Exporters
An exporter is a plugin package that sends what Prism observes about model attempts to a service you choose. Prism ships two official exporters in the plugin catalog: OpenTelemetry export and Spend alerts. Exporters see the same metadata as any observer package: model, provider, token counts, cost, latency and outcome. They never see prompt or response text.
Set up an exporter
Section titled “Set up an exporter”- Open Plugins, find the package in the catalog and install it.
- Select the package. Fill in its settings and its keys. Keys are stored in Prism’s sealed vault: after you save one, the console shows only that it is saved. Use Replace to enter a new value or Clear to remove it.
- Choose Send test message. Prism starts the package for this one call, with the settings in the form, and shows the answer of the service, for example “Telegram accepted the test message.”
- Choose Enable. From now on every observed attempt is delivered to the package.
The package card shows the result of the last send. When the service refuses a message or span, Prism still marks the delivery as handled and shows the service’s answer with your keys removed from it. Spans and messages that were refused are not sent again, except spend alerts, which are tried again with the next attempt.
Removing a package deletes its stored keys.
What an exporter may do
Section titled “What an exporter may do”- It sends only to the hosts its package declares (for example
api.telegram.organdhooks.slack.com) or to the one address you type into its settings. The address must behttps://and must not point at the Prism computer itself; addresses on your LAN or tailnet are allowed. - Prism resolves the address before each send and refuses it if the name now points at the Prism computer or at a cloud metadata address.
- It sends at most eight requests per delivery, GET or POST, each at most 64 KiB, and Prism does not follow redirects.
- It builds its own requests. Prism adds no keys or headers of its own.
OpenTelemetry export
Section titled “OpenTelemetry export”OpenTelemetry export sends one span per model attempt over OTLP/HTTP with JSON bodies to the collector address you enter.
| Setting | Value |
|---|---|
| Collector address | The OTLP base address, without /v1/traces. Langfuse Cloud: https://cloud.langfuse.com/api/public/otel (EU) or https://us.cloud.langfuse.com/api/public/otel (US). |
| Service name | Optional; the service.name of the spans, default prism. |
| Authorization header | Optional key. For Langfuse: Basic followed by base64 of public_key:secret_key. |
| Extra headers | Optional key: one Name: value per line, for collectors that read a key from another header. |
Spans follow the OpenTelemetry GenAI conventions (gen_ai.system,
gen_ai.request.model, gen_ai.usage.input_tokens,
gen_ai.usage.output_tokens) and carry Prism’s cost as prism.cost.*. The
charged cost and the catalog-price estimate are separate attributes. When the
charged cost is complete and in US dollars it is also sent as
gen_ai.usage.cost, which Langfuse shows as the generation’s cost.
All attempts of one request are in one trace. Each attempt is sent once, as one span, when it has finished (success, error or cancelled); a streaming attempt that has only started is not sent. Repeated sends of the same attempt keep the same span ID.
Langfuse takes OTLP over HTTP with JSON or protobuf bodies (not gRPC) and
authenticates with Authorization: Basic followed by base64 of
public_key:secret_key.
Spend alerts
Section titled “Spend alerts”Spend alerts adds up the day’s model spend and sends a message to a Telegram chat or a Slack channel:
- when spend crosses a share of your daily budget, 50, 80 and 100 % unless you choose other thresholds, once per threshold and day;
- with the first attempt of a new day, a summary of the previous day.
| Setting | Value |
|---|---|
| Channel | Telegram or Slack. |
| Telegram chat ID | The chat the bot writes to; add the bot to the chat first. |
| Daily budget, Budget currency | Without a budget only the daily summary is sent. |
| Alert at | Percent thresholds, default [50, 80, 100]. |
| Count against the budget | The charged cost (default) or the catalog-price estimate. |
| Time zone | When a day starts, for example Europe/Berlin; default UTC. |
| Telegram bot token | Key, from @BotFather. |
| Slack webhook address | Key, an incoming webhook address starting with https://hooks.slack.com/. |
Charged cost and catalog-price estimates are never added together, and currencies are never converted. The budget is compared with the chosen amount in the budget currency only; spend in other currencies is listed in the message but not counted against the budget.