but do
-- not contain the term "database".
SELECT * FROM docs WHERE docs MATCH 'sqlite -database';

4. The relative precedence of the set operations is different. In particular, using the standard query syntax the "OR" operator has a higher precedence than "AND". The precedence of operators when using the standard query syntax is:

1. The following example illustrates precedence of operators using the standard query syntax:

-- Search for documents that contains at least one of the terms "database"
-- and "sqlite", and also contains the term "library". Because of the differences
-- in operator precedences, this query would have a different interpretation using
-- the enhanced query syntax.
SELECT * FROM docs WHERE docs MATCH 'sqlite OR database library';

4. auxiliary functions - Snippet, Offsets and Matchinfo

The FTS3 module provides three special SQL scalar functions that may be useful to the developers of full-text query systems: "snippet", "offsets" and "matchinfo". The purpose of the "snippet" and "offsets" functions is to allow the user to identify the location of queried terms in the returned documents. The "matchinfo" function provides the user with metrics that may be useful for filtering or sorting query results according to relevance.

The first argument to all three special SQL scalar functions must be the the special hidden column of an FTS3 table that has the same name as the table (see above). For example, given an FTS3 table named "mail":

SELECT offsets(mail) FROM mail WHERE mail MATCH <full-text query expression>;
SELECT snippet(mail) FROM mail WHERE mail MATCH <full-text query expression>;
SELECT matchinfo(mail) FROM mail WHERE mail MATCH <full-text query expression>;

The three auxiliary functions are only useful within a SELECT statement that uses the FTS3 table's full-text index. If used within a SELECT that uses the "query by rowid" or "linear scan" strategies, then the snippet and offsets both return an an empty string, and the matchinfo function returns a blob value zero bytes in size.

All three auxiliary functions extract a set of "matchable phrases" from the FTS3 query expression to work with. The set of matchable phrases for a given query consists of all phrases (including unquoted tokens and token prefixes) in the expression except those that are prefixed with a unary "-" operator (standard syntax) or are part of a sub-expression that is used as the right-hand operand of a NOT operator.

With the following provisos, each series of tokens in the FTS3 table that matches one of the matchable phrases in the query expression is known as a "phrase match":

1. If a matchable phrase is part of a series of phrases connected by NEAR operators in the FTS3 query expression, then each phrase match must be sufficiently close to other phrase matches of the relevant types to satisfy the NEAR condition.
2. If the matchable phrase in the FTS3 query is restricted to matching data in a specified FTS3 table column, then only phrase matches that occur within that column are considered.

4.1. The Offsets Function

For a SELECT query that uses the full-text index, the offsets() function returns a text value containing a series of space-separated integers. For each term in each phrase match of the current row, there are four integers in the returned list. Each set of four integers is interpreted as follows:

The following block contains examples that use the offsets function.

CREATE VIRTUAL TABLE mail USING fts3(subject, body);
INSERT INTO mail VALUES('hello world', 'This message is a hello world message.');
INSERT INTO mail VALUES('urgent: serious', 'This mail is seen as a more serious mail');

-- The following query returns a single row (as it matches only the first
-- entry in table "mail". The text returned by the offsets function is
-- "0 0 6 5 1 0 24 5".
--
-- The first set of four integers in the result indicate that column 0
-- contains an instance of term 0 ("world") at byte offset 6. The term instance
-- is 5 bytes in size. The second set of four integers shows that column 1
-- of the matched row contains an instance of term 0 ("world") at byte offset
-- 24. Again, the term instance is 5 bytes in size.
SELECT offsets(mail) FROM mail WHERE mail MATCH 'world';

-- The following query returns also matches only the first row in table "mail".
-- In this case the returned text is "1 0 5 7 1 0 30 7".
SELECT offsets(mail) FROM mail WHERE mail MATCH 'message';

-- The following query matches the second row in table "mail". It returns the
-- text "1 0 28 7 1 1 36 4". Only those occurrences of terms "serious" and "mail"
-- that are part of an instance of the phrase "serious mail" are identified; the
-- other occurrences of "serious" and "mail" are ignored.
SELECT offsets(mail) FROM mail WHERE mail MATCH '"serious mail"';

4.2. The Snippet Function

The snippet function is used to create formatted fragments of document text for display as part of a full-text query results report. The snippet function may be passed between one and six arguments, as follows:

The snippet function first attempts to find a fragment of text consisting of |N| tokens within the current row that contains at least one phrase match for each matchable phrase matched somewhere in the current row, where |N| is the absolute value of the sixth argument passed to the snippet function. If the text stored in a single column contains less than |N| tokens, then the entire column value is considered. Text fragments may not span multiple columns.

