Completions chain

Build a Markov chain with document glyphs attached to each transition.

completions.New() creates an in-memory completion chain. Ingest text to learn transitions. Call Complete() to rank next-word candidates by glyph similarity.

Create a chain

import { completions } from "@koda.oss/glyph";

const chain = completions.New({
  order: 3,
  create: { size: 128, normalize: true },
});

Option	Default	Meaning
order	3	Markov order (state = previous order - 1 tokens)
create	{}	Options forwarded to Create() on ingest and complete

Ingest documents

chain.Ingest("moon-doc", "goodbye moon farewell night");
chain.Ingest("sun-doc", "goodbye sun hello day");

Ingest(key, text) requires a key. That key is stored on every transition and returned as source.key on completion results.

Per ingest:

1. Fingerprint the full document → glyph G
2. Tokenize with stripped unigrams (CreateUnigrams / TextStrip) — letters and digits only
3. For each Markov transition, store G and key on that edge

Chain tokens are stripped, not just filtered. Punctuation and symbols do not enter the Markov state.

Storage model

Map<stateKey, Map<nextToken, { count, sources[] }>>

Field	Meaning
stateKey	Previous order - 1 tokens joined by space ("" when order: 1)
count	How many times this transition was seen
sources	{ key, glyph, weight } per ingested document

Duplicate Ingest calls with the same key merge weights on matching transitions.

Chain methods

Method	Behavior
Ingest(key, text)	Add transitions from a keyed document
Complete(prefix, options?)	Rank next-token candidates
Clear()	Remove all states
Size()	Number of distinct state keys