Retrieval Relevance

Overview

The Retrieval Relevance evaluator determines whether the external information retrieved during a step is relevant to the request it was meant to serve. It is source-agnostic: the retrieved information may come from a vector-database / semantic search, a tool or function call, an MCP server, a web search, or a database query. It scores the retrieved information as a whole — holistically, per retrieval step — against the request.

When to Use

Use the Retrieval Relevance evaluator when you need to:

Diagnose RAG quality - Check whether retrieved documents actually bear on the user’s question
Evaluate tool- and MCP-based retrieval - Judge whether a tool call, MCP query, or web search returned information relevant to the request, not just whether it succeeded
Compare retrieval strategies - Measure the relevance of what a retriever, reranker, or agent surfaces across different pipelines

This evaluator judges the retrieved information against the request — it is independent of any final answer. To judge whether the answer is grounded in the context, use the Faithfulness Evaluator.

Supported Levels

The level of an evaluator determines the scope of the evaluation in OpenTelemetry terms. Some evaluations are applicable to individual spans, some to full traces or sessions, and some are applicable at multiple levels.

Level	Supported	Notes
Span	Yes	Apply to the span that performed the retrieval. Provide the request as `input` and the retrieved information as `context`.

Relevant span kinds: RETRIEVER and RERANKER spans, TOOL spans that return information (knowledge base, web search, MCP, SQL), and LLM spans that retrieved information themselves (e.g. server-side / native web search, where results are embedded in the message content). Action tools with side effects (e.g. send_email) and pure LLM turns are not retrieval steps and should not be scored.

Input Requirements

The Retrieval Relevance evaluator requires two inputs:

Field	Type	Description
`input`	`string`	The request the retrieval was serving
`context`	`string`	The external information retrieved during the step, with all returned items joined together

Formatting Tips

For best results:

Use the user’s request as input. For tool and SQL steps, prefer the user’s request (e.g. the trace root’s input.value) over a reformulated tool argument or a generated SQL query.

Join multiple retrieved items into a single context string with clear separators (see Input Mapping below):

Our return policy allows returns within 30 days of purchase.

Refunds are processed within 5 business days.

Use human-readable strings rather than raw JSON where possible.

Output Interpretation

The evaluator returns a Score object with the following properties:

Property	Value	Description
`label`	`"relevant"` or `"irrelevant"`	Classification result
`score`	`1.0` or `0.0`	Numeric score (1.0 = relevant, 0.0 = irrelevant)
`explanation`	`string`	LLM-generated reasoning for the classification
`direction`	`"maximize"`	Higher scores are better
`metadata`	`object`	Additional information such as the model name. When tracing is enabled, includes the `trace_id` for the evaluation.

Interpretation:

Relevant (1.0): The retrieved information contains content that materially helps address the request. If any meaningful part of the retrieved information helps, the step is relevant — even when the set is partial or mixed with unrelated material.
Irrelevant (0.0): The retrieved information does not help address the request — it is off-topic, about a different entity or time period, only tangentially related, empty, or an error.

Usage Examples

Python
TypeScript

from phoenix.evals import LLM
from phoenix.evals.metrics import RetrievalRelevanceEvaluator

# Initialize the LLM client
llm = LLM(provider="openai", model="gpt-4o")

# Create the evaluator
relevance_eval = RetrievalRelevanceEvaluator(llm=llm)

# Inspect the evaluator's requirements
print(relevance_eval.describe())

# Evaluate a single example
eval_input = {
    "input": "What is the capital of France?",
    "context": "Paris is the capital and largest city of France."
}

scores = relevance_eval.evaluate(eval_input)
print(scores[0])
# Score(name='retrieval_relevance', score=1.0, label='relevant', ...)

import { createRetrievalRelevanceEvaluator } from "@arizeai/phoenix-evals";
import { openai } from "@ai-sdk/openai";

// Create the evaluator
const retrievalRelevanceEvaluator = createRetrievalRelevanceEvaluator({
  model: openai("gpt-4o"),
});

// Evaluate an example
const result = await retrievalRelevanceEvaluator.evaluate({
  input: "What is the capital of France?",
  context: "Paris is the capital and largest city of France.",
});

console.log(result);
// { score: 1, label: "relevant", explanation: "..." }

Using Input Mapping

When your data has different field names or requires transformation, use input mapping. This is especially useful for combining multiple retrieved items into a single context string.