If such a text fragment can be found, it is returned with the following modifications:

• If the text fragment does not begin at the start of a column value, the "ellipses" text is prepended to it.
• If the text fragment does not finish at the end of a column value, the "ellipses" text is appended to it.
• For each token in the text fragment that is part of a phrase match, the "start match" text is inserted into the fragment before the token, and the "end match" text is inserted immediately after it.

If more than one such fragment can be found, then fragments that contain a larger number of "extra" phrase matches are favoured. The start of the selected text fragment may be moved a few tokens forward or backward to attempt to concentrate the phrase matches toward the center of the fragment.

Assuming N is a positive value, if no fragments can be found that contain an phrase match corresponding to each matchable phrase, the snippet function attempts to find two fragments of approximately N/2 tokens that between them contain at least one phrase match for each matchable phrase matched by the current row. If this fails, attempts are made to find three fragments of N/3 tokens each and finally four N/4 token fragments. If a set of four fragments cannot be found that encompasses the required phrase matches, the four fragments of N/4 tokens that provide the best coverage are selected.

If N is a negative value, and no single fragment can be found containing the required phrase matches, the snippet function searches for two fragments of |N| tokens each, then three, then four. In other words, if the specified value of N is negative, the sizes of the fragments is not decreased if more than one fragment is required to provide the desired phrase match coverage.

After the M fragments have been located, where M is between two and four as described in the paragraphs above, they are joined together in sorted order with the "ellipses" text separating them. The three modifications enumerated earlier are performed on the text before it is returned.

Note: In this block of examples, newlines and whitespace characters have
been inserted into the document inserted into the FTS3 table, and the expected
results described in SQL comments. This is done to enhance readability only,
they would not be present in actual SQLite commands or output.

-- Create and populate an FTS3 table.
CREATE VIRTUAL TABLE text USING fts3();
INSERT INTO text VALUES('
  During 30 Nov-1 Dec, 2-3oC drops. Cool in the upper portion, minimum temperature 14-16oC 
  and cool elsewhere, minimum temperature 17-20oC. Cold to very cold on mountaintops, 
  minimum temperature 6-12oC. Northeasterly winds 15-30 km/hr. After that, temperature 
  increases. Northeasterly winds 15-30 km/hr.     
');

-- The following query returns the text value:
--
--   "<b>...</b>cool elsewhere, minimum temperature 17-20oC. <b>Cold</b> to very 
--    <b>cold</b> on mountaintops, minimum temperature 6<b>...</b>".
--
SELECT snippet(text) FROM text WHERE text MATCH 'cold';

-- The following query returns the text value:
--
--   "...the upper portion, [minimum] [temperature] 14-16oC and cool elsewhere,
--    [minimum] [temperature] 17-20oC. Cold..."
--
SELECT snippet(text, '[ ']', '...') FROM text WHERE text MATCH '"min* tem*"'

4.3. The Matchinfo Function

The matchinfo function returns a blob value. If used within a query that uses the full-text index (not a "query by rowid" or "linear scan"), then the blob consists of (2 + C * P * 3) 32-bit unsigned integers in machine byte-order, where C is the number of columns in the FTS3 table being queried, and P is the number of matchable phrases in the query.

Phrases and columns are both numbered from left to right starting from zero.

For example:

-- Create and populate an FTS3 table with two columns:
CREATE VIRTUAL TABLE t1 USING fts3(a, b);
INSERT INTO t1 VALUES('transaction default models default', 'Non transaction reads');
INSERT INTO t1 VALUES('the default transaction', 'these semantics present');
INSERT INTO t1 VALUES('single request', 'default data');

-- The following query returns a single row consisting of a single blob
-- value 80 bytes in size (20 32-bit integers). If each block of 4 bytes in
-- the blob is interpreted as an unsigned integer in machine byte-order,
-- the integers will be:
--
--     3 2  1 3 2  0 1 1  1 2 2  0 1 1  0 0 0  1 1 1
--
-- The row returned corresponds to the second entry inserted into table t1.
-- The first two integers in the blob show that the query contained three
-- phrases and the table being queried has two columns. The next block of
-- three integers describes column 0 (in this case column "a") and phrase
-- 0 (in this case "default"). The current row contains 1 hit for "default"
-- in column 0, of a total of 3 hits for "default" that occur in column
-- 0 of any table row. The 3 hits are spread across 2 different rows.
--
-- The next set of three integers (0 1 1) pertain to the hits for "default"
-- in column 1 of the table (0 in this row, 1 in all rows, spread across 
-- 1 rows).
--
SELECT matchinfo(t1) FROM t1 WHERE t1 MATCH 'default transaction "these semantics"';

