@mastra/core/dist/docs/references/reference-evals-run-evals.md

> Discover all available pages from the documentation index: https://mastra.ai/llms.txt

# runEvals

The `runEvals` function enables batch evaluation of agents and workflows by running multiple test cases against scorers concurrently. This is essential for systematic testing, performance analysis, and validation of AI systems.

## Usage example

```typescript
import { runEvals } from '@mastra/core/evals'
import { myAgent } from './agents/my-agent'
import { myScorer1, myScorer2 } from './scorers'

const result = await runEvals({
  target: myAgent,
  data: [
    { input: 'What is machine learning?' },
    { input: 'Explain neural networks' },
    { input: 'How does AI work?' },
  ],
  scorers: [myScorer1, myScorer2],
  targetOptions: { maxSteps: 5 },
  concurrency: 2,
  onItemComplete: ({ item, targetResult, scorerResults }) => {
    console.log(`Completed: ${item.input}`)
    console.log(`Scores:`, scorerResults)
  },
})

console.log(`Average scores:`, result.scores)
console.log(`Processed ${result.summary.totalItems} items`)
```

### Multi-turn evaluation

```typescript
import { runEvals } from '@mastra/core/evals'
import { checks } from '@mastra/evals/checks'
import { weatherAgent } from './agents/weather-agent'

const result = await runEvals({
  target: weatherAgent,
  data: [
    {
      inputs: [
        'What is the weather in Brooklyn?',
        'What about tomorrow?',
        'Compare the two forecasts.',
      ],
    },
  ],
  scorers: [checks.calledTool('get_weather', { times: 2 }), checks.includes('Brooklyn')],
})
```

### With gates and thresholds

```typescript
import { runEvals } from '@mastra/core/evals'
import { checks } from '@mastra/evals/checks'
import { faithfulnessScorer } from './scorers'

const result = await runEvals({
  target: myAgent,
  data: [{ input: 'What is the weather in Brooklyn?' }],
  gates: [checks.calledTool('get_weather'), checks.noToolErrors()],
  scorers: [{ scorer: faithfulnessScorer, threshold: 0.7 }, checks.includes('Brooklyn')],
})

result.verdict // 'passed' | 'scored' | 'failed'
result.gateResults // [{ id, passed, score }]
result.thresholdResults // [{ id, passed, averageScore, threshold }]
```

## Parameters

**target** (`Agent | Workflow`): The agent or workflow to evaluate.

**data** (`RunEvalsDataItem[]`): Array of test cases with input data and optional ground truth.

**scorers** (`ScorerEntry[] | AgentScorerConfig | WorkflowScorerConfig`): Scorers to use. Each entry is either a bare MastraScorer or { scorer, threshold } for threshold tracking. An AgentScorerConfig object separates agent-level and trajectory scorers. A WorkflowScorerConfig object specifies scorers for the workflow, individual steps, and trajectory. Optional when at least one gate is provided (gate-only runs).

**gates** (`MastraScorer[]`): Scorers that must score 1.0 for the run to pass. If any gate averages below 1.0 across data items, the verdict is failed. Gates run before regular scorers on each data item. When provided, scorers may be omitted.

**targetOptions** (`AgentExecutionOptions | WorkflowRunOptions`): Options forwarded to the target during execution. For agents: options passed to agent.generate() (e.g. maxSteps, modelSettings, instructions). For workflows: options passed to run.start() (e.g. perStep, outputOptions, initialState). For multi-turn agent runs (inputs/turns), runEvals generates and injects the shared thread and a resource, so memory.thread is optional; provide memory.resource to reuse a specific resource.

**concurrency** (`number`): Number of test cases to run concurrently. (Default: `1`)

**onItemComplete** (`function`): Callback function called after each test case completes. Receives item, target result, and scorer results.

## Data item structure

**input** (`string | string[] | CoreMessage[] | any`): Input data for the target. For agents: messages or strings. For workflows: workflow input data. Optional when inputs is provided.

**inputs** (`(string | string[] | CoreMessage[] | any)[]`): Multi-turn inputs. Each entry is one turn (same shape as input) sent sequentially to the agent on the same thread. Scorers see the accumulated output from all turns. Only supported for Agent targets. When provided, input can be omitted. Mutually exclusive with turns.

