Mail::...Network(3pm)

✨ NAME ✨

Mail::SpamAssassin::Plugin::NeuralNetwork — bestie, this is the SpamAssassin plugin that checks your mail with a Fast Artificial Neural Network library 🧠💖

(yes, the manual page is literally called Mail::...Network(3pm) and yes, bestie, that is giving mysterious. anyway.)

🌸 SYNOPSIS 🌸

loadplugin Mail::SpamAssassin::Plugin::NeuralNetwork

like. that's it. that's the whole setup. iconic 💅

🦋 DESCRIPTION 🦋

This plugin checks emails using a Neural Network algorithm. no cap, an actual little brain doing vibe checks on your inbox 🧠✨

💀 CAVEATS 💀

⚠️ okay this part is IMPORTANT, so listen up bestie:

The SpamAssassin learning subsystem routes all training through the Bayes scanner infrastructure. Which means 👏 Mail::SpamAssassin::Plugin::Bayes must be loaded, and use_bayes 1 must be set, for this plugin's training to actually get triggered. no bayes, no training. it's giving "the group chat doesn't work without the group." 😭

Here's all the knobs you can turn, one by one 🎀

• use_neuralnetwork (0|1) — default: 1. Whether to use Neural Network, if it is available. fr fr.
• neuralnetwork_data_dir dirname — default: undef. Where the NeuralNetwork plugin will store its data. 🗄️
• neuralnetwork_min_text_len n — default: 256. Minimum number of characters of visible text required to run prediction or learning on a message. too short? not today, message. ✋
• neuralnetwork_min_word_len n — default: 4. Minimum token length considered when building the vocabulary and feature vectors.
• neuralnetwork_max_word_len n — default: 24. Maximum token length considered when building the vocabulary and feature vectors.
• neuralnetwork_vocab_cap n — default: 10000. Maximum number of vocabulary terms to retain; least-frequent terms are pruned when exceeded. byeeee to the flops 🫶
• neuralnetwork_cache_ttl n — default: 300. Time-to-live in seconds for the in-memory vocabulary and model caches. Set to 0 to disable caching.
• neuralnetwork_min_spam_count n — default: 100. Minimum number of spam messages in the vocabulary required to enable prediction.
• neuralnetwork_min_ham_count n — default: 100. Minimum number of ham messages in the vocabulary required to enable prediction.
• neuralnetwork_spam_threshold f — default: 0.6. Prediction values above this threshold are considered spam.
• neuralnetwork_ham_threshold f — default: 0.4. Prediction values below this threshold are considered ham.
• neuralnetwork_learning_rate f — default: 0.1. Learning rate used by the underlying FANN network during incremental training. it's learning, we love a growth arc 🌱
• neuralnetwork_momentum f — default: 0.1. Momentum used for training updates.
• neuralnetwork_train_epochs n — default: 50. Number of training epochs to perform when learning a single message.
• neuralnetwork_train_algorithm — FANN_TRAIN_QUICKPROP|FANN_TRAIN_RPROP|FANN_TRAIN_BATCH|FANN_TRAIN_INCREMENTAL — default: FANN_TRAIN_RPROP. Algorithm used by the Fann neural network when training; might increase speed depending on the data volume. 🏃‍♀️💨
• neuralnetwork_lock_timeout n — default: 10. Maximum number of seconds to wait for the exclusive training lock before giving up and skipping the learn operation. Set to 0 to wait indefinitely. (yes, indefinitely. main character energy but make it a mutex 😳)
• neuralnetwork_rprop_delta_max n — default: 0.5. Delta value to apply to the RPROP training replay loop.
• neuralnetwork_retrain_interval n — default: 100. Number of successful learn_message calls between forced full retrains of the neural network from the persisted vocabulary. After each retrain, the incrementally-trained network is replaced with the freshly rebuilt one. Set to 0 to disable periodic retraining and keep online learning only.
• neuralnetwork_stopwords words — default: "the and for with that this from there their have be not but you your". Space-separated list of stopwords to ignore when tokenizing text. the it-girls of words that don't get a seat at the table 💅
• neuralnetwork_autolearn 0|1 — default: 0. When SpamAssassin declares a message a clear spam or ham during the message scan, and launches the auto-learn process, the message is autolearned as spam/ham in the same way as during the manual learning. Value 0 for this option disables the auto-learn process for this plugin.
• neuralnetwork_autolearn_vocab_only 0|1 — default: 0. When set to 1 and auto-learn is enabled, autolearned messages update the vocabulary and training buffer but skip FANN model training and saving. This avoids slow I/O operations and temporary files during spamd processing. The model is rebuilt from the accumulated vocabulary on the next manual sa-learn run. ngl this one is the "do it later" queen 👑
• neuralnetwork_dsn — default: none. The DBI dsn of the database to use.
  • For SQLite, the database will be created automatically if it does not already exist; the supplied path and file must be read/writable by the user running spamassassin or spamd.
  • For MySQL/MariaDB or PostgreSQL, see sql-directory for database table creation clauses.
  • You will need to have the proper DBI module for your database. For example DBD::SQLite, DBD::mysql, DBD::MariaDB or DBD::Pg.
  • Minimum required SQLite version is 3.24.0 (available from DBD::SQLite 1.59_01).
  • Examples:

neuralnetwork_dsn dbi:SQLite:dbname=/var/lib/spamassassin/NeuralNetwork.db

• neuralnetwork_username — default: none. The username that should be used to connect to the database. Not used for SQLite.
• neuralnetwork_password — default: none. The password that should be used to connect to the database. Not used for SQLite. 🙈 shhh
• neuralnetwork_min_vocab_hits n — default: 10. Minimum number of tokens in the email that must exist in the vocabulary for prediction to run.

🔥 EVAL RULES 🔥

these are the little rules you drop into your config, and they SLAP 💅✨

• check_neuralnetwork_spam() — Body eval rule. Returns true when the neural network prediction score exceeds neuralnetwork_spam_threshold (default 0.6).
• check_neuralnetwork_ham() — Body eval rule. Returns true when the neural network prediction score is below neuralnetwork_ham_threshold (default 0.4).
• check_neuralnetwork(low, high) — Body eval rule accepting two optional floating-point arguments. Returns true when the raw prediction score falls within the inclusive range [low, high]. Defaults: low = 0.0, high = 1.0.

Use this rule to define finer-grained confidence tiers. 👑 confidence TIERS, bestie, we love a ranking system:

body      NN_CONFIDENT_SPAM  eval:check_neuralnetwork(0.75, 1.0)
describe  NN_CONFIDENT_SPAM  Email classified as spam with high confidence by Neural Network
score     NN_CONFIDENT_SPAM  2.0

body      NN_PROBABLE_SPAM   eval:check_neuralnetwork(0.55, 0.75)
describe  NN_PROBABLE_SPAM   Email classified as probable spam by Neural Network
score     NN_PROBABLE_SPAM   1.0

body      NN_PROBABLE_HAM    eval:check_neuralnetwork(0.0, 0.4)
describe  NN_PROBABLE_HAM    Email classified as ham by Neural Network
score     NN_PROBABLE_HAM    -1.0

and that's the whole page, bestie! 🎀 this all came from the User Contributed Perl Documentation, running on perl v5.40.1, dated 2026-07-23, and the page signs off as Mail::...Network(3pm) 🌈🦋 go forth and let the neural net read your mail (respectfully) 💖🫶