FTS: Full Text Search (fts) Plugin

As the amount and importance of information stored in email messages is increasing in people’s everyday lives, searching through those messages is becoming ever more important. At the same time mobile clients add their own restrictions for what can be done on the client side. The ever diversifying mail client software also tests the limits of the IMAP protocol and current server implementations.

Furthermore, the IMAP protocol requires some rather complicated and expensive searching capabilities. For example, the protocol requires arbitrary substring matching. Some newer mobile clients (e.g. Apple iOS) rely on this functionality.

Without a high-performance index, Dovecot must fall back to a slow sequential search through all messages (default behavior). If storage latencies are high, this searching may not be completed in a reasonable time, or resource utilization may be too large, especially in mailboxes with large messages.

Dovecot maintains these FTS indexing engines:

Name	Description
Dovecot Pro FTS	Dovecot native, object storage optimized driver. Only available as part of Dovecot Pro.
fts_solr plugin	Interface to Apache Solr; stores data remotely.
fts_flatcurve plugin	Xapian based driver; stores data locally.

Searching In Dovecot

When a FTS indexing backend is not present, searches use a slow sequential search through all message data. This is both computationally and time expensive. It is desirable to pre-index data so that searches can be executed against this index.

There is a subtle but important distinction between searching through message headers and searching through message bodies.

Searching through message bodies (via the standard IMAP 'SEARCH TEXT/BODY' commands) makes use of the FTS indexes.

On the other hand, searching through message headers benefits from Dovecot's standard index and cache files (dovecot.index and dovecot.index.cache), which often contain the necessary information. It is possible to redirect header searches to FTS indexes via a configuration option (fts_enforced).

Triggers for FTS indexing are configurable. It can be started on demand when searching, or automatically when new messages arrive or as a batch job.

By default the FTS indexes are updated only while searching, so neither LDA/LMTP nor an IMAP 'APPEND' command updates the indexes immediately. This means that if a user has received a lot of mail since the last indexing (i.e., the last search operation), it may take a while to index all the new mails before replying to the search command. Dovecot sends periodic “* OK Indexed n% of the mailbox” updates which can be caught by client implementations to implement a progress bar.

Updating the FTS index as messages arrive makes for a more responsive user experience, especially for users who don’t search often, but have a lot of mail. On the other hand, it increases overall system load regardless of whether or not the indexes will ever be used by the user.

Dovecot FTS Architecture

Dovecot splits the full text search functionality into two parts: a common tokenization library (lib-fts) and backend indexing engine responsible for storing the tokens produced by the common library persistently.

Some of the FTS backends do their own internal tokenization, although it's possible to configure them to use the lib-fts tokenization as well.

See Tokenization for more details about configuring the tokenization.

All backends are implemented as plugins that extend the base fts plugin's functionality.

Settings

fts_autoindex

Default	no
Value	boolean
See Also	fts_autoindex_exclude fts_autoindex_max_recent_msgs

If enabled, index mail as it is delivered or appended.

fts_autoindex_exclude

Default	[None]
Value	string
See Also	fts_autoindex
Changes	Changed: 2.4.0 This setting now honors namespaces for mailbox names.

To exclude a mailbox from automatic indexing, it can be listed in this setting.

To exclude additional mailboxes, add sequential numbers to the end of the plugin name.

Use either mailbox names or special-use flags (e.g. Trash).

For example:

plugin {
  fts_autoindex_exclude = Junk
  fts_autoindex_exclude2 = Trash
  fts_autoindex_exclude3 = External Accounts/*
}

This setting matches also the namespace prefix in folder names.

Namespaces match as follows:

• The full folder name, including the namespace prefix.

  For example fts_autoindex_exclude = Public/incoming would match the incoming folder in the Public/ namespace.

• For inbox=yes namespace, the folder name without the namespace prefix.

  For example fts_autoindex_exclude = incoming would match the incoming folder in the INBOX namespace, but not in the Public/ namespace.

• The folder names support * and ? wildcards.

  Namespace prefixes must NOT be specified and will not match for:

  • the INBOX folder
  • special-use flags (e.g. Trash)

fts_autoindex_max_recent_msgs

Default	[None]
Value	unsigned integer
See Also	fts_autoindex

To exclude infrequently accessed mailboxes from automatic indexing, set this value to the maximum number of Recent flagged messages that exist in the mailbox.

A value of 0 means to ignore this setting.

Mailboxes with more flagged Recent messages than this value will not be autoindexed, even though they get deliveries or appends. This is useful for, e.g., inactive Junk folders.

