email-content

Extract the conversationally relevant content from arbitrary email messages.

Given a raw RFC 5322 message, the library separates what the sender actually wrote from the reply history their mail client embedded automatically, and returns both, classified, with a confidence and a machine readable reason for every decision.

It is built for a chat style application that uses ordinary email as its transport, where the other participants use whatever client they like and do not know their mail is being read as chat.

The rule everything follows from

Hiding a sentence the sender wrote is a silent data loss the reader may never notice. Showing a paragraph of redundant quoted history is a cosmetic annoyance. The first is treated as at least ten times worse than the second, and every design decision follows from that.

In practice that means:

• Content is only ever hidden when something structural says it is history: an attribution line, a client quote container, or a match against a message that demonstrably existed earlier. Quote characters alone are never enough.
• Nothing is deleted. Hidden segments are returned in full next to the visible ones, and the original message is always available. "Hidden" is a rendering hint.
• Ambiguous content is labelled UNKNOWN and stays visible.
• No generative model touches the sender's words. The pipeline selects and classifies existing spans; it never rewrites, summarises or paraphrases.

Install

npm install @postalsys/email-content

Runtime requirements. The library is plain CommonJS JavaScript with no compiled dependencies, and runs on Node 20 or later. Nothing in the dependency tree is native and none of it is ESM only. There are exactly three direct runtime dependencies, each present for correctness rather than convenience:

| dependency | why | | --- | --- | | @zone-eu/mailsplit | The only local MIME parser that walks the tree non-destructively, emitting raw header and body bytes in document order with IMAP style part numbers. That is what makes every segment traceable to exact source bytes. | | libmime | Header encoded-word decoding, charset handling and format=flowed unfolding. | | node-html-parser | A forgiving HTML parser that exposes a source range on every node, which is what lets an HTML block point back at the offsets it came from. |

The corpus tooling shipped in the same package (corpus, annotate, evaluate, benchmark) needs Node 22.5 or later, because it uses the built in node:sqlite. That module is loaded lazily, so it never touches a host that only calls analyze() or extract(). Running those commands on an older runtime reports the requirement rather than failing obscurely.

Bundling with @yao-pkg/pkg

Verified working as a single file binary on macOS arm64, Linux x64 and Windows x64, built against Node 24 and each producing byte identical classifications.

No pkg configuration is required. The detector pattern files are read at runtime through path.join(__dirname, 'patterns'), a form pkg detects statically and bundles automatically. Adding them to pkg.assets also works but is unnecessary.

If the patterns ever fail to bundle, the library throws at load rather than degrading silently, because the alternative is a binary that recognises no reply history at all and reports no error.

Build from source

Requires Node 24 for the development workflow.

npm install
npm run build
npm test

Library

The primary entry point takes content you have already parsed. Most callers hold { text, html, subject } from mailparser, postal-mime, EmailEngine or an IMAP client, and never have a raw message to hand.

const { analyze } = require('@postalsys/email-content');

const result = await analyze({
    text: plainTextBody,   // optional
    html: htmlBody,        // optional
    subject: 'Re: Deploy window',     // optional
    from: 'Bob <[email protected]>',    // optional, improves signature detection
    inReplyTo: '<[email protected]>' // optional
});

for (const segment of result.segments) {
    console.log(segment.type, segment.confidence, segment.text);
}

Either body may be omitted. When both are present the HTML drives the classification, because that is where quote structure survives, and the plain text is used as a cross check.

If you do have a raw RFC 5322 message, extract() parses it and runs the same pipeline:

const { extract } = require('@postalsys/email-content');
const result = await extract(fs.readFileSync('message.eml'));

Result

Both entry points return the same shape. extract() adds nothing except optional MIME diagnostics.

