Rerank

Rerank runs the same document through three different summarization methods (TextRank, LexRank, and BART) and lets you compare their output side by side instead of committing to whichever one a library happens to default to. TextRank and LexRank aren't wrappers around an existing package; they're implemented directly in TypeScript and run entirely client-side, scoring sentences by how similar they are to every other sentence in the document.

Both build a sentence-similarity matrix from cosine similarity over tokenized sentences, then treat it as a graph and rank nodes the way PageRank ranks web pages: a sentence is "important" if it's similar to other important sentences. TextRank's version runs the classic damped iteration to convergence:

const scores = new Array(sentences.length).fill(1);
const damping = 0.85;
const iterations = 50;

for (let iter = 0; iter < iterations; iter++) {
  const newScores = [...scores];
  for (let i = 0; i < sentences.length; i++) {
    let sum = 0;
    for (let j = 0; j < sentences.length; j++) {
      if (i !== j) {
        const totalSim = similarityMatrix[j].reduce((a, b) => a + b, 0);
        if (totalSim > 0) {
          sum += (similarityMatrix[j][i] / totalSim) * scores[j];
        }
      }
    }
    newScores[i] = (1 - damping) + damping * sum;
  }
  scores.splice(0, scores.length, ...newScores);
}

LexRank instead thresholds the similarity matrix before row-normalizing it into a proper transition matrix, which is the main mathematical difference between the two approaches. BART, being an actual transformer model, calls out to a hosted inference API and falls back to an extractive method if that request fails, so a summary always comes back, even offline.