OpenSearch Relevance Agent

A multi-agent system inside OpenSearch Dashboards that diagnoses and tunes search relevance through natural-language conversation, shipped experimentally in OpenSearch 3.6 as part of the OpenSearch Agent Server.


What It Does

Targets three stated obstacles to relevance work: ambiguous user queries, incomplete or noisy metadata, and moving a fix from a test environment to production. The pitch is that it “reduces relevance diagnosis from days to hours — no deep search expertise required”, with a human kept in the loop.

An orchestrator coordinates three specialised agents:

AgentRole
User Behavior AnalysisFinds relevance gaps from UBI data where available, query patterns where not
Hypothesis GeneratorTurns those findings into data-driven tuning strategies
EvaluatorTests the strategies against offline evaluation sets

UBI sharpens the analysis but is not required to start.

Architecture

  • Strands SDK — the agent framework it is built on, deployed into an existing OpenSearch environment.
  • AG-UI — the standard behind the Dashboards chat surface.
  • Model Context Protocol — every agent reaches the engine exclusively through the OpenSearch MCP server, characterised as a secure translator between the AI and the search engine.

The notable constraint: metric computation is offloaded from the LLM to deterministic tools, so numbers are measured rather than estimated by a model. The announcement cites “standard relevance metrics” and impact quantification via Search Relevance Workbench without naming specific measures.

Scope in 3.6

Supported: query-DSL-level optimizations — refining search fields, adjusting weights, tuning boost functions.

Roadmap: online interleaving tests in production, schema evolution recommendations, vector and hybrid parameter optimization, automated LTR training, and multi-platform data sources over MCP.

Articles