text stringlengths 0 59.1k |
|---|
type: "agent", |
}) |
.score(({ payload, params }) => { |
const keyword = params.keyword as string; |
const matched = payload.output?.toLowerCase().includes(keyword.toLowerCase()); |
return { score: matched ? 1 : 0 }; |
}) |
.build(); |
// Usage: |
scorers: { |
keyword: { |
scorer: keywordScorer, |
params: { keyword: "refund" }, |
}, |
} |
``` |
## VoltOps Integration |
When a VoltOps client is configured globally, live scorer results are forwarded automatically: |
```ts |
import VoltAgent, { Agent, VoltAgentObservability } from "@voltagent/core"; |
import { VoltOpsClient } from "@voltagent/sdk"; |
const voltOpsClient = new VoltOpsClient({ |
publicKey: process.env.VOLTAGENT_PUBLIC_KEY, |
secretKey: process.env.VOLTAGENT_SECRET_KEY, |
}); |
const observability = new VoltAgentObservability(); |
new VoltAgent({ |
agents: { support: agent }, |
observability, |
voltOpsClient, // enables automatic forwarding |
}); |
``` |
The framework creates evaluation runs, registers scorers, appends results, and finalizes summaries. Each batch of scores (per agent interaction) becomes a separate run in VoltOps. |
## Sampling Strategies |
### Ratio Sampling |
Sample a percentage of interactions: |
```ts |
sampling: { type: "ratio", rate: 0.1 } // 10% of traffic |
``` |
Use for high-volume agents where scoring every interaction is expensive. |
### Count Sampling |
Sample every Nth interaction: |
```ts |
sampling: { type: "count", rate: 100 } // every 100th interaction |
``` |
Use when you need predictable sampling intervals or rate-limiting. |
### Per-Scorer Sampling |
Override sampling for specific scorers: |
```ts |
eval: { |
sampling: { type: "ratio", rate: 1 }, // default: score all |
scorers: { |
moderation: { |
scorer: moderationScorer, |
sampling: { type: "ratio", rate: 1 }, // always run moderation |
}, |
helpfulness: { |
scorer: helpfulnessScorer, |
sampling: { type: "ratio", rate: 0.05 }, // 5% for expensive LLM judge |
}, |
}, |
} |
``` |
## Error Handling |
If a scorer throws an exception, the result is marked `status: "error"` and the error message is captured in `errorMessage`. Other scorers continue executing. |
```ts |
.score(({ payload, params }) => { |
if (!params.keyword) { |
throw new Error("keyword parameter is required"); |
} |
// ... |
}) |
``` |
The error appears in observability storage and VoltOps telemetry. |
## Best Practices |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.