Any folders excluded from automatic indexing will still be indexed, if a search on them is performed.

Example:

plugin {
  fts_autoindex_max_recent_msgs = 999
}

fts_decoder

Default	[None]
Value	string

Decode attachments to plaintext using this service and index the resulting plaintext.

See the decode2text.sh script included in Dovecot for how to use this.

Example:

plugin {
  fts_decoder = decode2text
}

service decode2text {
  executable = script /usr/lib/dovecot/decode2text.sh
  user = vmail
  unix_listener decode2text {
	mode = 0666
  }
}

This setting and fts_tika cannot be used simultaneously.

fts_enforced

Default	no
Value	string
Allowed Values	yesnobody

Require FTS indexes to perform a search? This controls what to do when searching headers and what to do on error situations.

When searching from message body, the FTS index is always (attempted to be) updated to contain any missing mails before the search is performed.

no

Searching from message headers won't update FTS indexes. For header searches, the FTS indexes are used for searching the mails that are already in it, but the unindexed mails are searched via dovecot.index.cache (or by opening the emails if the headers aren't in cache).

If FTS lookup or indexing fails, both header and body searches fallback to searching without FTS (i.e. possibly opening all emails). This may timeout for large mailboxes and/or slow storage.

yes

Searching from message headers updates FTS indexes, the same way as searching from body does. If FTS lookup or indexing fails, the search fails.

body

Searching from message headers won't update FTS indexes (the same behavior as with no). If FTS lookup or indexing fails, the search fails.

Only the yes value guarantees consistent search results. In other cases it's possible that the search results will be different depending on whether the search was performed via FTS index or not.

fts_filters

Default	no
Value	string
See Also	FTS Tokenization

The list of filters to apply.

Language specific filter chains can be specified with fts_filters_<lang> (e.g. fts_filters_en).

See filters for configuration information.

fts_header_excludes

Default	[None]
Value	string

The list of headers to, respectively, include or exclude.

• The default is the preexisting behavior, i.e. index all headers.
• includes take precedence over excludes: if a header matches both, it is indexed.
• The terms are case insensitive.
• An asterisk * at the end of a header name matches anything starting with that header name.
• The asterisk can only be used at the end of the header name. Prefix and infix usage of asterisk are not supported.

Example

plugin {
  fts_header_excludes = Received DKIM-* X-* Comments
  fts_header_includes = X-Spam-Status Comments
}

• Received headers, all DKIM- headers and all X- experimental headers are excluded, with the following exceptions:

  • Comments and X-Spam-Status are indexed anyway, as they match both excludes and includes lists.
  • All other headers are indexed.

Example

plugin {
  fts_header_excludes = *
  fts_header_includes = From To Cc Bcc Subject Message-ID In-* X-CustomApp-*
}

• No headers are indexed, except those specified in the includes.

fts_header_includes

Default	[None]
Value	string
See Also	fts_header_excludes

fts_index_timeout

Default	[None]
Value	unsigned integer

When the full text search backend detects that the index isn't up-to-date, the indexer is told to index the messages and is given this much time to do so. If this time limit is reached, an error is returned, indicating that the search timed out during waiting for the indexing to complete: NO [INUSE] Timeout while waiting for indexing to finish. Note the fts_enforced setting does not change this behavior.

A value of 0 means no timeout.

fts_language_config

Default	<textcat dir>
Value	string
See Also	fts_languages

Path to the textcat/exttextcat configuration file, which lists the supported languages.

This is recommended to be changed to point to a minimal version of a configuration that supports only the languages listed in fts_languages.

Doing this improves language detection performance during indexing and also makes the detection more accurate.

Example:

plugin {
  fts_language_config = /usr/share/libexttextcat/fpdb.conf
}

fts_languages

Default	<textcat dir>
Value	string
See Also	fts_language_config

A space-separated list of languages that the full text search should detect.

At least one language must be specified.

The language listed first is the default and is used when language recognition fails.

The filters used for stemming and stopwords are language dependent.

Note: For better performance it's recommended to synchronize this setting with the textcat configuration file; see fts_language_config.

Example:

plugin {
  fts_languages = en de
}

fts_message_max_size

Default	[None]
Value	size
Changes	Added: 2.4.0

Maximum body size that is processed by fts. 0 means unlimited.

fts_tika

Default	[None]
Value	string
Changes	Changed: 2.4.0 Basic authentication support (via URL) is added.

URL for Apache Tika decoder for attachments.

Example:

plugin {
  fts_tika = http://tikahost:9998/tika/
}

This setting and fts_decoder cannot be used simultaneously.

fts_tokenizers

Default	generic email-address
Value	string
See Also	FTS Tokenization

The list of tokenizers to use.

