← All packages

@lacspace/condense

v1.0.1LLM Efficiency2 @lacspace deps

Extractive multi-source condenser — turn several articles on one story into a short, deduplicated, token-budgeted digest that keeps the numbers, quotes and named entities, so an LLM only rewrites a fraction of the text. BM25 sentence ranking, near-duplicate removal, per-source attribution with char offsets, Devanagari-aware (।/॥, ०-९, रु). Deterministic, isomorphic.

npm i @lacspace/condense

Usage

condense.ts
import { condense } from "@lacspace/condense";

const digest = condense(
  [
    { text: kathmanduPost, label: "Kathmandu Post" },
    { text: himalayanTimes, label: "Himalayan Times" },
    { text: onlineKhabar, label: "Online Khabar" },
  ],
  { tokenBudget: 1500, gazetteer: ["Nepal Rastra Bank", "\u0928\u0947\u092a\u093e\u0932 \u0930\u093e\u0937\u094d\u091f\u094d\u0930 \u092c\u0948\u0902\u0915"] },
);

digest.text;       // kept sentences grouped per source under [S1 Kathmandu Post] headers
digest.tokens;     // \u2264 tokenBudget
digest.droppedDup; // near-duplicate sentences removed across sources
// Feed digest.text to the model instead of six full articles.

Exports 2

condensesplitSentences

Keywords

condensesummarizeextractive-summarymulti-documentbm25token-budgetdedupeprompt-compressionragdevanagari

More in LLM Efficiency

@lacspace/screen

A cheap lexical content screener and LLM admission gate — score text against your own weighted, multi-language term lexicons (with negation, proximity windows and a named-entity gazetteer) and get a clear / review / block decision, so obviously-clean and obviously-flagged text never reaches an expensive model. Deterministic, auditable, isomorphic.

@lacspace/keyphrase

Extractive keyphrases, tags, hashtags, named entities and category votes — a zero-dependency RAKE + TF-IDF engine with built-in English and Nepali stopwords, Devanagari-aware, so you can stop asking an LLM to generate tags/hashtags/entities. Deterministic, isomorphic.

@lacspace/llm-cache

A content-hash cache for LLM calls — key by (model family, prompt version, normalized input) so identical requests, retries after a 429 and repeated rewrites never pay twice. Pluggable async store (in-memory LRU built in, Mongo/Redis/KV via an adapter), TTL, stale-if-error and a wrap() memoizer. Zero-dependency, isomorphic.

@lacspace/keypool

AI key pooling two ways. createAiChain: route-ordered multi-provider failover (Gemini, Groq, Cerebras, OpenAI, Anthropic, DeepSeek, OpenRouter, any OpenAI-compatible) that rests key+model pairs on 429 (Gemini daily quota until Pacific midnight), quarantines rejected keys (incl. Gemini's 400 'API key not valid'), cools down after N failures, shrinks Groq max_tokens to fit TPM, regrows empty reasoning replies, JSON mode with repair, per-purpose routes and caps, a call ledger hook, and a shared store (memory or Redis). createKeypool: RPM/RPD/TPM/TPD rotation and accounting. Built on @lacspace/ai.