@lacspace/extractive

Give an AI writer a compact brief, not the whole article — so it spends a fraction of the tokens. Deterministic extractive summarization (TextRank), key-facts extraction (numbers, money, percentages, dates, named entities), headline candidates and key phrases/hashtags — for English and Nepali (sentence splitting on the danda ।, decimal-safe). No LLM.

npm i @lacspace/extractive

import { brief, summarize, keyFacts } from "@lacspace/extractive";

const b = brief(longArticle);
// → {
//   summary: "…3–4 central sentences…",
//   headlineCandidates: ["…", "…"],
//   keyphrases: ["policy rate", "inflation", …],
//   hashtags: ["#policyrate", …],
//   keyFacts: { numbers, amounts, percentages, dates, entities },
//   sentenceCount: 18
// }
// Feed `b` to your writer instead of the full text — same facts, tiny prompt.

summarize(article, { maxSentences: 3 }).summary;   // extractive summary
keyFacts(article).percentages;                      // the hard figures to preserve

Why

Sending whole articles to an LLM for every post is the biggest token sink in a newsroom pipeline. extractive does the reading deterministically and hands the model a short, fact-anchored brief — the model only writes, it doesn't ingest.

API

• summarize(text, { maxSentences?, ratio?, lang? }) → { summary, sentences, ranked } — TextRank picks the most central sentences and returns them in original order. ranked is every sentence with its 0–1 score.
• keyFacts(text) → { numbers, amounts, percentages, dates, entities } — the figures and names a rewrite must not change (via @lacspace/factcheck-lite + @lacspace/keyphrase).
• headlineCandidates(text, { max? }) → string[] — the lead plus the shortest high-ranked sentences, trailing terminators stripped.
• brief(text, options?) → the bundle above, ready for a writer prompt.
• splitSentences(text) / tokenize(sentence) / textrank(tokenSets) — the building blocks, exported.
• describe() → a machine-readable capability + JSON-Schema command descriptor, so an AI "conductor" can pick options and drive the package without generating content.

All deterministic and Devanagari-aware. Reuses @lacspace/keyphrase, @lacspace/factcheck-lite and @lacspace/trend-detect (stopwords) rather than duplicating them.

Licence

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

No third-party names (1.1.0)

scrubSources(text, { outlets?, keep? }) → { text, removed, remaining, clean } strips credit lines (Source:, Photo:, स्रोत:, तस्बिर: …), datelines like (Reuters) -, and attribution phrases (according to the Kathmandu Post, कान्तिपुरका अनुसार, … अनलाइनखबरले जनाएको छ) for ~150 Nepali and international outlets and stock libraries, keeping the facts. Names it can't remove safely are listed in remaining so a validator can hold the post; mentionsOutlet(text) is the quick gate. Run it on the writer's input and again on the final copy.

Nepali case endings (1.2.0)

Outlet names are now matched with Nepali case endings attached (को/का/की/ले/मा/बाट/लाई/सँग/द्वारा/मार्फत …) and in Romanised form (Kantipur ko, Setopatima). New attribution patterns: कान्तिपुरको रिपोर्ट अनुसार, रातोपाटीमा प्रकाशित समाचार अनुसार, अनलाइनखबरले जनाएअनुसार, उनले कान्तिपुरलाई बताए → उनले बताए, सेतोपाटीका संवाददाता → संवाददाता, Kantipur ko report anusar. A trailing … बढेको अनलाइनखबरले जनाएको छ। now keeps its छ.

Outlet names that are also everyday words (AMBIGUOUS_OUTLETS: उज्यालो, शिलापत्र, नयाँ पत्रिका, Dawn …) are removed only inside credit lines, attributions and datelines. A bare word like नागरिक is never touched; only नागरिक दैनिक / नागरिक न्यूज count as outlets.

neutralize: true (or { ne, en }) swaps any mention that survives for a neutral noun and keeps the case ending: सेतोपाटीका कर्मचारी → सञ्चारमाध्यमका कर्मचारी. It is off by default, so survivors are reported in remaining instead.

1.2.1: an inflected noun after the outlet keeps its case ending: सेतोपाटीमा प्रकाशित लेखमा उल्लेख छ। → एक लेखमा उल्लेख छ।. This also covers समाचारमा, रिपोर्टमा, भिडियोमा, कार्यक्रममा, कान्तिपुरको रिपोर्टमा and …को समाचारले.