This setting can be overridden for specific languages by using fts_tokenizers_<lang> (e.g. fts_tokenizers_en).

See tokenizers for configuration information.

Configuration

FTS Indexing Triggers

Missing mails are always added to FTS indexes when using IMAP SEARCH command that attempts to access the FTS indexes.

Automatic FTS indexing can also be done during mail delivery, IMAP APPEND and other ways of adding mails to mailboxes using fts_autoindex.

Indexing can also be triggered manually:

$ doveadm index -u user@domain -q INBOX

Enforce FTS

When FTS indexing fails, Dovecot falls back on using the built-in search, which does not have indexes for mail bodies.

This could end up opening all the mails in the mailbox, which often isn't wanted.

To disable this functionality, enable fts_enforced.

Indexing Attachments

Attachments can be indexed either via a script that translates the attachment to UTF-8 plaintext or Apache Tika server.

• fts_decoder: Use decoder script.
• fts_tika: Use Apache Tika server.

Note

fts_decoder and fts_tika cannot be used simultaneously.

Rescan

Dovecot keeps track of indexed messages in the dovecot.index files. If this becomes out of sync with the actual FTS indexes (either too many or too few mails), you'll need to do a rescan and then index missing mails:

$ doveadm fts rescan -u user@domain
$ doveadm index -u user@domain -q '*'

Note that currently most FTS backends don't implement the rescan. Instead, they simply delete all the FTS indexes. This may change in the future versions.

Languages

Language names are given as ISO 639-1 alpha 2 codes.

Stemming support indicates whether the snowball filter can be used.

Stopwords support indicates whether a stopwords file is distributed with Dovecot.

Currently supported languages:

Language Code	Language	Stemming	Stopwords
da	Danish	Yes	Yes
de	German	Yes	Yes
en	English	Yes	Yes
es	Spanish	Yes	Yes
fi	Finnish	Yes	Yes
fr	French	Yes	Yes
it	Italian	Yes	Yes
ja	Japanese (Requires Dovecot Pro)	No	No
nl	Dutch	Yes	Yes
no	Norwegian (Bokmal & Nynorsk detected)	Yes	Yes
pt	Portuguese	Yes	Yes
ro	Romanian	Yes	Yes
ru	Russian	Yes	Yes
sv	Swedish	Yes	Yes
tr	Turkish	Yes	Yes

Tokenization

Dovecot contains tokenization support that can be used by FTS drivers.

The lib-fts tokenization library works in the following way:

1. Language detection: When indexing, the text language is attempted to be detected.

   If the detection fails, the first listed language is used.

   When searching, the search is done using all the configured languages.

2. Tokenization: The text is split to tokens (individual words).

   • Whitespace and other nonindexable characters are dropped.
   • Base64 sequences are looked for and skipped.

3. Filtering: Tokens are normalized:

   • Normalization / lowercasing
   • Stemming

4. Stopwords: A configurable list of words not to be indexed

Languages

fts_languages lists languages FTS should detect.

At least one language must be listed.

The first language is the default language used in case detection fails.

Each added language makes the indexing and searching slightly slower, so it's recommended not to add too many languages unnecessarily.

The language detection performance can be improved by limiting the number of languages available for textcat, see fts_language_config.

Example:

plugin {
  fts_languages = en de
}

Filter and Tokenizer Order

The filters and tokenizers are created in the order they are declared in their respective settings in the configuration file. They form a chain, where the first filter or tokenizer is the parent or grandparent of the rest. The direction of the data flow needs some special attention.

In filters, the data flows from parent to child, so tokens are first passed to the grandparent of all filters and then further down the chain. For some filtering chains the order is important. E.g. the snowball stemmer wants all input in lower case, so the filter lower casing the tokens will need to be listed before it.

In tokenizers however, the data however flows from child to parent. This means that the tokenizer listed 'last' gets the processed data 'first'.

So, for filters data flows “left to right” through the filters listed in the configuration. In tokenizers the order is “right to left”.

Base64 Detection

Base64 sequences are looked for in the tokenization buffer and skipped when detected.

A base64 sequence is detected by:

• An optional leader character comprised in leader-characters set,
• A run of characters, all comprised in the base64-characters set, at least minimum-run-length long,
• An end-of-buffer, or a trailer character comprised in trailer-characters set,

where:

• leader-characters are: [ \t\r\n=:;?]
• base64-characters are: [0-9A-Za-z/+]
• trailer-characters are: [ \t\r\n=:;?]
• minimum-run-length is: 50
• minimum-run-count is: 1

Thus, (even single) 50-chars runs of characters in the base64 set are recognized as base64 and ignored in indexing.