The matchinfo function is much faster than either the snippet or offsets functions. This is because the implementation of both snippet and offsets is required to retrieve the documents being analyzed from disk, whereas all data required by matchinfo is available as part of the same portions of the full-text index that are required to implement the full-text query itself. This means that of the following two queries, the first may be an order of magnitude faster than the second:

SELECT docid, matchinfo(tbl) FROM tbl WHERE tbl MATCH <query expression>;
SELECT docid, offsets(tbl) FROM tbl WHERE tbl MATCH <query expression>;

The matchinfo function provides much of the information required to calculate probabilistic "bag-of-words" relevancy scores such as Okapi BM25/BM25F that may be used to order results in a full-text search application. Also often used in such functions is the length or relative length of each document or document field. Unfortunately, this information is not made available by the matchinfo function as it would require loading extra data from the database, potentially slowing matchinfo() down by an order of magnitude. One solution is for the application to store the lengths of each document or document field in a separate table for use in calculating relevancy scores. Appendix A of this document, "search application tips", contains an example of using the matchinfo() function efficiently.

5. Tokenizers

An FTS3 tokenizer is a set of rules for extracting terms from a document or basic FTS3 full-text query.

Unless a specific tokenizer is specified as part of the CREATE VIRTUAL TABLE statement used to create the FTS3 table, the default tokenizer, "simple", is used. The simple tokenizer extracts tokens from a document or basic FTS3 full-text query according to the following rules:

• A term is a contiguous sequence of eligible characters, where eligible characters are all alphanumeric characters, the "_" character, and all characters with UTF codepoints greater than or equal to 128. All other characters are discarded when splitting a document into terms. They serve only to separate adjacent terms.

• All uppercase characters within the ASCII range (UTF codepoints less than 128), are transformed to their lowercase equivalents as part of the tokenization process. Thus, full-text queries are case-insensitive when using the simple tokenizer.

For example, when a document containing the text "Right now, they're very frustrated.", the terms extracted from the document and added to the full-text index are, in order, "right now they re very frustrated". Such a document would match a full-text query such as "MATCH 'Frustrated'", as the simple tokenizer transforms the term in the query to lowercase before searching the full-text index.

As well as the "simple" tokenizer, the FTS3 source code features a tokenizer that uses the Porter Stemming algorithm. This tokenizer uses the same rules to separate the input document into terms, but as well as folding all terms to lower case it uses the Porter Stemming algorithm to reduce related English language words to a common root. For example, using the same input document as in the paragraph above, the porter tokenizer extracts the following tokens: "right now thei veri frustrat". Even though some of these terms are not even English words, in some cases using them to build the full-text index is more useful than the more intelligible output produced by the simple tokenizer. Using the porter tokenizer, the document not only matches full-text queries such as "MATCH 'Frustrated'", but also queries such as "MATCH 'Frustration'", as the term "Frustration" is reduced by the Porter stemmer algorithm to "frustrat" - just as "Frustrated" is. So, when using the porter tokenizer, FTS3 is able to find not just exact matches for queried terms, but matches against similar English language terms. For more information on the Porter Stemmer algorithm, please refer to the page linked above.

Example illustrating the difference between the "simple" and "porter" tokenizers:

-- Create a table using the simple tokenizer. Insert a document into it.
CREATE VIRTUAL TABLE simple USING fts3(tokenize=simple);
INSERT INTO simple VALUES('Right now they''re very frustrated');

-- The first of the following two queries matches the document stored in
-- table "simple". The second does not.
SELECT * FROM simple WHERE simple MATCH 'Frustrated');
SELECT * FROM simple WHERE simple MATCH 'Frustration');

-- Create a table using the porter tokenizer. Insert the same document into it
CREATE VIRTUAL TABLE porter USING fts3(tokenize=porter);
INSERT INTO porter VALUES('Right now they''re very frustrated');