**turns** (`EvalTurn[]`): Multi-turn conversation with per-turn assertions. Each turn is an object { input, gates?, scorers? } sent sequentially on the same thread; its gates/scorers evaluate only that turn's input and output. Per-turn outcomes are reported in turnResults and folded into the overall verdict. Only supported for Agent targets. Mutually exclusive with input and inputs.

**groundTruth** (`any`): Expected or reference output for comparison during scoring.

**expectedTrajectory** (`TrajectoryExpectation`): Expected trajectory configuration for trajectory scoring. Includes expected steps, ordering, efficiency budgets, blacklists, and tool failure tolerance. Passed to trajectory scorers as run.expectedTrajectory. Overrides the static defaults in scorer constructors.

**requestContext** (`RequestContext`): Request Context to pass to the target during execution.

**tracingContext** (`TracingContext`): Tracing context for observability and debugging.

**startOptions** (`WorkflowRunOptions`): Per-item workflow run options (e.g. initialState, perStep, outputOptions). Merged on top of targetOptions, so per-item values take precedence. Only applicable when the target is a workflow.

## Agent scorer configuration

For agents, use `AgentScorerConfig` to separate agent-level scorers from trajectory scorers:

**agent** (`MastraScorer[]`): Scorers that receive the raw agent output (MastraDBMessage\[]). Use for evaluating response quality, content, etc.

**trajectory** (`MastraScorer[]`): Scorers that receive a pre-extracted Trajectory object. When storage is configured, the pipeline extracts a hierarchical trajectory from observability traces (including nested tool calls and model generations). Otherwise, it falls back to extracting tool calls from agent messages.

## Workflow scorer configuration

For workflows, use `WorkflowScorerConfig` to specify scorers at different levels:

**workflow** (`MastraScorer[]`): Scorers to evaluate the entire workflow output.

**steps** (`Record<string, MastraScorer[]>`): Object mapping step IDs to arrays of scorers for evaluating individual step outputs.

**trajectory** (`MastraScorer[]`): Scorers that receive a pre-extracted Trajectory from the workflow execution. When storage is configured, the pipeline extracts a hierarchical trajectory from observability traces (including nested agent runs and tool calls within workflow steps). Otherwise, it falls back to extracting step results from the workflow output.

## Returns

**scores** (`Record<string, any>`): Average scores across all test cases, organized by scorer name.

**summary** (`object`): Summary information about the experiment execution.

**summary.totalItems** (`number`): Total number of test cases processed.

**verdict** (`'passed' | 'scored' | 'failed'`): Present when gates or threshold-bearing scorers are provided. passed = all gates and thresholds met. scored = gates passed but a threshold was missed. failed = at least one gate did not score 1.0.

**gateResults** (`GateResult[]`): Per-gate results averaged across all data items. Each entry has id, passed (boolean), and score (0–1).

**thresholdResults** (`ThresholdResult[]`): Per-threshold-scorer results averaged across all data items. Each entry has id, passed, averageScore, and threshold.

**turnResults** (`TurnResult[]`): Present when any data item uses turns. Each entry has index (zero-based turn), optional gateResults, thresholdResults, and scores (bare-scorer averages keyed by scorer id), aggregated by turn index across data items.

## EvalTurn

A single turn in a `turns` array. Its `gates`/`scorers` evaluate only that turn's input and output:

**input** (`string | string[] | CoreMessage[] | any`): The input sent to the agent for this turn.

**gates** (`MastraScorer[]`): Gates that must score 1.0 for this turn. A failing turn gate makes the overall verdict failed.

**scorers** (`ScorerEntry[]`): Scorers (optionally with thresholds) evaluated against this turn only. A missed per-turn threshold (with gates passing) makes the verdict scored.

## ScorerEntry

A scorer entry in the `scorers` array can be either a bare scorer or a scorer with a threshold:

**scorer** (`MastraScorer`): The scorer instance.

**threshold** (`number | { min?: number; max?: number }`): A number implies minimum threshold (score at or above passes). Use { min, max } for range-based checks — e.g. { max: 0.3 } for scorers like hallucination where a high score is bad. Both min and max must be between 0 and 1.