If a base64 sequence happens to be split across different chunks of data, part of it might not be detected as base64. In this case, the undetected base64 fragment is still indexed. However, this happens rarely enough that it does not significantly impact the quality of the filter.

So far the above rule seems to give good results in base64 indexing avoidance. It also performs well in removing base64 fragments inside headers, like ARC-Seal, DKIM-Signature, X-SG-EID, X-SG-ID, including header-encoded parts (e.g. =?us-ascii?Q?...?= sequences).

Filter Configuration

Filters affect how data is indexed.

They are configured through fts_filters.

Example:

plugin {
  fts_filters = normalizer-icu snowball stopwords
  fts_filters_en = lowercase snowball english-possessive stopwords
}

Available filters:

lowercase

Change all text to lower case. Supports UTF8, when compiled with libicu and the library is installed. Otherwise only ASCII characters are lowercased.

stopwords

Filter certain common and short words, which are usually useless for searching.

WARNING

Using stopwords with multiple languages configured WILL cause some searches to fail. The recommended solution is to NOT use the stopword filter when multiple languages are present in the configuration.

Stopwords specific settings:

stopwords_dir

• Default: <DATADIR>/stopwords

Path to the directory containing stopword files. Stopword files are looked up in ”<path>”/stopwords_<lang>.txt.

See languages for the list of stopword files that are currently distributed with Dovecot.

More languages can be obtained from Apache Lucene, Snowball stemmer, or https://github.com/stopwords-iso/.

snowball

Stemming tries to convert words to a common base form. A simple example is converting “cars” to “car” (in English).

This stemmer is based on the Snowball stemmer library.

normalizer-icu

Normalize text using libicu. This is potentially very resource intensive.

Caveat for Norwegian

The default normalizer filter does not modify U+00F8 (Latin Small Letter O with Stroke). In some configurations it might be desirable to rewrite it to, e.g., o. Same goes for the upper case version. This can be done by passing a modified id setting to the normalizer filter.

Similar cases can exist for other languages as well.

normalizer-icu specific settings:

stopwords_dir

• Default: Any-Lower; NFKD; [: Nonspacing Mark :] Remove; [\\x20] Remove

Description of the normalizing/transliterating rules to use.

See Normalizer Format for syntax.

english-possessive

Remove trailing 's from English possessive form tokens. Any trailing single ' characters are already removed by tokenizing, whether this filter is used or not.

The snowball filter also removes possessive suffixes from English, so if using snowball this filter is not needed.

TIP

snowball likely produces better results, so this filter is advisable only when snowball is not available or cannot be used due to extreme CPU performance requirements.

contractions

Removes certain contractions that can prefix words. The idea is to only index the part of the token that conveys the core meaning.

Only works with French, so the language of the input needs to be recognized by textcat as French.

It filters “qu'”, “c'”, “d'”, “l'”, “m'”, “n'”, “s'” and “t'”.

Do not use at the same time as generic tokenizer with algorithm=tr29 wb5a=yes.

Tokenizer Configuration

Tokenizers affect how input data is parsed.

Available tokenizers:

generic

Input data, such as email text and headers, need to be divided into words suitable for indexing and searching. The generic tokenizer does this.

Generic-specific settings:

maxlen

• Default: 30

Maximum length of token, before an arbitrary cut off is made.

algorithm

• Default: simple

Accepted values are simple or tr29. It defines the method for looking for word boundaries.

simple is faster and will work for many texts, especially those using latin alphabets, but leaves corner cases.

tr29 implements a version of Unicode technical report 29 word boundary lookup. It might work better with, e.g., texts containing Katakana or Hebrew characters, but it is not possible to use a single algorithm for all existing languages.

wb5a

• Default: no

Unicode TR29 rule WB5a setting to the tr29 tokenizer. Splits prefixing contracted words from base word. E.g. “l'homme” → “l” “homme”. Together with a language specific stopword list unnecessary contractions can thus be filtered away. This is disabled by default and only works with the TR29 algorithm.

Enable by fts_tokenizer_generic = algorithm=tr29 wb5a=yes.

email-address

This tokenizer preserves email addresses as complete search tokens, by bypassing the generic tokenizer, when it finds an address. It will only work as intended if it is listed after other tokenizers.

Email Address tokenizer specific settings:

maxlen

• Default: 254

Maximum length of token, before an arbitrary cut off is made.