@agentproto/telemetry-langfuse
v0.2.1
Published
Langfuse ingestion REST API sink for @agentproto/telemetry. Dependency-free: maps EvalEvent batches to Langfuse trace-create and score-create objects over fetch.
Maintainers
agentiknetReadme
@agentproto/telemetry-langfuse
A dependency-free sink that forwards agentproto eval events to the Langfuse public ingestion REST API. No Langfuse SDK is required — the implementation uses fetch directly.
Usage
import { langfuseTelemetry } from "@agentproto/telemetry-langfuse"
const sink = langfuseTelemetry({
publicKey: process.env.LANGFUSE_PUBLIC_KEY,
secretKey: process.env.LANGFUSE_SECRET_KEY,
baseUrl: "https://cloud.langfuse.com",
environment: "production",
})
runEval({ suite, sink })
await sink.flush()Event mapping
| EvalEvent kind | Langfuse action | Notes |
|---|---|---|
| eval.started | trace-create | One trace per runId; name is eval:${suiteId}. |
| eval.case.scored | score-create | One score per case/scorer tuple. |
| eval.finished | score-create | Two scores: eval.meanValue and eval.passRate. |
| unknown | ignored | Forward compatibility. |
Events are buffered in memory until flush() POSTs them as a single batch to POST ${baseUrl}/api/public/ingestion with Basic auth.
Security
Credentials are used only for Basic auth headers. They are never logged or returned from flush().