took place the release of the spam filtering system rspamd 3.9 , providing means for evaluating messages according to various criteria, including rules, statistical methods and black lists, on the basis of which the final weight of the message is formed to make a decision on the need to block. RSPAMD supports almost all the possibilities implemented in SpamasssSin, and has a number of features that allow you to filter mail on average 10 times faster than SpamasssSin, as well as ensure the best filtration quality. The system code is written in the language of SI and spreads under the license Apache 2.0.
RSPAMD is built using event-oriented architecture (Event-Driven) and is originally designed for use in highly loaded systems, allowing you to process hundreds of messages per second. The rules for identifying signs of spam are highly flexible and can contain regular expressions in the simplest form, and in more difficult situations they can be decorated in LUA. The expansion of functionality and the addition of new types of inspections is realized through the modules that can be created in SI and LUA. For example, modules are available for checking the sender using SPF, confirming the sender domain via DKIM, forming requests to the DNSBL lists. To simplify the setting, creating rules and tracking statistics, an administrative Web interface is provided.
In the new version:
- Improved settings of the Bayesian classifier. The default window is reduced from 5 to 2 words, which allowed to achieve increased performance and 4 times reduce the consumption of a place in the storage without degradation of the level of spam classification. To test the work of the classifier with different settings, a utility is proposed
“Raspamadm classifier_test”. - added GPT module, using the API OpenAi GPT to classify the text through a request to large language models , such as GPT-3.5 Turbo and GPT-4O. The accuracy of the classification of spam using the new module is lower than that of the Bayesian classifier, but its advantage is that it does not require preliminary training and can take into account the context in messages, while the effective work of the Bayesov classifier needs high -quality and balanced engine training. In addition to direct detection of spam in messages, the GPT module can be used to train the Bayesian classifier.
- The possibility of joint use of Known_Senders and Replies modules was implemented to mark the verified senders, using as a sign that they had previously sent answers.
- By default, a dynamic change in restrictions on the intensity of sending messages ( Dynamic Ratelimit ) related to one sender or recipient or recipient or recipient IP address.