Answer Similarity Scorer

The createAnswerSimilarityScorer() function creates a scorer that evaluates how similar an agent's output is to a ground truth answer. This scorer is specifically designed for CI/CD testing scenarios where you have expected answers and want to ensure consistency over time.

Parameters

model:

LanguageModel

The language model used to evaluate semantic similarity between outputs and ground truth.

options:

AnswerSimilarityOptions

Configuration options for the scorer.

AnswerSimilarityOptions

requireGroundTruth:

boolean

= true

Whether to require ground truth for evaluation. If false, missing ground truth returns score 0.

semanticThreshold:

number

= 0.8

Weight for semantic matches vs exact matches (0-1).

exactMatchBonus:

number

= 0.2

Additional score bonus for exact matches (0-1).

missingPenalty:

number

= 0.15

Penalty per missing key concept from ground truth.

contradictionPenalty:

number

= 1.0

Penalty for contradictory information. High value ensures wrong answers score near 0.

extraInfoPenalty:

number

= 0.05

Mild penalty for extra information not present in ground truth (capped at 0.2).

scale:

number

= 1

Score scaling factor.

This function returns an instance of the MastraScorer class. The .run() method accepts the same input as other scorers (see the MastraScorer reference), but requires ground truth to be provided in the run object.

.run() Returns

runId:

string

The id of the run (optional).

score:

number

Similarity score between 0-1 (or 0-scale if custom scale used). Higher scores indicate better similarity to ground truth.

reason:

string

Human-readable explanation of the score with actionable feedback.

preprocessStepResult:

object

Extracted semantic units from output and ground truth.

analyzeStepResult:

object

Detailed analysis of matches, contradictions, and extra information.

preprocessPrompt:

string

The prompt used for semantic unit extraction.

analyzePrompt:

string

The prompt used for similarity analysis.

generateReasonPrompt:

string

The prompt used for generating the explanation.

Scoring Details

The scorer uses a multi-step process:

Extract: Breaks down output and ground truth into semantic units
Analyze: Compares units and identifies matches, contradictions, and gaps
Score: Calculates weighted similarity with penalties for contradictions
Reason: Generates human-readable explanation

Score calculation: max(0, base_score - contradiction_penalty - missing_penalty - extra_info_penalty) × scale

Examples

Usage with runExperiment

This scorer is designed for use with runExperiment for CI/CD testing:

import { runExperiment } from "@mastra/core/scores";
import { createAnswerSimilarityScorer } from "@mastra/evals/scorers/llm";

const scorer = createAnswerSimilarityScorer({ model });

await runExperiment({
  data: [
    {
      input: "What is the capital of France?",
      groundTruth: "Paris is the capital of France",
    },
  ],
  scorers: [scorer],
  target: myAgent,
  onItemComplete: ({ scorerResults }) => {
    // Assert similarity score meets threshold
    expect(scorerResults["Answer Similarity Scorer"].score).toBeGreaterThan(
      0.8,
    );
  },
});

Perfect similarity example

In this example, the agent's output semantically matches the ground truth perfectly.

src/example-perfect-similarity.ts
import { runExperiment } from "@mastra/core/scores";
import { createAnswerSimilarityScorer } from "@mastra/evals/scorers/llm";
import { myAgent } from "./agent";

const scorer = createAnswerSimilarityScorer({ model: "openai/gpt-4o-mini" });

const result = await runExperiment({
  data: [
    {
      input: "What is 2+2?",
      groundTruth: "4",
    },
  ],
  scorers: [scorer],
  target: myAgent,
});

console.log(result.scores);

Perfect similarity output

The output receives a perfect score because both the agent's answer and ground truth are identical.

{
  "Answer Similarity Scorer": {
    score: 1.0,
    reason: "The score is 1.0/1 because the output matches the ground truth exactly. The agent correctly provided the numerical answer. No improvements needed as the response is fully accurate."
  }
}

High semantic similarity example

In this example, the agent provides the same information as the ground truth but with different phrasing.

src/example-semantic-similarity.ts
import { runExperiment } from "@mastra/core/scores";
import { createAnswerSimilarityScorer } from "@mastra/evals/scorers/llm";
import { myAgent } from "./agent";

const scorer = createAnswerSimilarityScorer({ model: "openai/gpt-4o-mini" });

const result = await runExperiment({
  data: [
    {
      input: "What is the capital of France?",
      groundTruth: "The capital of France is Paris",
    },
  ],
  scorers: [scorer],
  target: myAgent,
});

console.log(result.scores);

High semantic similarity output

The output receives a high score because it conveys the same information with equivalent meaning.

{
  "Answer Similarity Scorer": {
    score: 0.9,
    reason: "The score is 0.9/1 because both answers convey the same information about Paris being the capital of France. The agent correctly identified the main fact with slightly different phrasing. Minor variation in structure but semantically equivalent."
  }
}

Partial similarity example

In this example, the agent's response is partially correct but missing key information.

src/example-partial-similarity.ts
import { runExperiment } from "@mastra/core/scores";
import { createAnswerSimilarityScorer } from "@mastra/evals/scorers/llm";
import { myAgent } from "./agent";

const scorer = createAnswerSimilarityScorer({ model: "openai/gpt-4o-mini" });

const result = await runExperiment({
  data: [
    {
      input: "What are the primary colors?",
      groundTruth: "The primary colors are red, blue, and yellow",
    },
  ],
  scorers: [scorer],
  target: myAgent,
});

console.log(result.scores);

Partial similarity output

The output receives a moderate score because it includes some correct information but is incomplete.

