zerank-2

Planned

zeroentropy/zerank-2

zerank-2 is the reranker Notion AI ran in production before acquiring its maker. A Qwen3-4B cross-encoder that rescores retrieved documents against the query, it tops published NDCG@10 comparisons against commercial rerankers, with particular strength in finance, legal, medical and code retrieval. Released Apache-2.0.

This model is on the roadmap but is not being served yet, so it will return model_not_found from the API. It appears here because we publish what we intend to run as openly as what we already do — the catalogue should tell you where the platform is going, not just where it is.

What a reranker does

A reranker does not chat and does not generate text. You give it one question and a list of documents, and it scores how well each document answers that question — reading the pair together, rather than comparing two separate embeddings. That is why it catches relevance a vector search misses, and why it only makes sense on a shortlist: it is far more accurate per comparison, and far more expensive than a vector lookup.

Where it goes in a search pipeline

1. Retrieve broadly. Your vector search or keyword index returns the top 50–100 candidates. Fast, cheap, approximate.
2. Rerank. Send the query and those candidates here. You get them back ordered by true relevance.
3. Use the top few. Put the best 3–5 in your prompt instead of the best 20. Better answers, a shorter prompt, and a smaller bill from whichever model writes the reply.

What people use it for

RAG that cites the right page — the most common use. Retrieval finds twenty plausible chunks; the reranker picks the ones that actually contain the answer.
Support and helpdesk search — matching a customer’s words to articles written in different words.
Agent tool results — trimming a large search or file-read result down to what is worth spending context on.
Site and catalogue search — reordering results before they reach the page.

Reading the scores

Every result carries a relevance_score between 0 and 1 and an index pointing back into the array you sent, so you can reorder your own objects without round-tripping the text. Results come back sorted, best first. Scores are calibrated to the model author’s own published scale, so thresholds you tuned elsewhere carry over unchanged.

Comparing options first? Our reranking API covers what the job costs elsewhere and how the units differ.