nlpkit.js

Dependency-free NLP building blocks for JavaScript and TypeScript. The first module is a variable-order n-gram model for next-word suggestions.

Install

npm install @vlchnk/nlpkit.js

Train and predict

import { buildNgramModel, suggestNextWords } from "@vlchnk/nlpkit.js/ngram";

const model = buildNgramModel(
  [
    { text: "Good morning! How can I help you?" },
    { text: "Good evening! How can I help you?", weight: 2 },
  ],
  { maxOrder: 4, topK: 10, minCount: 1 },
);

const result = suggestNextWords(model, "How can I he");

// result.candidates[0].w === "help"
// Replace text in [result.replaceStart, result.replaceEnd] with that word.

Training and prediction use the same Unicode-aware normalization. Prediction tries the longest available context first and backs off to shorter contexts if necessary. It also filters candidates by the partial word immediately before the cursor.

buildNgramModel returns plain JSON-compatible entries, so a model can be saved to a file, database, object storage, or Redis. chooseCandidates is provided for the Redis MGET pattern used by the original implementation, and Redis key helpers are exported separately without adding a Redis client dependency.

Model options

• maxOrder — maximum number of preceding words, default 4.
• topK — candidates retained per context, default 10.
• minCount — minimum weighted occurrence count, default 2.
• A record's optional positive weight multiplies all counts from that text.

The model is intended as a small, explainable baseline for operator-assist autocomplete. For large corpora, train offline and store each entry by its context key rather than rebuilding the model for every request.