Glossary

This page defines the terms that Relevan uses in its product, its API, and its documentation.

Learning loop

The learning loop is the cycle that makes Relevan search better from use:

1. An agent or a user searches an index.
2. The caller sends feedback about what happened after the search.
3. Relevan updates the ranking model of the index from the collected feedback, on a schedule.
4. The next search uses the updated ranking.

The ranking belongs to the index. Each index improves from its own traffic. The learning loop is on the Learn plan and above. See pricing.

Agent search

Agent search is search that an AI agent sends, in place of a person. An agent sends and reformulates queries more frequently than a person does, and it can read a structured response. Relevan tunes its responses for this behavior. A person can also search a Relevan index.

Click

A click is a feedback event: a real user picked a result. It is a real user action, but it does not confirm that the result met the need.

Conversion

A conversion is a feedback event for a closing event, such as a purchase or a hire. It confirms that the result was correct. It is the strongest feedback signal. Send a short label, such as purchase, with each conversion.

Feedback

Feedback is a report of what happened after a search. A caller sends it to POST /v1-beta/{orgId}/indexes/{indexName}/feedback. There are three types: click, LLM judgment, and conversion. Feedback trains the ranking through the learning loop. The feedback skill gives the steps.

Field use

A field use is a query capability that a mapping gives to a field. The values are searchable, filterable, sortable, aggregatable, autocompletable, and highlightable.

Index skill doc

The index skill doc is a Markdown document for one index: GET /v1/{orgId}/indexes/{indexName}/skill. Relevan generates it from the live mapping of the index. It gives the fields of the index and tells an agent how to search it. Fetch it again after a mapping change.

LLM judgment

An LLM judgment is a feedback event: an agent judged a result relevant, and no person confirmed it. It is the weakest feedback signal. An agent sends llm_judgment for each result that it uses.

Managed index

A managed index is a search index that Relevan operates for you. Relevan runs the indexing, the tuning, the upgrades, and the uptime. There is no cluster to size, shard, or upgrade.

Mapping

A mapping tells Relevan how to read the documents in an index. It gives each field a field use, and where needed a kind and a language. Relevan rejects a document until the index has a mapping. The mapping skill gives the steps.

Query ID

The query ID identifies one search. Each search response carries it in searchContext.queryId. Send it back with the feedback for that search, so Relevan can connect the feedback to the search.

Ranking signal

A ranking signal is a value in a document that shows which result is better, such as a price, a sales count, a stock level, or the date of the last activity. Relevan can use a signal only if the signal is a field in the index. The ranking signals skill tells you how to find the signals.

Search context

The search context is the searchContext object in each search response. It carries the query ID, the session ID, and suggestions that tell an agent how to change a query that gave poor results.

Session

A session groups the searches of one conversation. Each search response carries a searchContext.sessionId. Send it back on the next search of the same conversation. Omit it to start a new session.

Skill

A skill is a Markdown file that tells an agent how to do one task with Relevan. The API serves each skill at /skills/{name}/SKILL.md. Start with the setup skill.

Tuned search

Tuned search finds results by meaning, and also by exact words. Each index has tuned search on each plan.