{
  "segments": [                  // render these
    {
      "type": "NEW_CONTENT",
      "text": "No, wait until Monday.",
      "html": "<div>No, wait until Monday.</div>",   // UNTRUSTED, see the security section
      "confidence": 0.95,
      "visibility": "show",      // show | collapse | hide
      "reasons": ["new_content_default", "html_primary_representation"],
      "source": {
        "kind": "mime-part",     // "mime-part" | "raw-message"
        "mimePart": "1.2",       // IMAP style part path
        "representation": "html",
        "blockIndex": 0,         // position within this part, not within the whole message
        "start": 0,              // character offsets into the canonical part text
        "end": 79,
        "domPath": "/0/1"
      }
    }
  ],
  "hiddenSegments": [ /* same shape, retained in full */ ],
  "overallConfidence": 0.97,
  "warnings": [],
  "originalAvailable": true,
  "primaryRepresentation": "html",
  "extractorVersion": "<current package version>",
  "historyUsed": false,
  "isReply": true
}

Segment types: NEW_CONTENT, INLINE_QUOTE, REPLY_HISTORY, FORWARDED_CONTENT, SIGNATURE, DISCLAIMER, UNKNOWN.

Render segments. hiddenSegments holds what the extractor is confident is embedded reply history, in full, so a client can offer to reveal it. Nothing is ever discarded.

A few fields are easy to misread:

• originalAvailable is always true, and it is a statement about behaviour, not about this object. The library never mutates or consumes your source, so the original is always still recoverable. It is not embedded in the result: you must retain it.
• hiddenSegments holds only segments whose visibility is hide. Segments marked collapse stay in segments, because a collapsed segment is still rendered, just folded.
• primaryRepresentation names what drove the classification of the primary section. It is not an inventory of every MIME part represented: a message with several body sections can produce segments from more than one part.
• source.kind discriminates the two provenance cases. mime-part carries the part path, representation and block index. raw-message carries only a byte range, and is used when MIME parsing failed outright and there is no part structure to point at.
• isReply: false means "not known to be a reply". When parsing fails there are no threading headers to read, and the result carries a reply_status_unknown warning to say so.

Thread history

If earlier messages in the thread are available, pass them. A block whose text provably existed before this message was written is history by definition, which is far stronger evidence than any markup heuristic. Measured on the reference set, supplying history cuts the share of reply history left visible from 70.9 percent to 60.1 percent, with no loss of new content.

const result = await analyze(
    { text, html },
    { history: [{ text: parentText, html: parentHtml }] }   // nearest ancestor first
);

History is always optional. Extraction works without it and never depends on it.

It earns its place twice over. A block whose text is found in an earlier message is history by provenance rather than by markup, which both raises confidence on regions the structure already recognised and catches the ones it cannot see at all: clients that paste the previous message as ordinary paragraphs, with no quote prefix, no blockquote and no attribution line. Nothing in such a message says it is quoting, so only the thread can tell. That case is collapsed rather than hidden, because a match proves the text existed before and proves nothing about whether the reader still needs it.

If you are reading a mailbox you already have what this needs: resolve the parent by In-Reply-To or References, usually from Sent, and pass it. Callers that resolve threads themselves can build the documents once with buildHistoryDocument() and pass historyDocuments instead, which skips re-parsing the ancestor on every message.

analyze() and extract() never throw for malformed message input. A body that cannot be segmented is returned as a single visible UNKNOWN segment plus a warning.

Security: rendering HTML

segment.html is untrusted, attacker controlled markup, copied verbatim out of the message. It can contain event handler attributes, remote images and stylesheets that report the fact and time of reading, forms, frames, unsafe SVG, javascript: and data: URLs, and tracking links.

Classification is not sanitisation. This library decides whether a span of the message is new content, reply history or a footer. It says nothing about whether that span is safe to put into a browser, and it deliberately does not rewrite source markup, because rewriting would destroy the byte ranges every segment points at.

No sanitiser ships here, and that is deliberate too. Sanitisation is context dependent and security sensitive, and a weak one bundled into an extraction library would be worse than none: it would look like protection without being it.