-- Both of the following queries match the document stored in table "porter".
SELECT * FROM porter WHERE porter MATCH 'Frustrated');
SELECT * FROM porter WHERE porter MATCH 'Frustration');

If this extension is compiled with the SQLITE_ENABLE_ICU pre-processor symbol defined, then there exists a built-in tokenizer named "icu" implemented using the ICU library. The first argument passed to the xCreate() method (see fts3_tokenizer.h) of this tokenizer may be an ICU locale identifier. For example "tr_TR" for Turkish as used in Turkey, or "en_AU" for English as used in Australia. For example:

CREATE VIRTUAL TABLE thai_text USING fts3(text, tokenize=icu th_TH)

The ICU tokenizer implementation is very simple. It splits the input text according to the ICU rules for finding word boundaries and discards any tokens that consist entirely of white-space. This may be suitable for some applications in some locales, but not all. If more complex processing is required, for example to implement stemming or discard punctuation, this can be done by creating a tokenizer implementation that uses the ICU tokenizer as part of its implementation.

5.1. Custom (User Implemented) Tokenizers

As well as the built-in "simple", "porter" and (possibly) "icu" tokenizers, FTS3 exports an interface that allows users to implement custom tokenizers using C. The interface used to create a new tokenizer is defined and described in the fts3_tokenizer.h source file.

Registering a new FTS3 tokenizer is similar to registering a new virtual table module with SQLite. The user passes a pointer to a structure containing pointers to various callback functions that make up the implementation of the new tokenizer type. For tokenizers, the structure (defined in fts3_tokenizer.h) is called "sqlite3_tokenizer_module".

FTS3 does not expose a C-function that users call to register new tokenizer types with a database handle. Instead, the pointer must be encoded as an SQL blob value and passed to FTS3 through the SQL engine by evaluating a special scalar function, "fts3_tokenizer()". The fts3_tokenizer() function may be called with one or two arguments, as follows:

SELECT fts3_tokenizer(<tokenizer-name>);
SELECT fts3_tokenizer(<tokenizer-name>, <sqlite3_tokenizer_module ptr>);

Where is a string identifying the tokenizer and is a pointer to an sqlite3_tokenizer_module structure encoded as an SQL blob. If the second argument is present, it is registered as tokenizer and a copy of it returned. If only one argument is passed, a pointer to the tokenizer implementation currently registered as is returned, encoded as a blob. Or, if no such tokenizer exists, an SQL exception (error) is raised.

SECURITY WARNING: If the fts3 extension is used in an environment where potentially malicious users may execute arbitrary SQL, they should be prevented from invoking the fts3_tokenizer() function, possibly using the authorisation callback.

The following block contains an example of calling the fts3_tokenizer() function from C code:

/*
** Register a tokenizer implementation with FTS3.
*/
int registerTokenizer(
  sqlite3 *db,
  char *zName,
  const sqlite3_tokenizer_module *p
){
  int rc;
  sqlite3_stmt *pStmt;
  const char *zSql = "SELECT fts3_tokenizer(?, ?)";

  rc = sqlite3_prepare_v2(db, zSql, -1, &pStmt, 0);
  if( rc!=SQLITE_OK ){
    return rc;
  }

  sqlite3_bind_text(pStmt, 1, zName, -1, SQLITE_STATIC);
  sqlite3_bind_blob(pStmt, 2, &p, sizeof(p), SQLITE_STATIC);
  sqlite3_step(pStmt);

  return sqlite3_finalize(pStmt);
}

/*
** Query FTS3 for the tokenizer implementation named zName.
*/
int queryTokenizer(
  sqlite3 *db,
  char *zName,
  const sqlite3_tokenizer_module **pp
){
  int rc;
  sqlite3_stmt *pStmt;
  const char *zSql = "SELECT fts3_tokenizer(?)";

  *pp = 0;
  rc = sqlite3_prepare_v2(db, zSql, -1, &pStmt, 0);
  if( rc!=SQLITE_OK ){
    return rc;
  }

  sqlite3_bind_text(pStmt, 1, zName, -1, SQLITE_STATIC);
  if( SQLITE_ROW==sqlite3_step(pStmt) ){
    if( sqlite3_column_type(pStmt, 0)==SQLITE_BLOB ){
      memcpy(pp, sqlite3_column_blob(pStmt, 0), sizeof(*pp));
    }
  }

  return sqlite3_finalize(pStmt);
}

6. Data Structures

This section describes at a high-level the way the FTS3 module stores its index and content in the database. It is not necessary to read or understand the material in this section in order to use FTS3 in an application. However, it may be useful to application developers attempting to analyze and understand FTS3 performance characteristics, or to developers contemplating enhancements to the existing FTS3 feature set.

