Skip to Content
Relevan Beta is here 🎉
ResourcesCompareRelevan vs ZeroEntropy

Relevan vs ZeroEntropy

ZeroEntropy made reranker and embedding models, and a hosted search engine called zsearch. Notion acquired ZeroEntropy in July 2026. The hosted products closed on September 4, 2026, and the models are now open weights under Apache 2.0.

This page helps two readers: a team that compares a reranker model with Relevan, and a team that must move off zsearch. The facts about ZeroEntropy come from its public pages and from Notion, as of October 2026. The sources are at the end of this page.

Summary

ZeroEntropyRelevan
ApproachBetter modelsRelevance tuning for each index
StatusHosted products closed on September 4, 2026. New signups are off.Available
Main productThe zerank-2 reranker and the zembed-1 embedding modelA managed search index with an HTTP API
What improves the resultsOne model that serves all users, or a fine-tuned custom modelThe retrieval and ranking configuration of each index
How to run it nowSelf-host the open weights, or use an inference providerRelevan runs the index

What ZeroEntropy does

  • zerank-2 is a reranker. It scores the documents that you send with a query, and returns them in a new order.
  • zembed-1 is an embedding model for vector retrieval.
  • ZeroEntropy also fine-tuned custom models for customers, for example models that rewrite queries.
  • zsearch was a hosted search engine with collections, documents, and metadata filters.

On July 24, 2026, ZeroEntropy announced that Notion acquired it. The announcement says that all products stay supported until September 4, 2026, and are then sunset. It also says that new signups are off, and that the models are now Apache 2.0. ZeroEntropy publishes a migration guide.

Notion reports that zerank-2 made its unified search up to 30% faster, and cut the latency of its reranking step by 85%. These figures are customer-reported.

What Relevan does

Relevan is a managed search engine for your own records, such as products, profiles, and documents. You send JSON records and a mapping. Relevan runs the OpenSearch index, and agents search it through an HTTP API.

  • Each search response tells the agent why each result matched, and suggests what to try next.
  • Relevan tunes retrieval and ranking for each index. Compare Relevan with other search tools explains what Relevan tunes, and when.
  • The feedback endpoint records clicks, LLM judgments, and conversions for each search.
  • Relevan prices by plan, from a free plan. See pricing.

The main difference

A reranker improves the order of the candidates that it receives. It cannot add a document that retrieval did not return. Relevan tunes the retrieval of each index, so that more correct documents enter the candidates. The Relevan reranking model then reorders the first 30 hits with one extra query inside the cluster.

zerank-2 is one model for all users. The Relevan reranking model of an index is trained on the judged queries of that index. A larger general model can score text with more skill. A model of one index can learn the signals of that index, but it needs many judged queries to help.

When to choose the ZeroEntropy models

  • You already have a retrieval pipeline, and you want a stronger reranker or embedding model in it.
  • You can run models on your own GPUs, or you have an inference provider that serves the open weights.
  • You want an open-weight model under Apache 2.0, with no vendor contract.

When to choose Relevan

  • You used zsearch, and you need a hosted search engine in its place.
  • You do not want to run model inference or a search cluster.
  • You want retrieval and ranking tuned for each index.
  • Your agents need an explanation of each match and a suggestion for the next search.

Use both

You can rerank Relevan results with zerank-2 in your own code. Send the query and the text of the Relevan results to the model, and use the new order.

Sources

Read on October 29, 2026:

Last updated on
Relevan vs turbopufferData sources