## Examples

### Gates and verdict

Use `gates` for hard pass/fail requirements and `{ scorer, threshold }` for tracked quality metrics:

```typescript
import { runEvals } from '@mastra/core/evals'
import { checks } from '@mastra/evals/checks'

const result = await runEvals({
  target: weatherAgent,
  data: [{ input: 'What is the weather in Brooklyn?' }],
  gates: [checks.calledTool('get_weather'), checks.noToolErrors()],
  scorers: [
    { scorer: faithfulnessScorer, threshold: 0.7 }, // min threshold (number shorthand)
    { scorer: hallucinationScorer, threshold: { max: 0.3 } }, // max threshold (high = bad)
    { scorer: toneScorer, threshold: { min: 0.5, max: 0.9 } }, // range threshold
    checks.includes('Brooklyn'), // bare scorer, no threshold
  ],
})

if (result.verdict === 'failed') {
  console.log(
    'Gate failures:',
    result.gateResults?.filter(g => !g.passed),
  )
} else if (result.verdict === 'scored') {
  console.log(
    'Threshold misses:',
    result.thresholdResults?.filter(t => !t.passed),
  )
}
```

### Agent Evaluation

```typescript
import { createScorer, runEvals } from '@mastra/core/evals'

const myScorer = createScorer({
  id: 'my-scorer',
  description: "Check if Agent's response contains ground truth",
  type: 'agent',
}).generateScore(({ run }) => {
  const response = run.output[0]?.content || ''
  const expectedResponse = run.groundTruth
  return response.includes(expectedResponse) ? 1 : 0
})

const result = await runEvals({
  target: chatAgent,
  data: [
    {
      input: 'What is AI?',
      groundTruth: 'AI is a field of computer science that creates intelligent machines.',
    },
    {
      input: 'How does machine learning work?',
      groundTruth: 'Machine learning uses algorithms to learn patterns from data.',
    },
  ],
  scorers: [relevancyScorer],
  concurrency: 3,
})
```

### Agent trajectory evaluation

Use `AgentScorerConfig` to evaluate both the agent response and its tool-calling trajectory:

```typescript
import { runEvals } from '@mastra/core/evals'
import { createTrajectoryAccuracyScorerCode } from '@mastra/evals/scorers/code/trajectory'

const trajectoryScorer = createTrajectoryAccuracyScorerCode()

const result = await runEvals({
  target: chatAgent,
  data: [
    {
      input: 'What is the weather in London?',
      expectedTrajectory: {
        steps: [{ stepType: 'tool_call', name: 'weatherTool' }],
      },
    },
  ],
  scorers: {
    // agent: [responseQualityScorer], // Optional: add agent-level scorers
    trajectory: [trajectoryScorer],
  },
})

// result.scores.agent — average agent-level scores
// result.scores.trajectory — average trajectory scores
```

### Agent with `targetOptions`

Pass execution options like `maxSteps` or `modelSettings` to customize agent behavior during evaluation:

```typescript
const result = await runEvals({
  target: chatAgent,
  data: [{ input: 'Summarize this article' }, { input: 'Translate to French' }],
  scorers: [relevancyScorer],
  targetOptions: {
    maxSteps: 5,
    modelSettings: { temperature: 0 },
  },
})
```

### Workflow Evaluation

```typescript
const workflowResult = await runEvals({
  target: myWorkflow,
  data: [
    { input: { query: 'Process this data', priority: 'high' } },
    { input: { query: 'Another task', priority: 'low' } },
  ],
  scorers: {
    workflow: [outputQualityScorer],
    steps: {
      'validation-step': [validationScorer],
      'processing-step': [processingScorer],
    },
  },
  onItemComplete: ({ item, targetResult, scorerResults }) => {
    console.log(`Workflow completed for: ${item.inputData.query}`)
    if (scorerResults.workflow) {
      console.log('Workflow scores:', scorerResults.workflow)
    }
    if (scorerResults.steps) {
      console.log('Step scores:', scorerResults.steps)
    }
  },
})
```

### Workflow trajectory evaluation

Add trajectory scoring to workflow evaluations to validate step execution order:

```typescript
const workflowResult = await runEvals({
  target: myWorkflow,
  data: [
    {
      input: { query: 'Process this data' },
      expectedTrajectory: {
        steps: [
          { stepType: 'workflow_step', name: 'validate' },
          { stepType: 'workflow_step', name: 'process' },
          { stepType: 'workflow_step', name: 'output' },
        ],
      },
    },
  ],
  scorers: {
    workflow: [outputQualityScorer],
    steps: {
      validate: [validationScorer],
    },
    trajectory: [trajectoryScorer],
  },
})

// result.scores.trajectory — workflow trajectory scores
```