For each FTS3 virtual table in a database, three real (non-virtual) tables are created to store the underlying data. The real tables are named "%_content", "%_segdir" and "%_segments", where "%" is replaced by the name supplied by the user for the FTS3 virtual table.

The leftmost column of the "%_content" table is an INTEGER PRIMARY KEY field named "docid". Following this is one column for each column of the FTS3 virtual table as declared by the user, named by prepending the column name supplied by the user with "cN", where N is the index of the column within the table, numbered from left to right starting with 1. Data types supplied as part of the virtual table declaration are not used as part of the %_content table declaration. For example:

-- Virtual table declaration
CREATE VIRTUAL TABLE abc USING FTS3(a NUMBER, b TEXT, c);

-- Corresponding %_content table declaration
CREATE TABLE abc_content(docid INTEGER PRIMARY KEY, c0a, c1b, c2c);

The %_content table contains the unadulterated data inserted by the user into the FTS3 virtual table by the user. If the user does not explicitly supply a "docid" value when inserting records, one is selected automatically by the system.

The two remaining tables, %_segments and %_segdir, are used to store the full-text index. Conceptually, this index is a lookup table that maps each term (word) to the set of docid values corresponding to records in the %_content table that contain one or more occurrences of the term. To retrieve all documents that contain a specified term, the FTS3 module queries this index to determine the set of docid values for records that contain the term, then retrieves the required documents from the %_content table. Regardless of the schema of the FTS3 virtual table, the %_segments and %_segdir tables are always created as follows:

CREATE TABLE %_segments(
  blockid INTEGER PRIMARY KEY,       -- B-tree node id
  block blob                         -- B-tree node data
);

CREATE TABLE %_segdir(
  level INTEGER,
  idx INTEGER,
  start_block INTEGER,               -- Blockid of first node in %_segments
  leaves_end_block INTEGER,          -- Blockid of last leaf node in %_segments
  end_block INTEGER,                 -- Blockid of last node in %_segments
  root BLOB,                         -- B-tree root node
  PRIMARY KEY(level, idx)
);

The schema depicted above is not designed to store the full-text index directly. Instead, it is used to one or more b-tree structures. There is one b-tree for each row in the %_segdir table. The %_segdir table row contains the root node and various meta-data associated with the b-tree structure, and the %_segments table contains all other (non-root) b-tree nodes. Each b-tree is referred to as a "segment". Once it has been created, a segment b-tree is never updated (although it may be deleted altogether).

The keys used by each segment b-tree are terms (words). As well as the key, each segment b-tree entry has an associated "doclist" (document list). A doclist consists of zero or more entries, where each entry consists of:

• A docid (document id), and
• A list of term offsets, one for each occurrence of the term within the document. A term offset indicates the number of tokens (words) that occur before the term in question, not the number of characters or bytes. For example, the term offset of the term "war" in the phrase "Ancestral voices prophesying war!" is 3.

Entries within a doclist are sorted by docid. Positions within a doclist entry are stored in ascending order.

The contents of the logical full-text index is found by merging the contents of all segment b-trees. If a term is present in more than one segment b-tree, then it maps to the union of each individual doclist. If, for a single term, the same docid occurs in more than one doclist, then only the doclist that is part of the most recently created segment b-tree is considered valid.

Multiple b-tree structures are used instead of a single b-tree to reduce the cost of inserting records into FTS3 tables. When a new record is inserted into an FTS3 table that already contains a lot of data, it is likely that many of the terms in the new record are already present in a large number of existing records. If a single b-tree were used, then large doclist structures would have to be loaded from the database, amended to include the new docid and term-offset list, then written back to the database. Using multiple b-tree tables allows this to be avoided by creating a new b-tree which can be merged with the existing b-tree (or b-trees) later on. Merging of b-tree structures can be performed as a background task, or once a certain number of separate b-tree structures have been accumulated. Of course, this scheme makes queries more expensive (as the FTS3 code may have to look up individual terms in more than one b-tree and merge the results), but it has been found that in practice this overhead is often negligible.

6.1. Variable Length Integer (varint) Format

Integer values stored as part of segment b-tree nodes are encoded using the FTS3 varint format. This encoding is similar, but not identical, to the the SQLite varint format.

