compromise: Instructions file for Codex

compromise

A rule-based natural-language-processing library for English. Tokenizes text, tags parts-of-speech, and finds & transforms parts of the text. Runs offline in node and the browser, no dependencies. Not an LLM/neural model.

Docs below are plain markdown (the published observablehq.com notebooks do not render as readable text).

Docs

• Concepts: document/View/Term model, mutability, build tiers
• Match syntax: the .match() mini-language
• Tags: the complete part-of-speech tagset
• Tagging-differences: ways compromise differs from other taggers
• API: every method, signature, and description
• Recipes: copy-paste solutions to common tasks
• Development: pnpm scripts, and conventions

Quick mental model

import nlp from 'compromise'

let doc = nlp('she sells seashells by the seashore.')  // parse → a View of the whole document
doc.verbs().toPastTense() // select verbs, transform them (mutates doc)
const str = doc.text() // 'she sold seashells by the seashore.'

• nlp(text) returns a View. Almost every method returns a View, so calls chain.
• Find with .match(), .has(), .if(), or named selections like .people(), .numbers().
• Transform with .toPastTense(), .replace(), .tag(), .normalize(), etc.
• Output with .text(), .json(), .out('array'), .debug().

Rules that prevent most mistakes

1. Transforms mutate the document in place. The View they return is the selection, not the whole doc. Read the final result from the original variable:

   let doc = nlp('I walk to work')
   doc.verbs().toPastTense()
   doc.text()                       // ✅ 'I walked to work'
   // ❌ nlp('I walk to work').verbs().toPastTense().text()  →  'walked work' (selection only)
   

   Use .clone() to transform a copy without touching the original.