The order matters:

raw message
  -> MIME parse and classification      (this library)
  -> choose visible segments            (your product policy)
  -> sanitise each untrusted HTML segment   (a maintained sanitiser)
  -> render under a restrictive CSP

• Never assign segment.html to innerHTML, dangerouslySetInnerHTML or any equivalent.
• Sanitise after segmentation and immediately before rendering. Sanitising before extraction strips the Gmail and Outlook quote containers the detectors need, and costs accuracy.
• Block remote images, stylesheets, fonts, media, frames and forms. Loading them leaks that the mail was opened, from where, and when.
• Strip event attributes, and dangerous URL schemes such as javascript: and vbscript:. Decide a deliberate policy for data: URLs rather than allowing them by default.
• Rewrite or proxy links according to your product's policy.
• Use a Content Security Policy as defence in depth, not as the sanitiser. CSP does not neutralise every dangerous construct or navigation.
• Do not load remote resources while sanitising or previewing, on the server either.

Desktop mail clients and browser renderers have different threat models. Document which environment your product supports rather than assuming one set of rules covers both. Before rolling extraction out, prove in an integration test in the consuming application that active attributes, hostile URLs, forms and tracking images are all neutralised and that no remote request is made.

Running this on untrusted mail

Inbound mail is attacker controlled, and this library does its work synchronously on the caller's event loop. Every algorithm here is bounded or linear in its input, and the adversarial cases that were not are covered by scaling tests in test/adversarial.test.ts. That is a floor, not a guarantee: the library cannot interrupt its own synchronous work, so a future bug in a detector or in a dependency is a stalled event loop for whoever called it.

A service that renders mail from strangers should therefore put a boundary around extraction rather than rely on the library alone:

• Cap the raw message before extract(). Reject oversized payloads at the transport if you can, before buffering. maxMessageBytes is a backstop for what still gets through, not a substitute.
• Cap the bodies before analyze(), with maxTotalDecodedTextBytes and maxContentChars. Per part limits do not bound total work: many legal sized parts add up.
• Run untrusted parsing in a worker thread or subprocess. That is the only way to enforce a wall clock or memory limit, because terminating the worker is the only mechanism that can interrupt a synchronous call.
• Enforce a wall clock timeout by terminating that worker, not by racing a promise. A promise race resolves while the blocked loop keeps burning.
• Bound concurrency, so a burst of large messages cannot occupy every worker at once.
• Keep the original message. On timeout, memory pressure or any degraded result, fall back to rendering it. Extraction output is a rendering artifact that can always be recomputed; the raw message is the system of record.

A syntactically small input can still exploit an algorithmic bug, so size limits do not remove the need for isolation.

Options and limits

const result = await extract(raw, {
    parse: {
        maxMessageBytes: 64 * 1024 * 1024,        // whole raw message
        maxTotalDecodedTextBytes: 16 * 1024 * 1024, // decoded text across every part
        maxTextPartBytes: 1024 * 1024,            // decoded text per part
        maxParts: 500,
        maxHeadSize: 1024 * 1024
    },
    confidence: { collapseThreshold: 0.7 }        // partial: the rest keep their defaults
});

const result2 = await analyze({ text, html }, { maxContentChars: 8_000_000 });

Recommended ceilings, documented rather than enforced, because the host knows its own memory budget and a startup crash on a value you chose deliberately protects nobody: maxTextPartBytes around 32 MiB, maxParts around 10,000, maxHeadSize around 16 MiB.

Malformed input and invalid configuration are handled differently, on purpose.

| | behaviour | | --- | --- | | Malformed message | Degrades. Never throws. Returns what it can, plus a warning. | | Invalid option | Throws ConfigurationError before any processing. |

A bad message is the sender's doing and the caller cannot fix it. A bad option is the programmer's own mistake, it will repeat on every message, and returning a plausible looking degraded result for it would hide an operational fault behind what reads like ordinary output. ConfigurationError carries a stable code and the option name, and never the offending value, message content or credentials, so it is safe to log.

Values are rejected, never clamped: clamping makes the pipeline behave differently from what you asked for, silently, forever. NaN, infinities, negatives, zero where a positive is required, fractions where an integer is required, and an inverted collapseThreshold/hideThreshold pair all throw. A large but finite value is accepted.

