Hybrid search

Dense embeddings struggle with rare tokens — product IDs, error codes, proper nouns. Sparse retrievers like BM25 nail those but miss paraphrases. Combining the two reliably outperforms either alone.

function rrf(rankings: number[][], k = 60): Map<number, number> {
  const scores = new Map<number, number>();
  for (const ranking of rankings) {
    ranking.forEach((id, i) => {
      scores.set(id, (scores.get(id) ?? 0) + 1 / (k + i + 1));
    });
  }
  return scores;
}

RRF requires no score calibration between retrievers — it only uses ranks. It's the default fusion in Weaviate, Qdrant and OpenSearch.

After fusion, pass the top-50 results through a cross-encoder (Cohere Rerank, BGE-reranker) to re-score with full attention between query and document. Latency cost is real but quality lift is the single biggest win in modern RAG.