@lacspace/llm-cache

Never pay twice for the same LLM call. A content-hash cache keyed by (model family, prompt version, normalized input), so identical requests, retries after a 429, and repeated rewrites are served from cache. Pluggable async store (in-memory LRU built in; Mongo/Redis/KV via a tiny interface), TTL, stale-if-error, and a wrap() memoizer. Zero dependencies, isomorphic.

npm i @lacspace/llm-cache

import { createLlmCache, memoryStore } from "@lacspace/llm-cache";

const cache = createLlmCache({ store: memoryStore(), ttlMs: 24 * 3600_000, promptVersion: "v3" });

const pack = await cache.wrap(
  { input: condensedSources, model: "gemini", variant: { lang: "ne" } },
  () => callGemini(condensedSources),          // only runs on a miss
  { staleIfError: true },                       // 429? serve the last good result
);

cache.stats(); // { hits, misses, sets }

• Stable key across runtimes via a wide (128-bit) dependency-free content hash; whitespace-normalized input by default.
• wrap(input, fn) memoizes an async call: the retry that a rate-limit forces you into costs nothing.
• Pluggable store — implement { get, set, delete } (sync or async) over Mongo/Redis/KV to share the cache across PM2 processes or hosts; kvStore-style adapters are trivial.
• TTL + stale-if-error — expire results, but fall back to the last good one when the provider is down.

Bump promptVersion when your template changes and old results retire automatically. Exports contentHash, memoryStore, and the CacheStore interface.

Licence

Lacspace Free Licence v1.0 — free for personal and commercial use.