@mailwoman/normalize

Stage 1 of the Mailwoman runtime pipeline — deterministic input preprocessing.

Pure-function text normalization that prepares free-text address strings for downstream parsing stages. Every transform produces a load-bearing offsetMap so downstream stages can map normalized-string spans back to raw-string character offsets.

import { normalize } from "@mailwoman/normalize"

const result = normalize("123  Main   St.")
// result.normalized → "123 Main St."
// result.offsetMap  → maps each normalized char back to raw

What it does

| Transform | Purpose | | ----------------------------- | --------------------------------------------------------------------------- | | NFC normalization | Unicode canonical composition | | Punctuation normalization | Smart-quotes → straight, fullwidth → ASCII, elision/apostrophe preservation | | Whitespace collapse | Multi-space, tab, non-breaking → single space; leading/trailing trim | | Abbreviation expansion | Opt-in — "St." → "Street", "Ave" → "Avenue" etc. | | CJK normalization | CJK-specific whitespace and punctuation handling |

API

// Full normalization pipeline (NFC → punctuation → whitespace)
normalize(input: string, opts?: NormalizeOpts): NormalizedInput

// Individual transforms (if you need only one)
applyNfc(input: string): NormalizedInput
applyPunctuation(input: string): NormalizedInput
collapseWhitespace(input: string): NormalizedInput
expandAbbreviations(input: string, opts?: ExpandOpts): NormalizedInput
applyCjkNormalization(input: string): CjkResult

// Offset map utilities
composeMaps(inner: OffsetMap, outer: OffsetMap): OffsetMap
identityMap(length: number): OffsetMap

Pipeline position

raw string → normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → ...

Stage 1 in the Staged Pipeline Contract. No runtime dependencies.

Design

• Pure functions, no side effects, no ML. The output is byte-for-byte deterministic for the same input.
• offsetMap is load-bearing. Every transform tracks how normalized positions map back to raw input positions. This is essential for the parser to report spans in the original string.
• Configurable via NormalizeOpts: toggle expandAbbreviations, normalizeCase, and cjk.

Related

• @mailwoman/query-shape — Stage 1.5, structural priors that consume the normalized output
• Staged Pipeline Contract
• Tokenization concepts

License

AGPL-3.0-only