Python
TypeScript

from phoenix.evals import LLM
from phoenix.evals.metrics import RetrievalRelevanceEvaluator

llm = LLM(provider="openai", model="gpt-4o")
relevance_eval = RetrievalRelevanceEvaluator(llm=llm)

# Example with a query and multiple retrieved documents
eval_input = {
    "query": "What is the return policy?",
    "retrieved": {
        "documents": [
            "Our return policy allows returns within 30 days.",
            "Refunds are processed within 5 business days."
        ]
    }
}

# Use input mapping with a lambda to concatenate documents
input_mapping = {
    "input": "query",
    "context": lambda x: "\n\n".join(x["retrieved"]["documents"])
}

scores = relevance_eval.evaluate(eval_input, input_mapping)

For more details on input mapping options, see Input Mapping.

import { bindEvaluator, createRetrievalRelevanceEvaluator } from "@arizeai/phoenix-evals";
import { openai } from "@ai-sdk/openai";

const retrievalRelevanceEvaluator = createRetrievalRelevanceEvaluator({
  model: openai("gpt-4o"),
});

// Bind with input mapping for different field names
const boundEvaluator = bindEvaluator(retrievalRelevanceEvaluator, {
  inputMapping: {
    input: "query",
    context: (data) => data.documents.join("\n\n"),
  },
});

const result = await boundEvaluator.evaluate({
  query: "What is the return policy?",
  documents: [
    "Our return policy allows returns within 30 days.",
    "Refunds are processed within 5 business days."
  ],
});

For more details on input mapping options, see Input Mapping.

Configuration

For LLM client configuration options, see Configuring the LLM.

Viewing and Modifying the Prompt

You can view the latest versions of our prompt templates on GitHub. The evaluators are designed to work well in a variety of contexts, but we highly recommend modifying the prompt to be more specific to your use case. Feel free to adapt them.

Python
TypeScript

from phoenix.evals.metrics import RetrievalRelevanceEvaluator
from phoenix.evals import LLM, ClassificationEvaluator

llm = LLM(provider="openai", model="gpt-4o")
evaluator = RetrievalRelevanceEvaluator(llm=llm)

# View the prompt template
print(evaluator.prompt_template)

# Create a custom evaluator based on the built-in template
custom_evaluator = ClassificationEvaluator(
    name="retrieval_relevance",
    prompt_template=evaluator.prompt_template,  # Modify as needed
    llm=llm,
    choices={"relevant": 1.0, "irrelevant": 0.0},
    direction="maximize",
)

import { RETRIEVAL_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG, createRetrievalRelevanceEvaluator } from "@arizeai/phoenix-evals";
import { openai } from "@ai-sdk/openai";

// View the prompt template
console.log(RETRIEVAL_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG.template);

// Create a custom evaluator with a modified template
const customEvaluator = createRetrievalRelevanceEvaluator({
  model: openai("gpt-4o"),
  promptTemplate: RETRIEVAL_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG.template, // Modify as needed
});

Using with Phoenix

Evaluating Traces

Run evaluations on traces collected in Phoenix and log results as annotations:

Running Experiments

Use the Retrieval Relevance evaluator in Phoenix experiments:

Using Evaluators in Experiments

API Reference

Python: RetrievalRelevanceEvaluator
TypeScript: createRetrievalRelevanceEvaluator

Faithfulness Evaluator - Evaluate whether a response is grounded in the retrieved context
Correctness Evaluator - Evaluate whether an answer is correct

Get Started

Tracing

Evaluation

Datasets & Experiments

Prompts

Settings

Concepts

Resources

Retrieval Relevance

Overview

When to Use

Supported Levels

Input Requirements

Formatting Tips

Output Interpretation

Usage Examples

Using Input Mapping

Configuration

Viewing and Modifying the Prompt

Using with Phoenix

Evaluating Traces

Running Experiments

API Reference

​Overview

​When to Use

​Supported Levels

​Input Requirements

​Formatting Tips

​Output Interpretation

​Usage Examples

​Using Input Mapping

​Configuration

​Viewing and Modifying the Prompt

​Using with Phoenix

​Evaluating Traces

​Running Experiments

​API Reference

​Related

Overview

When to Use

Supported Levels

Input Requirements

Formatting Tips

Output Interpretation

Usage Examples

Using Input Mapping

Configuration

Viewing and Modifying the Prompt

Using with Phoenix

Evaluating Traces

Running Experiments

API Reference

Related