Compare Relevan with other search tools

This page shows where Relevan fits among the tools that give AI agents search over your own data. Each linked page compares Relevan with one vendor.

The facts about other vendors come from their public pages, as of October 2026. Each page lists its sources. A figure that a vendor publishes about its own product is marked as vendor-reported.

Four approaches

Teams that build agent search over their own data use one or more of these approaches:

Approach	What improves	Example vendors
Better models	The embedding model or the reranker model that scores documents. One model serves many customers.	ZeroEntropy
Integrated agent retrieval	A hosted pipeline that parses files, indexes them, and runs a search loop for the agent.	Mixedbread
Search infrastructure	The database that stores the index and runs the queries. Your team writes and tunes the ranking.	turbopuffer, Chroma, OpenSearch, Elasticsearch
Relevance tuning for each index	The retrieval and ranking configuration of one index, from the documents and the traffic of that index.	Relevan

A reranker model, from any vendor, reorders the candidates that retrieval returns. It cannot add a document that retrieval did not return. Relevan tunes the retrieval of each index, and it can also rerank the first page with a model of that index.

What Relevan tunes

Relevan runs each search in phases: it reads the query, enriches it, retrieves candidates from OpenSearch, and reranks the first page. Relevan keeps a separate configuration for each index. The configuration reaches two phases.

The retrieval phase decides which documents become candidates, and their order. For each index, the configuration sets:

The weight of each field.
The BM25 parameters k1 and b.
The fuzziness, and the match threshold for long queries.
Hybrid retrieval, which adds a vector query to the keyword query.
Entity matching, for example the names of places.

The rerank phase can use a learned model of the index. The model is a gradient-boosted tree model (LambdaMART). It reorders the first 30 hits of the first page. One extra OpenSearch query calculates the model inputs on those candidates, and no call goes to an external model.

How the configuration changes

When an index first holds 1,000 documents, Relevan reads a sample and sets a baseline field weight. A product catalog gets more weight on the name field. Prose gets less weight on the name field. Other data gets no change.
After that, Relevan can assess an index from a sample of its documents and from its search logs. The logs give the zero-result rate, the low-confidence rate, and the rate of query rewrites. The assessment gives a verdict and the evidence for each setting. Relevan applies the changes that the evidence supports.
Relevan trains the reranking model of an index from judged queries of that index. The labels can be human labels, LLM grades, or clicks. Relevan keeps a model only when its cross-validated gain in nDCG@10 is 0.005 or more.
Training does not run on each search. Relevan trains a model and checks it before the model changes the results.

The feedback endpoint records what happened after each search: a click, an LLM judgment, or a conversion. The glossary defines each type.

Click and feedback data

A learned ranking model needs data about what happened after each search. With a self-run engine, your team builds and runs this path. It has these parts:

A log of each search, with a query ID, the results, and the position of each result.
An endpoint or a client library that collects clicks, judgments, and conversions.
A stream or a queue that carries the events, with buffering, retries, and alarms.
Storage, partitioned by customer, index, and date.
A retention policy, and the removal of personal data.
A join of each event to the search that it is about.
A weight for each event type. A click is weaker evidence than a conversion.
A job that turns events into judgment lists, with a correction for position bias.
Training, evaluation, and deployment of the model, and a new run when the data or the mapping changes.

Relevan supplies parts 1 to 4 and part 6 through its API, and the retention window of part 5:

Each search response has a query ID. Relevan logs each search in the User Behavior Insights (UBI) query schema, with up to 50 results and their positions and scores.
The feedback endpoint takes a click, an LLM judgment, or a conversion. Each event refers to a query ID, a session ID, or a customer ID, and it can give the result ID and its position.
Relevan gives each event type a trust level. The caller cannot set it. The caller marks each event as production traffic or as an evaluation run.
Relevan hashes the customer ID before it logs a search.
A stream carries searches and feedback to storage, partitioned by organization, index, date, and event type.
Relevan keeps search and feedback events for 13 months by default.
A scorecard job reads the logged searches. It calculates the zero-result rate, the low-confidence rate, and the rate of query rewrites for an index.

One part is not automatic today. Relevan turns judged queries into training data in an offline step, with checks before a model goes live.

The cost argument

When retrieval misses the correct document, a reranker cannot recover it. The agent must search again, or the system must send more candidates to the reranker. Each extra search adds results for the agent to read, and each extra candidate adds reranker compute.

Relevan tunes retrieval for each index, so that the correct document can enter the candidates on the first search. The expected effects are fewer searches, fewer results read, and less reranker compute. Relevan has not measured these cost effects yet. The next section gives the evidence that exists.

Evidence

Reranking model, offline. With human labels and hybrid retrieval, the model of each index improved nDCG@10 on two public datasets. On WANDS, it went from 0.758 to 0.770 on 480 queries (p = 0.002). On Amazon ESCI, it went from 0.555 to 0.571 on 1,000 queries (p < 0.001). A model trained on other datasets gave no gain on three of four test sets.
Agent-built indexes, October 2026. Agents built an index on Relevan, Typesense, and Algolia with no human help, then tuned it with 100 labeled queries. On 380 held-out WANDS queries, Relevan went from 0.670 to 0.690 nDCG@10. Typesense reached 0.708 and Algolia reached 0.706. After tuning, the differences were not significant.
Homepage demo. The homepage shows a scripted episode against a real search function. The agent that reads the explanations in each Relevan response uses 2 searches in place of 3, and reads 6 results in place of 22. This difference comes from the explanations in the response. It does not come from learning.