embedding-cache-lite

Content-hash cache for embedding vectors. Wraps any embedder function so identical text (per model) only gets embedded once — cuts redundant API calls in RAG ingestion pipelines that re-run over overlapping content.

Install

npm install embedding-cache-lite

Quick start

import { EmbeddingCache } from 'embedding-cache-lite';

const cache = new EmbeddingCache(
  (text) => openai.embeddings.create({ input: text, model: 'text-embedding-3-small' }).then(r => r.data[0].embedding),
  { model: 'text-embedding-3-small', maxEntries: 10000 }
);

const vector = await cache.embed(chunk.text); // only calls the API on a real miss
console.log(cache.stats()); // { hits, misses }

Why key on text + model

The same text embedded with two different models produces different, non-comparable vectors — caching purely by text content would silently return the wrong model's vector if you ever switch models or run two side by side. Keying on a hash of model:text keeps that safe by construction.

API

• new EmbeddingCache(embedder, options?) — options: { model?, maxEntries? }
• .embed(text) — returns the cached vector or calls embedder and caches the result
• .has(text) / .size() / .stats() / .clear()

Eviction is simple LRU once maxEntries is exceeded.

License

MIT