An encoded FTS3 varint consumes between one and ten bytes of space. The number of bytes required is determined by the sign and magnitude of the integer value encoded. More accurately, the number of bytes used to store the encoded integer depends on the position of the most significant set bit in the 64-bit twos-complement representation of the integer value. Negative values always have the most significant bit set (the sign bit), and so are always stored using the full ten bytes. Positive integer values may be stored using less space.

The final byte of an encoded FTS3 varint has its most significant bit cleared. All preceding bytes have the most significant bit set. Data is stored in the remaining seven least significant bits of each byte. The first byte of the encoded representation contains the least significant seven bits of the encoded integer value. The second byte of the encoded representation, if it is present, contains the seven next least significant bits of the integer value, and so on. The following table contains examples of encoded integer values:

6.2. Segment B-Tree Format

Segment b-trees are prefix-compressed b+-trees. There is one segment b-tree for each row in the %_segdir table (see above). The root node of the segment b-tree is stored as a blob in the "root" field of the corresponding row of the %_segdir table. All other nodes (if any exist) are stored in the "blob" column of the %_segments table. Nodes within the %_segments table are identified by the integer value in the blockid field of the corresponding row. The following table describes the fields of the %_segdir table:

Apart from the root node, the nodes that make up a single segment b-tree are always stored using a contiguous sequence of blockids. Furthermore, the nodes that make up a single level of the b-tree are themselves stored as a contiguous block, in b-tree order. The contiguous sequence of blockids used to store the b-tree leaves are allocated starting with the blockid value stored in the "start_block" column of the corresponding %_segdir row, and finishing at the blockid value stored in the "leaves_end_block" field of the same row. It is therefore possible to iterate through all the leaves of a segment b-tree, in key order, by traversing the %_segments table in blockid order from "start_block" to "leaves_end_block".

6.2.1. Segment B-Tree Leaf Nodes

The following diagram depicts the format of a segment b-tree leaf node.

Segment B-Tree Leaf Node Format

The first term stored on each node ("Term 1" in the figure above) is stored verbatim. Each subsequent term is prefix-compressed with respect to its predecessor. Terms are stored within a page in sorted (memcmp) order.

6.2.2. Segment B-Tree Interior Nodes

The following diagram depicts the format of a segment b-tree interior (non-leaf) node.

Segment B-Tree Interior Node Format

6.3. Doclist Format

A doclist consists of an array of 64-bit signed integers, serialized using the FTS3 varint format. Each doclist entry is made up of a series of two or more integers, as follows:

1. The docid value. The first entry in a doclist contains the literal docid value. The first field of each subsequent doclist entry contains the difference between the new docid and the previous one (always a positive number).
2. Zero or more term-offset lists. A term-offset list is present for each column of the FTS3 virtual table that contains the term. A term-offset list consists of the following:
   1. Constant value 1. This field is omitted for any term-offset list associated with column 0.
   2. The column number (1 for the second leftmost column, etc.). This field is omitted for any term-offset list associated with column 0.
   3. A list of term-offsets, sorted from smallest to largest. Instead of storing the term-offset value literally, each integer stored is the difference between the current term-offset and the previous one (or zero if the current term-offset is the first), plus 2.
3. Constant value 0.

FTS3 Doclist Format

FTS3 Doclist Entry Format

For doclists for which the term appears in more than one column of the FTS3 virtual table, term-offset lists within the doclist are stored in column number order. This ensures that the term-offset list associated with column 0 (if any) is always first, allowing the first two fields of the term-offset list to be omitted in this case.

Appendix A: Search Application Tips

FTS3 is primarily designed to support Boolean full-text queries - queries to find the set of documents that match a specified criteria. However, many (most?) search applications require that results are somehow ranked in order of "relevance", where "relevance" is defined as the likelihood that the user who performed the search is interested in a specific element of the returned set of documents. When using a search engine to find documents on the world wide web, the user expects that the most useful, or "relevant", documents will be returned as the first page of results, and that each subsequent page contains progressively less relevant results. Exactly how a machine can determine document relevance based on a users query is a complicated problem and the subject of much ongoing research.