Exceeding an input limit is not a configuration error. It emits input_limit_exceeded naming the option that fired, returns in full whatever the limit did allow through, and never throws.

Validation runs on every call, so it is kept cheap and deterministic. It is not amortised: if you need it paid once, hold the validated options yourself with validateParseOptions() and validateConfidenceConfig().

The command line validates the same way and exits with status 2, naming the option, rather than letting --concurrency -1 or --timeout -1 reach a worker pool or a SQL query.

Command line

# Classify one message
email-content extract message.eml
email-content extract message.eml --json
email-content extract message.eml --explain      # per block evidence, features and reasons

# Label blocks with an LLM instead of a human (opt in, sends content to OpenAI)
email-content annotate --sample 60 --model gpt-5.4-mini
email-content disagreements --direction expensive

# Scan a corpus of .eml files into a local SQLite database
email-content corpus scan /path/to/eml-folder
email-content corpus scan /path/to/eml-folder --sample 2000 --seed phase1
email-content corpus scan /path/to/eml-folder --concurrency 14

# Report on it
email-content corpus stats --out docs/corpus-report.md
email-content corpus stats --json
email-content corpus threads                     # in-corpus thread parent resolution

# Throughput and catastrophic error rates
email-content benchmark --sample 3000

What it gets right

The two cases that define the problem:

> Should we deploy this today?

No, wait until Monday.

> Should I notify the customer?

Yes, please.

Both questions are INLINE_QUOTE and stay visible: they are the context the answers depend on.

Yes, Monday works.

On Friday, Alice wrote:
> Can we meet on Monday?

The trailing quote is REPLY_HISTORY and is hidden, because an attribution line anchors it and no authored content follows.

Beyond that, the deterministic baseline handles Gmail, Outlook desktop, Outlook Web, Apple Mail, Thunderbird, Proton, Yahoo and Zimbra quote structures; forwarded messages with a comment above them; bottom posted replies, which are collapsed rather than hidden because the reader is more likely to need the context; format=flowed; legacy charsets; and messages whose plain text part has lost its quote markers entirely, which is now the majority case for recent mail.

It also, deliberately, does not fire on ordinary prose that merely looks like email structure. Text containing From: and Subject: lines is only treated as an embedded header block when the values look like header values, so a support message explaining which headers to include stays visible.

Accuracy

Measured against a 452 message stratified reference set annotated by gpt-5.6-sol under the conventions in docs/annotation-guide.md, with thread history enabled.

The reference is not ground truth, but it is close. Two strong models (gpt-5.5 and gpt-5.6-sol) given identical input agree with each other on 97.2 percent of blocks under the current conventions, and 98.5 percent once three messages are excluded in which a single message-level judgment flips every block at once. 392 of the 452 messages agree on every single block. That is the practical ceiling; see docs/annotation.md.

| property | measured | | --- | ---: | | block label agreement | 83.7% | | new content characters retained | 100.0% | | new content blocks retained | 100.0% | | hidden content genuinely history | 99.8% | | reply history correctly hidden | 63.4% | | messages losing all new content | 0 | | messages losing an inline answer | 0 |

Not one character the reference considers newly written is hidden, across 452 messages and 221,256 characters of new content.

Per label F1: REPLY_HISTORY 93.3, FORWARDED_CONTENT 82.4, NEW_CONTENT 75.4, SIGNATURE 52.0, INLINE_QUOTE 36.9, DISCLAIMER 33.7.

Recall is the number to read here, not F1. Precision is measured against a reference that labels every block, while the extractor deliberately leaves anything it cannot anchor visible, so an honest "I am not sure" scores as a miss. REPLY_HISTORY recall is 89.5 percent. NEW_CONTENT label recall is 92.7 percent, and the difference from the 100 percent character retention is entirely blocks the extractor labels UNKNOWN, which are always shown: an honest "not sure" costs label agreement, never content.

