The classifier can be configured and customized to do things like:

  • automatically delete torrents you don’t want in your index
  • add custom tags to torrents you’re interested in
  • customize the keywords and file extensions used for determining a torrent’s content type
  • specify completely custom logic to classify and perform other actions on torrents

Skip to practical use cases and examples


After a torrent is crawled or imported, some further processing must be done to gather metadata, have a guess at the torrent’s contents and finally index it in the database, allowing it to be searched and displayed in the UI/API.

bitmagnet’s classifier is powered by a Domain Specific Language. The aim of this is to provide a high level of customisability, along with transparency into the classification process which will hopefully aid collaboration on improvements to the core classifier logic.

The classifier is declared in YAML format. The application includes a core classifier that can be configured, extended or completely replaced with a custom classifier. This page documents the required format.

Source precedence

bitmagnet will attempt to load classifier source code from all the following locations. Any discovered classifier source will be merged with other sources in the following order of precedence:

Note that multiple sources will be merged, not replaced. For example, keywords added to the classifier configuration will be merged with the core keywords.

The merged classifier source can be viewed with the CLI command bitmagnet classifier show.

See Running the CLI


A JSON schema for the classifier is available; some editors and IDEs will be able to validate the structure of your classifier document by specifying the $schema attribute:


The classifier schema can also be viewed by running the cli command bitmagnet classifier schema.

See Running the CLI

The classifier declaration comprises the following components:


A workflow is a list of actions that will be executed on all torrents when they are classified. When no custom configuration is provided, the default workflow will be run. To use a different workflow instead, specify the classifier.workflow configuration option with the name of your custom workflow.


An action is a piece of workflow to be executed. All actions either return an updated classification result or an error.

For example, the following action will set the content type of the current torrent to audiobook:

set_content_type: audiobook

The following action will return an unmatched error:


And the following action will delete the current torrent being classified (returning a delete error):


These actions aren’t much use on their own - we’d want to check some conditions are satisfied before setting a content type or deleting a torrent, and for this we’d use the if_else action. For example, the following action will set the content type to audiobook if the torrent name contains audiobook-related keywords, and will otherwise return an unmatched error:

  condition: "torrent.baseName.matches(keywords.audiobook)"
    set_content_type: audiobook
  else_action: unmatched

The following action will delete a torrent if its name matches the list ofbanned keywords:

  condition: "torrent.baseName.matches(keywords.banned)"
  if_action: delete

Actions may return the following types of error:

  • An unmatched error indicates that the current action did not match for the current torrent
  • A delete error indicates that the torrent should be deleted
  • An unhandled error may occur, for example if the TMDB API was unreachable

Whenever an error is returned, the current classification will be terminated.

Note that a workflow should never return an unmatched error. We expect to iterate through a series of checks corresponding to each content type. If the current torrent does not match the content type being checked, we’ll proceed to the next check until we find a match; if no match can be found, the content type will be unknown. To facilitate this, we can use the find_match action.

The find_match action is a bit like a try/catch block in some programming languages; it will try to match a particular content type, and if an unmatched error is returned, it will catch the unmatched error proceed to the next check. For example, the following action will attempt to classify a torrent as an audiobook, and then as an ebook. If both checks fail, the content type will be unknown:

  # match audiobooks:
  - if_else:
      condition: "torrent.baseName.matches(keywords.audiobook)"
        set_content_type: audiobook
      else_action: unmatched
  # match ebooks:
  - if_else:
      condition: ", f.extension in extensions.ebook ? f.size : - f.size).sum() > 0"
        set_content_type: ebook
      else_action: unmatched

For a full list of available actions, please refer to the JSON schema.


Conditions are used in conjunction with the if_else action, in order to execute an action if a particular condition is satisfied.

The conditions in the examples above use CEL (Common Expression Language) expressions.

The CEL environment

CEL is already a well-documented language, so this page won’t go into detail about the CEL syntax. In the context of the bitmagnet classifier, the CEL environment exposes a number of variables:

  • torrent: The current torrent being classified (protobuf type: bitmagnet.Torrent)
  • result: The current classification result (protobuf type: bitmagnet.Classification)
  • keywords: A map of strings to regular expressions, representing named lists of keywords
  • extensions: A map of strings to string lists, representing named lists of extensions
  • contentType: A map of strings to enum values representing content types (e.g.,
  • fileType: A map of strings to enum values representing file types (e.g.,
  • flags: A map of strings to the configured values of flags
  • kb, mb, gb: Variables defined for convenience, equal to the number of bytes in a kilobyte, megabyte and gigabyte respectively

For more details on the protocol buffer types, please refer to the protobuf schema.

Boolean logic (or, and & not)

In addition to CEL expressions, conditions may be declared using the boolean logic operators or, and and not. For example, the following condition evaluates to true, if either the torrent consists mostly of file extensions very commonly used for music (e.g. flac), OR if the torrent both has a name that includes music-related keywords, and consists mostly of audio files:

  - ", f.extension in ? f.size : - f.size).sum() > 0"
  - and:
      - "torrent.baseName.matches("
      - ", f.fileType == ? f.size : - f.size).sum() > 0"

Note that we could also have specified the above condition using just one CEL expression, but breaking up complex conditions like this is more readable.


The classifier includes lists of keywords associated with different types of torrents. These aim to provide a simpler alternative to regular expressions, and the classifier will compile all keyword lists to regular expressions that can be used within CEL expressions. In order for a keyword to match, it must appear as an isolated token in the test string - that is, it must be either at the beginning or preceded by a non-word character, and either at the end or followed by a non-word character.