One very simple scheme might be to count the number of instances of the users search terms in each result document. Those documents that contain many instances of the terms are considered more relevant than those with a small number of instances of each term. In an FTS3 application, the number of term instances in each result could be determined by counting the number of integers in the return value of the offsets function. The following example shows a query that could be used to obtain the ten most relevant results for a query entered by the user:

-- This example (and all others in this section) assumes the following schema
CREATE VIRTUAL TABLE documents USING fts3(title, content);

-- Assuming the application has supplied an SQLite user function named "countintegers"
-- that returns the number of space-separated integers contained in its only argument,
-- the following query could be used to return the titles of the 10 documents that contain
-- the greatest number of instances of the users query terms. Hopefully, these 10
-- documents will be those that the users considers more or less the most "relevant".
SELECT title FROM documents 
  WHERE documents MATCH <query>
  ORDER BY countintegers(offsets(document)) DESC
  OFFSET 0 LIMIT 10

The query above could be made to run faster by using the FTS3 matchinfo function to determine the number of query term instances that appear in each result. The matchinfo function is much more efficient than the offsets function. Furthermore, the matchinfo function provides extra information regarding the overall number of occurrences of each query term in the entire document set (not just the current row) and the number of documents in which each query term appears. This may be used (for example) to attach a higher weight to less common terms which may increase the overall computed relevancy of those results the user considers more interesting.

-- If the application supplies an SQLite user function called "rank" that
-- interprets the blob of data returned by matchinfo and returns a numeric
-- relevancy based on it, then the following SQL may be used to return the
-- titles of the 10 most relevant documents in the dataset for a users query.
SELECT title FROM documents 
  WHERE documents MATCH <query>
  ORDER BY rank(matchinfo(document)) DESC
  OFFSET 0 LIMIT 10

The SQL query in the example above uses less CPU than the first example in this section, but still has a non-obvious performance problem. SQLite satisfies this query by retrieving the value of the "title" column and matchinfo data from the FTS3 module for every row matched by the users query before it sorts and limits the results. Because of the way SQLite's virtual table interface works, retrieving the value of the "title" column requires loading the entire row from disk (including the "content" field, which may be quite large). This means that if the users query matches several thousand documents, many megabytes of "title" and "content" data may be loaded from disk into memory even though they will never be used for any purpose.

The SQL query in the following example block is one solution to this problem. In SQLite, when a sub-query used in a join contains a LIMIT clause, the results of the sub-query are calculated and stored in temporary table before the main query is executed. This means that SQLite will load only the docid and matchinfo data for each row matching the users query into memory, determine the docid values corresponding to the ten most relevant documents, then load only the title and content information for those 10 documents only. Because both the matchinfo and docid values are gleaned entirely from the full-text index, this results in dramatically less data being loaded from the database into memory.

SELECT title FROM documents JOIN ( 
    SELECT docid, rank(matchinfo(document)) AS rank 
    FROM documents
    WHERE documents MATCH <query>
    ORDER BY rank DESC 
    OFFSET 0 LIMIT 10
) AS ranktable USING(docid)
ORDER BY ranktable.rank DESC

The next block of SQL enhances the query with solutions to two other problems that may arise in developing search applications using FTS3:

1. The snippet function cannot be used with the above query. Because the outer query does not include a "WHERE ... MATCH" clause, the snippet function may not be used with it. One solution is to duplicate the WHERE clause used by the sub-query in the outer query. The overhead associated with this is usually negligible.

2. The relevancy of a document may depend on something other than just the data available in the return value of matchinfo. For example each document in the database may be assigned a static weight based on factors unrelated to its content (origin, author, age, number of references etc.). These values can be stored by the application in a separate table that can be joined against the documents table in the sub-query so that the rank function may access them.

This version of the query is very similar to that used by the sqlite.org documentation search application.

-- This table stores the static weight assigned to each document in FTS3 table
-- "documents". For each row in the documents table there is a corresponding row
-- with the same docid value in this table.
CREATE TABLE documents_data(docid INTEGER PRIMARY KEY, weight);

-- This query is similar to the one in the block above, except that:
--
--   1. It returns a "snippet" of text along with the document title for display. So
--      that the snippet function may be used, the "WHERE ... MATCH ..." clause from
--      the sub-query is duplicated in the outer query.
--
--   2. The sub-query joins the documents table with the document_data table, so that
--      implementation of the rank function has access to the static weight assigned
--      to each document.
SELECT title, snippet(documents) FROM documents JOIN ( 
    SELECT docid, rank(matchinfo(document), documents_data.weight) AS rank
    FROM documents JOIN documents_data USING(docid)
    WHERE documents MATCH <query>
    ORDER BY rank DESC 
    OFFSET 0 LIMIT 10
) AS ranktable USING(docid)
WHERE documents MATCH <query>
ORDER BY ranktable.rank DESC

All the example queries above return the ten most relevant query results. By modifying the values used with the OFFSET and LIMIT clauses, a query to return (say) the next ten most relevant results is easy to construct. This may be used to obtain the data required for a search applications second and subsequent pages of results.