With thread history available the numbers are materially better than the aggregate suggests, and a chat application always has the thread it is rendering:

| | messages | agreement | history hidden | | --- | ---: | ---: | ---: | | history available | 305 | 88.3% | 66.3% | | history unavailable | 147 | 69.8% | 45.9% |

Limitations

• It still errs towards showing too much. Around 37 percent of reply history remains displayed. That is the cheap direction of error and it is deliberate, but it is the largest remaining gap.
• Disclaimer recall is 20.9 percent at 86.4 percent precision. The generic notification footer shapes are covered; what remains is per-sender templates, which hand written rules cannot reach without overfitting to one vendor. This is the clearest case for a learned model.
• Signature recall is 39.6 percent at 75.6 percent precision. Two independent signals are required, such as a closing greeting plus contact details, or the sender's own name plus either.
• INLINE_QUOTE precision is 27.9 percent. Blocks wrongly called inline quote stay visible, so this costs tidiness, not content. Under the settled depth-1 convention this is now a genuine disagreement with the reference, not a labelling artefact: the extractor treats any quoted run with authored text after it as answered, and the reference frequently reads the same run as a trailing copy.
• Outlook desktop remains the weakest recognised client, around 70 percent against 92 percent for Gmail and Apple Mail; unrecognised clients sit at 68 percent. Outlook Web reached 84 percent once its embedded header blocks were detected across sibling paragraphs.
• No learned model. This is the deterministic baseline. The labelled data, a settled labelling convention and the evaluation harness all exist to train one.

Privacy

The corpus this was built against is private email, and the project is built accordingly:

• Nothing contacts the network by default. The runtime dependency tree is three packages and none opens a socket. The one exception is email-content annotate, an opt in command that sends block text to an LLM API; see docs/annotation.md.
• Source .eml files are opened read only and never modified.
• No message body, subject or address is written to logs, to the terminal, or to the corpus database by any command except disagreements, which prints block text on purpose so a human can adjudicate it. Subjects and addresses are stored as salted hashes; sender domains are kept because stratified sampling needs them.
• "Processed locally" is not the same as "can never leave the machine". annotate sends block text to an LLM API when you opt into it, file paths and diagnostics appear in your own terminal and shell history, and the corpus database sits on whatever disk and backup you point it at.
• The corpus database is derived from private mail and is not anonymised. It is gitignored. See docs/threat-model.md for exactly what it can still reveal.

Documentation

| document | contents | | --- | --- | | docs/architecture.md | Pipeline stages, dependency choices and why, safety invariants | | docs/adr-001-document-global-classification.md | Why multipart sections are classified as one sequence, and the two index spaces that follow | | docs/implementation-plan.md | Phases, decisions taken, open questions | | docs/threat-model.md | What is stored, what it reveals, hostile content handling | | docs/corpus-report.md | Statistics over the working corpus (generated, gitignored) | | docs/benchmark.md | Throughput, memory, and where the time goes | | docs/annotation.md | Using an LLM as annotator: method, results, limitations |

Project layout

src/mime/        MIME parsing, charset decoding, normalised message model
src/canonical/   Plain text and HTML canonicalisation
src/segment/     Block segmentation and feature extraction
src/detect/      Detectors, with pattern sets as JSON data
src/classify/    Rules, reconciliation, confidence
src/corpus/      Scanner, worker pool, SQLite store, statistics
src/cli/         Command line entry points
test/            Tests and hand written fixtures

Pattern sets in src/detect/patterns/ can be overridden at runtime by pointing EMAIL_CONTENT_PATTERNS at a directory containing replacements.

Status

Working and measured: the MIME parser, canonicalisation, block segmentation, the deterministic rule baseline, section aware multipart body planning, thread history matching, the corpus scanner over 320k messages, LLM assisted annotation, the evaluation harness, the CLI, and a test suite covering fuzzing, hostile input scaling bounds and the two content loss regressions. Run npm test for the current count rather than trusting a number written here, which goes stale the moment a test is added.

Planned, not built. Described in docs/implementation-plan.md:

• the local review application for adjudicating annotator disagreements;
• the corpus redaction workflow;
• a learned sequence model.

Anything in this README describing those is design intent, not a shipped feature.