Reserved characters in the syntax are:

  • parentheses ( and ) enclose a group
  • | is an OR operator
  • * is a wildcard operator
  • ? makes the previous character or group optional
  • + specifies one or more of the previous character
  • # specifies any number
  • ` ` specifies any non-word or non-number character

For example, to define some music- and audiobook-related keywords:

  music: # define music-related keywords
    - music # all letters are case-insensitive, and must be defined in lowercase unless escaped
    - discography
    - album
    - \V.?\A # escaped letters are case-sensitive; matches "VA", "V.A" and "V.A.", but not "va"
    - various artists # matches "various artists" and "Various.Artists"
  audiobook: # define audiobook-related keywords
    - (audio)?books?
    - (un)?abridged
    - narrated
    - novels?
    - (auto)?biograph(y|ies) # matches "biography", "autobiographies" etc.

If you’d rather use plain old regular expressions, the CEL syntax supports that too, for example torrent.baseName.matches("^myregex$").


The classifier includes lists of file extensions associated with different types of content. For example, to identify torrents of type comic by their file extensions, the extensions are first declared:

    - cb7
    - cba
    - cbr
    - cbt
    - cbz

The extensions can now be used as part of a condition within an if_else action:

  condition: ", f.extension in extensions.comic ? f.size : - f.size).sum() > 0"
    set_content_type: comic
  else_action: unmatched


Flags can be used to configure workflows. In order to use a flag in a workflow, it must first be defined. For example, the core classifier defines the following flags that are used in the default workflow:

  tmdb_enabled: bool
  delete_content_types: content_type_list
  delete_xxx: bool

These flags can be referenced within CEL expressions, for example to delete adult content if the delete_xxx flag is set to true:

  condition: "flags.delete_xxx && result.contentType =="
  if_action: delete


The classifier can be customized by providing a classifier.yml file in a supported location as described above. If you only want to make some minor modifications, it may be convenient to specify these using the main application configuration instead, by providing values in either config.yml or as environment variables. The application configuration exposes some but not all properties of the classifier.

For example, in your config.yml you could specify:

  # specify a custom workflow to be used:
  workflow: custom
  # add to the core list of music keywords:
      - my-custom-music-keyword
  # add a file extension to the list of audiobook-related extensions:
      - abc
  # auto-delete all comics
      - comics

Or as environment variables you could specify:

TMDB_ENABLED=false \ # disable the TMDB API integration
  CLASSIFIER_WORKFLOW=custom \ # specify a custom workflow to be used
  CLASSIFIER_DELETE_XXX=true \ # auto-delete all adult content
  bitmagnet worker run --all


The classifier source is compiled on initial load, and all structural and syntax errors should be caught at compile time. If there are errors in your classifier source, bitmagnet should exit with an error message indicating the location of the problem.

Testing on individual torrents

You can test the classifier on an individual torrent or torrents using the bitmagnet process CLI command:

bitmagnet process --infoHash=aaaaaaaaaaaaaaaaaaaa --infoHash=bbbbbbbbbbbbbbbbbbbb

See Running the CLI

Reclassify all torrents

Read how to reclassify all torrents.

Practical use cases and examples

Auto-delete specific content types

The default workflow provides a flag that allows for automatically deleting specific content types. For example, to delete all comic, software and xxx torrents:

    - comic
    - software
    - xxx

Auto-deleting adult content has been one of the most requested features. For convenience, this is exposed as the configuration option classifier.delete_xxx, and can be specified with the environment variable CLASSIFIER_DELETE_XXX=true.

Auto-delete torrents containing specific keywords

Any torrents containing keywords in the banned list will be automatically deleted. This is primarily used for deleting CSAM content, but the list can be extended to auto-delete any other keywords:

    - my-hated-keyword

Disable the TMDB API integration

The tmdb_enabled flag can be used to disable the TMDB API integration:

  tmdb_enabled: false

For convenience, this is also exposed as the configuration option tmdb.enabled, and can be specified with the environment variable TMDB_ENABLED=false.

The apis_enabled flag has the same effect, disabling TMDB and any future API integrations:

  apis_enabled: false

API integrations can also be disabled for individual classifier runs, without disabling them globally, by passing the --apisDisabled flag to the reprocess command.

Extend the default workflow with custom logic

Custom workflows can be added in the workflows section of the classifier document. It is possible to extend the default workflow by using the run_workflow action within your custom workflow, for example:

    - <my custom action to be executed before the default workflow>
    - run_workflow: default
    - <my custom action to be executed after the default workflow>

A concrete example of this is adding tags to torrents based on custom criteria.

Use tags to create custom torrent categories

Is there a category of torrent you’re interested in that isn’t captured by one of the core content types? Torrent tags are intended to capture custom categories and content types.

Let’s imagine you’d like to surface torrents containing interesting documents. The interesting documents have specific file extensions, and their filenames contain specific keywords. Let’s create a custom action to tag torrents containing interesting documents:

# define file extensions for the documents we're interested in:
    - doc
    - docx
    - pdf
# define keywords that must be present in the filenames of the interesting documents:
    - interesting
    - fascinating
# extend the default workflow with a custom workflow to tag torrents containing interesting documents:
    # first run the default workflow:
    - run_workflow: default
    # then add the tag to any torrents containing interesting documents:
    - if_else:
        condition: "torrent.files.filter(f, f.extension in extensions.interesting_documents && f.basePath.matches(keywords.interesting_documents)).size() > 0"
          add_tag: interesting-documents

To specify that the custom workflow should be used, remember to specify the classifier.workflow configuration option, e.g. CLASSIFIER_WORKFLOW=custom bitmagnet worker run --all.