# API reference ```js import SpamScanner from 'spamscanner'; // ESM const SpamScanner = require('spamscanner'); // CommonJS const scanner = new SpamScanner(options); const result = await scanner.scan(rawMessage, scanOptions); ``` Create one scanner and reuse it: it loads the model once and caches DNS and language model answers. ## Options Every option can be passed to the constructor. Most can also be passed to `scan()` for one message, which merges them over the constructor's. | Option | Default | Meaning | | ----------------------- | -------------------------------- | --------------------------------------------------------------------------------------------------------- | | `threshold` | `5` | Score at which a message is spam | | `rejectThreshold` | `15` | Score at which `action` is `reject` | | `scores` | `{}` | Points per test: setting keys or test names ([tests and scores](https://spamscanner.net/docs/scoring/)) | | `classifier` | bundled model | A `Classifier`, a model object, a model file path, or `false` | | `classifierOptions` | `{}` | Options for the bundled model (see [Classifier](https://spamscanner.net/docs/api/#classifier)) | | `allowedLanguages` | `[]` | ISO 639-1 codes; other languages get `LANGUAGE_NOT_ALLOWED` | | `phishing.cloudflare` | `true` | Ask Cloudflare's filtering resolvers about link hosts | | `phishing.adult` | `true` | Also ask the family resolver, which blocks adult sites | | `phishing.maxHosts` | `25` | Link hosts looked up per message | | `phishing.homograph` | `{}` | `brands` (replace the built-in list), `extraBrands`, `allowlist` (domains never flagged), `strictMode` | | `dnsbl` | `{ip: [], domain: []}` | DNS blocklist zones for the client IP and for link domains | | `dns` | `{servers: null, timeout: 3000}` | Name servers for DNS checks (default: the system's) | | `attachments` | `true` | Inspect attachments | | `macros` | `true` | Flag macros, active PDFs and RTF objects | | `arbitrary` | `true` | Run the [rules](https://spamscanner.net/docs/scoring/#rules) | | `authentication` | `false` | `true`, or `{dnsServers, timeout, mta, weights}`: SPF, DKIM, DMARC, ARC | | `reputation` | `false` | `{allowlist, denylist, apiUrl, headers, timeout}` | | `allowlist`, `denylist` | | IP addresses, domains or addresses; shorthand for `reputation` | | `clamav` | `false` | `true` (default socket), `{socket}` or `{host, port}` | | `llm` | `null` | [Language model settings](https://spamscanner.net/docs/llm/#any-server-port-and-authentication), with `mode`, `minScore`, `maxScore` | | `toxicity` | `false` | `{model, threshold = 0.7}`: a model whose `classify([text])` answers like `@tensorflow-models/toxicity` | | `nsfw` | `false` | `{model, threshold = 0.6}`: a model whose `classify(imageBuffer)` answers like `nsfwjs` | | `maxLength` | `100000` | Characters of body text read | | `timeout` | `10000` | Milliseconds allowed for each network check | | `session` | `{}` | Default SMTP session details | `scan()` also takes `session`: ```js await scanner.scan(raw, { session: { remoteAddress: '203.0.113.5', // the client's IP address resolvedClientHostname: 'mx.example.com', // its verified reverse DNS name helo: 'mx.example.com', envelope: { mailFrom: {address: 'alice@example.com'}, rcptTo: [{address: 'bob@example.org'}], }, }, }); ``` A server that has already checked SPF, DKIM, DMARC and ARC (with [mailauth](https://github.com/postalsys/mailauth)) can pass its results as `authenticationResults`, and they are scored instead of checked again: ```js import {authenticate} from 'mailauth'; const auth = await authenticate(raw, {ip, helo, sender, mta}); await scanner.scan(raw, {authenticationResults: auth}); ``` ## Methods | Method | Returns | | ------------------------------------ | ---------------------------------------------------------------------------------- | | `scan(source, options)` | The result. `source` is a Buffer, string, Uint8Array or readable stream | | `scanFile(path, options)` | The result for a message file | | `learn(source, 'spam' \| 'ham')` | Teaches the classifier one message | | `unlearn(source, 'spam' \| 'ham')` | Undoes `learn` | | `saveModel(path, options)` | Writes the classifier to a file (`options`: `minCount`, `maxFeatures`, `metadata`) | | `getClassifier()` | The `Classifier` in use, or `null` | | `getClassification(features)` | The classifier's verdict for features from `getFeatures` | | `getFeatures(mail)` | `{features, words, language, script, links, hosts, text, hidden}` of a message | | `getTokens(text, locale)` | The words of a text | | `getTokensAndMailFromSource(source)` | `{tokens, features, mail}` | | `parse(source)` | `{raw, mail}`, with `mail` from mailparser | In `getFeatures`, `text` is the text a mail client shows, which the words are read from, and `hidden` how many characters of text the HTML hides. `scanner.metrics` holds `totalScans`, `averageTime` and `lastScanTime`. ## The result ```js { isSpam: true, score: 16.25, threshold: 5, rejectThreshold: 15, action: 'reject', // 'accept', 'tag' or 'reject' message: 'Spam (BAYES_999, PHISHING_LOOKALIKE_DOMAIN, DECEPTIVE_LINK, FROM_NAME_BRAND)', tests: [ {name: 'BAYES_999', score: 6.25, description: 'Classifier spam probability 99.9%'}, {name: 'PHISHING_LOOKALIKE_DOMAIN', score: 5, description: '"paypa1-secure.top" imitates paypal by swapping characters'}, // ... ], results: { classification: {probability, category, spam, ham, clues, coverage}, phishing: [], // lookalike domains, deceptive links, Cloudflare and URIBL findings uncheckedHosts: 0, // link hosts beyond phishing.maxHosts, not looked up attachments: [], // every attachment finding executables: [], // the executable findings among them macros: [], // macros, active PDFs, RTF objects viruses: [], // ClamAV findings: {filename, virus: [names], message} arbitrary: [], // rules strong enough to mark spam alone obfuscation: {invisible, mixed, styled}, authentication: null, // mailauth's results with a score, when checked reputation: null, dnsbl: [], // {zone, value} for each listing language: {language, script, notAllowed}, llm: null, // the language model's verdict, when asked toxicity: [], nsfw: [], idnHomographAttack: {detected, domains, riskScore}, }, links: ['http://paypa1-secure.top/login'], language: 'en', tokens: ['dear', 'customer', /* ... */], mail: {/* the parsed message, from mailparser */}, version: '7.0.0', metrics: {totalTime: 42}, } ``` Findings in `phishing`, `attachments`, `executables`, `macros`, `viruses` and `arbitrary` are objects with a `type` and a `message`. `String(finding)` is the message. ## Exports ```js import SpamScanner, { Classifier, getFeatures, segmentWords, detectLanguage, LLMClassifier, PROVIDERS, RECOMMENDED_MODELS, CLASSIFIER_MODELS, DEFAULT_SCORES, scoreResults, spamHeaders, rewriteMessage, loadModel, saveModel, defaultModelPath, loadDefaultModel, train, evaluate, readExamples, MilterServer, createHttpServer, createTcpServer, createSpamdServer, DEFAULTS, VERSION, } from 'spamscanner'; import ArfParser from 'spamscanner/arf'; ``` ### Classifier ```js const classifier = new Classifier({spamCutoff: 0.99, hamCutoff: 0.2}); classifier.learn(getFeatures(mail).features, 'spam'); classifier.classify(features); // {probability, category, spam, ham, clues, coverage} ``` | Option | Default | Meaning | | -------------------------------------- | -------------- | -------------------------------------------------------------------- | | `strength`, `unknown` | `0.45`, `0.5` | Robinson's s and x: how strongly rare features are pulled toward 0.5 | | `minDistance` | `0.1` | Features closer to 0.5 than this are ignored | | `maxClues` | `150` | Strongest features combined per message | | `hamCutoff`, `spamCutoff` | `0.2`, `0.99` | Bounds of the `unsure` range | | `minLanguageExamples`, `languageShare` | `1000`, `0.02` | Messages of each class needed in a language for full confidence | | `languagePrior` | `true` | Weigh words against their own language's counts | `merge(other)`, `toJSON(options)`, `Classifier.fromJSON(json)`, `size` and `entries()` are also available. ### Training ```js import {train, evaluate, saveModel} from 'spamscanner'; const {classifier, spam, ham} = await train({ spam: ['/mail/Junk'], ham: ['/mail/Archive', '/mail/archive.mbox'], datasets: [{file: 'extra.csv', textColumn: 'body', labelColumn: 'kind'}], }); saveModel(classifier, 'model.json'); ``` `evaluate(classifier, examples)` takes `{label, features}` items, or `readExamples(sources)`, which reads the same sources as `train`, and returns counts, precision, recall, F1, accuracy and false positive, false negative and unsure rates. ### Servers ```js import SpamScanner, {MilterServer, createHttpServer, createSpamdServer} from 'spamscanner'; const scanner = new SpamScanner({authentication: true}); const milter = new MilterServer(scanner, {reject: true, rejectCode: 451, subjectTag: '[SPAM]'}); milter.on('scan', ({session, result}) => console.log(session.remoteAddress, result.score)); await milter.listen(7831); createHttpServer(scanner, {token: process.env.SPAMSCANNER_TOKEN}).listen(7832, '127.0.0.1'); createSpamdServer(scanner).listen(783, '127.0.0.1'); ``` ### ARF reports `spamscanner/arf` reads and writes abuse feedback reports (RFC 5965), the format mailbox providers use to report spam complaints: ```js import ArfParser from 'spamscanner/arf'; const report = await ArfParser.parse(rawReport); console.log(report.feedbackType, report.sourceIp, report.originalHeaders.subject); const raw = ArfParser.create({ feedbackType: 'abuse', userAgent: 'MyService/1.0', from: 'abuse@example.com', to: 'fbl@example.net', originalMessage, sourceIp: '192.0.2.1', }); ``` `ArfParser.tryParse()` returns `null` instead of throwing for messages that are not reports, and `isArfMessage(mail)` checks a parsed message.