{
  "Answer Similarity Scorer": {
    score: 0.6,
    reason: "The score is 0.6/1 because the answer captures some key elements but is incomplete. The agent correctly identified red and blue as primary colors. However, it missed the critical color yellow, which is essential for a complete answer."
  }
}

Contradiction example

In this example, the agent provides factually incorrect information that contradicts the ground truth.

src/example-contradiction.ts
import { runExperiment } from "@mastra/core/scores";
import { createAnswerSimilarityScorer } from "@mastra/evals/scorers/llm";
import { myAgent } from "./agent";

const scorer = createAnswerSimilarityScorer({ model: "openai/gpt-4o-mini" });

const result = await runExperiment({
  data: [
    {
      input: "Who wrote Romeo and Juliet?",
      groundTruth: "William Shakespeare wrote Romeo and Juliet",
    },
  ],
  scorers: [scorer],
  target: myAgent,
});

console.log(result.scores);

Contradiction output

The output receives a very low score because it contains factually incorrect information.

{
  "Answer Similarity Scorer": {
    score: 0.0,
    reason: "The score is 0.0/1 because the output contains a critical error regarding authorship. The agent correctly identified the play title but incorrectly attributed it to Christopher Marlowe instead of William Shakespeare, which is a fundamental contradiction."
  }
}

CI/CD Integration example

Use the scorer in your test suites to ensure agent consistency over time:

src/ci-integration.test.ts
import { describe, it, expect } from "vitest";
import { runExperiment } from "@mastra/core/scores";
import { createAnswerSimilarityScorer } from "@mastra/evals/scorers/llm";
import { myAgent } from "./agent";

describe("Agent Consistency Tests", () => {
  const scorer = createAnswerSimilarityScorer({ model: "openai/gpt-4o-mini" });

  it("should provide accurate factual answers", async () => {
    const result = await runExperiment({
      data: [
        {
          input: "What is the speed of light?",
          groundTruth:
            "The speed of light in vacuum is 299,792,458 meters per second",
        },
        {
          input: "What is the capital of Japan?",
          groundTruth: "Tokyo is the capital of Japan",
        },
      ],
      scorers: [scorer],
      target: myAgent,
    });

    // Assert all answers meet similarity threshold
    expect(result.scores["Answer Similarity Scorer"].score).toBeGreaterThan(
      0.8,
    );
  });

  it("should maintain consistency across runs", async () => {
    const testData = {
      input: "Define machine learning",
      groundTruth:
        "Machine learning is a subset of AI that enables systems to learn and improve from experience",
    };

    // Run multiple times to check consistency
    const results = await Promise.all([
      runExperiment({ data: [testData], scorers: [scorer], target: myAgent }),
      runExperiment({ data: [testData], scorers: [scorer], target: myAgent }),
      runExperiment({ data: [testData], scorers: [scorer], target: myAgent }),
    ]);

    // Check that all runs produce similar scores (within 0.1 tolerance)
    const scores = results.map(
      (r) => r.scores["Answer Similarity Scorer"].score,
    );
    const maxDiff = Math.max(...scores) - Math.min(...scores);
    expect(maxDiff).toBeLessThan(0.1);
  });
});

Custom configuration example

Customize the scorer behavior for specific use cases:

src/custom-config.ts
import { runExperiment } from "@mastra/core/scores";
import { createAnswerSimilarityScorer } from "@mastra/evals/scorers/llm";
import { myAgent } from "./agent";

// Configure for strict exact matching with high scale
const strictScorer = createAnswerSimilarityScorer({
  model: "openai/gpt-4o-mini",
  options: {
    exactMatchBonus: 0.5, // Higher bonus for exact matches
    contradictionPenalty: 2.0, // Very strict on contradictions
    missingPenalty: 0.3, // Higher penalty for missing info
    scale: 10, // Score out of 10 instead of 1
  },
});

// Configure for lenient semantic matching
const lenientScorer = createAnswerSimilarityScorer({
  model: "openai/gpt-4o-mini",
  options: {
    semanticThreshold: 0.6, // Lower threshold for semantic matches
    contradictionPenalty: 0.5, // More forgiving on minor contradictions
    extraInfoPenalty: 0, // No penalty for extra information
    requireGroundTruth: false, // Allow missing ground truth
  },
});

const result = await runExperiment({
  data: [
    {
      input: "Explain photosynthesis",
      groundTruth:
        "Photosynthesis is the process by which plants convert light energy into chemical energy",
    },
  ],
  scorers: [strictScorer, lenientScorer],
  target: myAgent,
});

console.log("Strict scorer:", result.scores["Answer Similarity Scorer"].score); // Out of 10
console.log("Lenient scorer:", result.scores["Answer Similarity Scorer"].score); // Out of 1

Parameters​

model:

options:

AnswerSimilarityOptions​

requireGroundTruth:

semanticThreshold:

exactMatchBonus:

missingPenalty:

contradictionPenalty:

extraInfoPenalty:

scale:

.run() Returns​

runId:

score:

reason:

preprocessStepResult:

analyzeStepResult:

preprocessPrompt:

analyzePrompt:

generateReasonPrompt:

Scoring Details​

Examples​

Usage with runExperiment​

Perfect similarity example​

Perfect similarity output​

High semantic similarity example​

High semantic similarity output​

Partial similarity example​

Partial similarity output​

Contradiction example​

Contradiction output​

CI/CD Integration example​

Custom configuration example​

Parameters

AnswerSimilarityOptions

.run() Returns

Scoring Details

Examples

Usage with runExperiment

Perfect similarity example

Perfect similarity output

High semantic similarity example

High semantic similarity output

Partial similarity example

Partial similarity output

Contradiction example

Contradiction output

CI/CD Integration example

Custom configuration example