@mailwoman/core

The foundation of the Mailwoman address parser — types, tokenization, classification primitives, solver, decoder, and the staged pipeline coordinator. Ships ~9 MB of provenance-tracked reference dictionaries (libpostal, Who's On First, chromium-i18n) consumed by the resolver and classifiers.

import { createRuntimePipeline, AddressParser, ComponentTag, Classification, Span } from "@mailwoman/core"

const pipeline = createRuntimePipeline({ locale: "en-US" })
const result = pipeline.parse("1600 Amphitheatre Parkway, Mountain View, CA 94043")

What's inside

| Module | Purpose | | --------------------- | --------------------------------------------------------------------------------------------------------------------------------------- | | types/ | Core type system: ComponentTag, Span, Classification, ClassificationMap, LocaleTag | | tokenization/ | Tokenizer primitives, whitespace/punctuation rules, token classification | | classification/ | Classification data structure, ClassificationMap, span overlap resolution | | decoder/ | Span proposal → tree projection, BIO decoding, reconcile/merge strategies, confidence calibration | | pipeline/ | createRuntimePipeline — the staged pipeline coordinator that wires normalize → query-shape → locale-gate → ... → classifier → decoder | | solver/ | Rule-based solver (the v0 rules engine), Solution, Solver | | parser/ | AddressParser — high-level parse entry point (consumed by mailwoman CLI) | | resources/ | ~9 MB of shipped reference data: libpostal dictionaries, WOF place data, chromium-i18n address formats |

Key exports

// Types
export type { ComponentTag, Span, Classification, ClassificationMap, LocaleTag }

// Pipeline
export { createRuntimePipeline, type RuntimePipeline, type PipelineOpts }

// Classification
export { Classification, ClassificationMap }
export { treeToClassification, classificationToTree }

// Decoder
export { decodeBioSpans, viterbiDecode, reconcileSpans }
export { createCalibrator, type Calibrator } // isotonic confidence calibration

// Solver (v0 rules)
export { Solver, Solution }

// Tokenization
export { tokenize, Token, TokenClass }

// Resources
export { loadDictionary, getAvailableLanguages }

Pipeline architecture

Mailwoman's runtime pipeline is a staged coordinator that chains pure-function stages with typed handoffs:

normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder

Each stage is published as its own @mailwoman/* package and wired together by the pipeline coordinator in this package. The design ensures every stage is independently testable, benchmarkable, and replaceable.

Reference data

This package ships immutable, provenance-tracked dictionaries consumed by the resolver and rule-based classifiers:

• libpostal — multilingual street types, place names, directional/ordinal tokens
• Who's On First — place hierarchy and geography
• chromium-i18n — per-country address format templates

The dictionaries are ~9 MB total and are loaded lazily.

Related

• mailwoman — the user-facing CLI + AddressParser
• @mailwoman/normalize — Stage 1 of the pipeline
• @mailwoman/neural — neural classifier (ONNX runtime)
• What Mailwoman Is
• Staged Pipeline Contract

License

AGPL-3.0-only