The next block contains an example rank function that uses matchinfo data implemented in C. Instead of a single weight, it allows a weight to be externally assigned to each column of each document. It may be registered with SQLite like any other user function using sqlite3_create_function.

/*
** SQLite user defined function to use with matchinfo() to calculate the
** relevancy of an FTS3 match. The value returned is the relevancy score
** (a real value greater than or equal to zero). A larger value indicates 
** a more relevant document.
**
** The overall relevancy returned is the sum of the relevancies of each 
** column value in the FTS3 table. The relevancy of a column value is the
** sum of the following for each reportable phrase in the FTS3 query:
**
**   (<hit count> / <global hit count>) * <column weight>
**
** where <hit count> is the number of instances of the phrase in the
** column value of the current row and <global hit count> is the number
** of instances of the phrase in the same column of all rows in the FTS3
** table. The <column weight> is a weighting factor assigned to each
** column by the caller (see below).
**
** The first argument to this function must be the return value of the FTS3 
** matchinfo() function. Following this must be one argument for each column 
** of the FTS3 table containing a numeric weight factor for the corresponding 
** column. Example:
**
**     CREATE VIRTUAL TABLE documents USING fts3(title, content)
**
** The following query returns the docids of documents that match the full-text
** query <query> sorted from most to least relevant. When calculating
** relevance, query term instances in the 'title' column are given twice the
** weighting of those in the 'content' column.
**
**     SELECT docid FROM documents 
**     WHERE documents MATCH <query> 
**     ORDER BY rank(matchinfo(documents), 1.0, 0.5) DESC
*/
static void rankfunc(sqlite3_context *pCtx, int nVal, sqlite3_value **apVal){
  int *aMatchinfo;                /* Return value of matchinfo() */
  int nCol;                       /* Number of columns in the table */
  int nPhrase;                    /* Number of phrases in the query */
  int iPhrase;                    /* Current phrase */
  double score = 0.0;             /* Value to return */

  assert( sizeof(int)==4 );

  /* Check that the number of arguments passed to this function is correct.
  ** If not, jump to wrong_number_args. Set aMatchinfo to point to the array
  ** of unsigned integer values returned by FTS3 function matchinfo. Set
  ** nPhrase to contain the number of reportable phrases in the users full-text
  ** query, and nCol to the number of columns in the table.
  */
  if( nVal<1 ) goto wrong_number_args;
  aMatchinfo = (unsigned int *)sqlite3_value_blob(apVal[0]);
  nPhrase = aMatchinfo[0];
  nCol = aMatchinfo[1];
  if( nVal!=(1+nCol) ) goto wrong_number_args;

  /* Iterate through each phrase in the users query. */
  for(iPhrase=0; iPhrase<nPhrase; iPhrase++){
    int iCol;                     /* Current column */

    /* Now iterate through each column in the users query. For each column,
    ** increment the relevancy score by:
    **
    **   (<hit count> / <global hit count>) * <column weight>
    **
    ** aPhraseinfo[] points to the start of the data for phrase iPhrase. So
    ** the hit count and global hit counts for each column are found in 
    ** aPhraseinfo[iCol*3] and aPhraseinfo[iCol*3+1], respectively.
    */
    int *aPhraseinfo = &aMatchinfo[2 + iPhrase*nCol*3];
    for(iCol=0; iCol<nCol; iCol++){
      int nHitCount = aPhraseinfo[3*iCol];
      int nGlobalHitCount = aPhraseinfo[3*iCol+1];
      double weight = sqlite3_value_double(apVal[iCol+1]);
      if( nHitCount>0 ){
        score += ((double)nHitCount / (double)nGlobalHitCount) * weight;
      }
    }
  }

  sqlite3_result_double(pCtx, score);
  return;

  /* Jump here if the wrong number of arguments are passed to this function */
wrong_number_args:
  sqlite3_result_error(pCtx, "wrong number of arguments to function rank()", -1);
}

*** DRAFT ***