### Workflow with per-item `startOptions`

Use `startOptions` on individual data items to customize each workflow run. Per-item values take precedence over `targetOptions`:

```typescript
const result = await runEvals({
  target: myWorkflow,
  data: [
    {
      input: { query: 'hello' },
      startOptions: { initialState: { counter: 1 } },
    },
    {
      input: { query: 'world' },
      startOptions: { initialState: { counter: 2 } },
    },
  ],
  scorers: [outputQualityScorer],
  targetOptions: { perStep: true },
})
```

### Multi-turn conversation evaluation

Use `inputs` to send sequential turns on a shared thread. Scorers see the accumulated output from all turns:

```typescript
const result = await runEvals({
  target: chatAgent,
  data: [
    {
      inputs: ['My favorite city is Brooklyn.', 'What is the weather in my favorite city?'],
    },
  ],
  gates: [checks.calledTool('get_weather')],
  scorers: [{ scorer: checks.similarity('Brooklyn weather forecast'), threshold: 0.5 }],
})

// result.verdict: 'passed' | 'scored' | 'failed'
```

Each turn runs `agent.generate()` with the same `threadId`, so the agent sees the full conversation history. `runEvals` also injects a `resourceId` (Mastra memory scopes messages by resource + thread), defaulting it to the generated thread. Pass `targetOptions.memory.resource` to pin a specific one. Cross-turn recall requires the agent to have a memory store configured. Otherwise turns run in isolation. Mix single-turn (`input`) and multi-turn (`inputs`) items in the same `data` array. When using `inputs`, `input` can be omitted.

Scoring uses the accumulated output from all turns as `run.output`, but only the first turn as `run.input`. Prefer output-based scorers (`checks.includes`, `checks.calledTool`, `checks.similarity`) for multi-turn. Input-relative scorers (e.g. faithfulness) only see the first turn's input. Trajectory scorers that read from the trace (`AgentScorerConfig.trajectory`) resolve against the last turn's span. Tool-call checks that read `run.output` (like `checks.calledTool`) still see every turn.

### Per-turn assertions

Use `turns` to attach `gates`/`scorers` to individual turns. Each per-turn assertion sees only that turn's input and output, so a later-turn regression can't be hidden by an earlier turn:

```typescript
const result = await runEvals({
  target: chatAgent,
  data: [
    {
      turns: [
        {
          input: 'What is the weather in Brooklyn?',
          gates: [checks.calledTool('get_weather')],
        },
        {
          input: 'What about tomorrow?',
          gates: [checks.calledTool('get_weather')], // must call again this turn
          scorers: [{ scorer: checks.similarity('tomorrow forecast'), threshold: 0.5 }],
        },
      ],
    },
  ],
})

result.verdict // folds in per-turn gate/threshold outcomes
result.turnResults // [{ index, gateResults, thresholdResults, scores }]
```

Per-turn gates/scorers evaluate only that turn (`run.input`/`run.output` are that turn's). A failing turn gate makes the verdict `failed`. A missed turn threshold (gates passing) makes it `scored`. Top-level `scorers`/`gates` still score the accumulated conversation as a whole. `turns` is Agent-only and can't be combined with `input` or `inputs`.

## Related

- [Multi-turn evals](https://mastra.ai/docs/evals/multi-turn): Conceptual guide to multi-turn evaluation
- [Gates and Verdicts](https://mastra.ai/docs/evals/gates-and-verdicts): Conceptual guide to severity semantics
- [Quick Checks](https://mastra.ai/reference/evals/checks): Zero-LLM composable micro-scorers
- [createScorer()](https://mastra.ai/reference/evals/create-scorer): Create custom scorers for experiments
- [MastraScorer](https://mastra.ai/reference/evals/mastra-scorer): Learn about scorer structure and methods
- [Trajectory Accuracy](https://mastra.ai/reference/evals/trajectory-accuracy): Built-in trajectory evaluation scorers
- [Scorer Utilities](https://mastra.ai/reference/evals/scorer-utils): Helper functions for extracting trajectory data
- [Custom Scorers](https://mastra.ai/docs/evals/custom-scorers): Guide to building evaluation logic
- [Scorers Overview](https://mastra.ai/docs/evals/overview): Understanding scorer concepts