Call data export
Zabio can send a record of each call your organization makes to your own observability tool, so your LLM calls sit beside the rest of your system. The records travel over OpenTelemetry: one OTLP span per call, sent over HTTPS as protobuf, with the attributes of the OpenTelemetry GenAI semantic conventions. Any tool that accepts OTLP traces over HTTP can receive them.
The export is off until you set it up. It never slows a call: records are sent in the background, a few seconds after the call ends.
Set it up
Section titled “Set it up”In the web app, open Settings, then Call data export. Owners and admins can set it up; members do not see the section.
- OTLP traces endpoint. The full address of your tool’s traces endpoint, used exactly as you write it. It usually ends in
/v1/traces. It must be publichttps, without a user name or password in it. - Headers. Up to 5 headers sent with every request, for example your tool’s
Authorizationheader. Zabio stores the values encrypted with your organization’s key and never shows them again. To change one, choose Replace headers and enter the full set again. - Send call records. Switch it on and save.
The section then shows when the last records were delivered and, if a delivery failed, why.
Langfuse
Section titled “Langfuse”| Field | Value |
|---|---|
| Endpoint | https://cloud.langfuse.com/api/public/otel/v1/traces (EU) or https://us.cloud.langfuse.com/api/public/otel/v1/traces (US) |
| Header | Authorization: Basic followed by the base64 of public key:secret key |
| Header | x-langfuse-ingestion-version: 4 |
Grafana Cloud
Section titled “Grafana Cloud”| Field | Value |
|---|---|
| Endpoint | the OTLP endpoint on your stack’s OpenTelemetry tile, followed by /v1/traces, for example https://otlp-gateway-prod-us-east-0.grafana.net/otlp/v1/traces |
| Header | Authorization: the value the tile shows for OTEL_EXPORTER_OTLP_HEADERS, starting with Basic |
Datadog
Section titled “Datadog”| Field | Value |
|---|---|
| Endpoint | the OTLP traces intake endpoint of your Datadog site, as Datadog’s OTLP intake page shows it for your site |
| Header | dd-api-key: your Datadog API key |
| Header | dd-otlp-source: llmobs and DD-LLMObs-Enabled: true, to see the calls in LLM Observability |
What is sent
Section titled “What is sent”One span per call, named after the operation and the model you asked for (for example chat auto/balanced). A request that needed more than one attempt sends one span per attempt, all in the same trace.
| Attribute | Meaning |
|---|---|
gen_ai.operation.name |
chat, embeddings, speech, transcription, image_generation, image_edit or video_generation |
gen_ai.provider.name |
the vendor that answered |
gen_ai.request.model, gen_ai.response.model |
the model you asked for and the model that answered |
gen_ai.usage.input_tokens, gen_ai.usage.output_tokens |
tokens in (cached tokens included) and out |
gen_ai.usage.cache_read.input_tokens, gen_ai.usage.cache_creation.input_tokens |
tokens read from and written to the prompt cache |
gen_ai.usage.reasoning.output_tokens |
reasoning tokens, when the vendor reports them |
gen_ai.response.finish_reasons |
stop, length, tool_call or content_filter |
gen_ai.request.max_tokens |
the output limit you asked for, when you set one |
gen_ai.request.stream |
true on a streamed call |
gen_ai.response.time_to_first_chunk |
seconds to the first token of a streamed call |
gen_ai.conversation.id, gen_ai.response.id |
the conversation and the response, on the Responses API |
error.type |
the error code, on a failed call |
zabio.call.id, zabio.key.id, zabio.project.id, zabio.session.id |
the ids of the call, its API key, its project and its session |
zabio.endpoint, zabio.attempt, zabio.role, zabio.outcome |
the endpoint, the attempt number, what the attempt was in its request, and how it ended |
zabio.routing.line |
the line Zabio applied on an auto call: cheap, balanced or best |
zabio.cost.billed_micros |
what the call cost you, in millionths of a US dollar |
zabio.cost.provider_micros, zabio.cost.service_fee_micros, zabio.cost.call_fee_micros, zabio.cost.share_discount_micros |
the lines of that cost, on calls you pay for |
zabio.cost.not_billed |
true on a call or an attempt you were not charged for |
zabio.cost.currency |
USD |
zabio.units.kind, zabio.units.count, zabio.units.media_size |
on audio, image and video calls: what the call was billed in (seconds of audio, characters, images, seconds of video), how many, and the size of an image or video |
zabio.byok, zabio.service_tier, zabio.reasoning.effort |
your own vendor key, the service tier and the reasoning setting |
zabio.firewall.action, zabio.firewall.detections |
what the AI firewall did on the request and how many detections it reported |
zabio.firewall.tokens_checked, zabio.firewall.billed_micros |
the AI firewall’s charge for the request |
zabio.search.count, zabio.search.engine, zabio.search.billed_micros |
the web searches the call ran: how many the vendor counted, who ran them and what they were billed. Never a query or a result |
The span starts when the call started and ends when it ended, so its duration is the call’s latency. Each request also names the sender (service.name is zabio), your organization (zabio.organization.id) and the region (zabio.region).
Prompts and answers
Section titled “Prompts and answers”By default no prompt and no answer is sent. To send them, switch on Include bodies. It applies only to calls made with an API key that has body logging on (see Teams): what is sent is the body your organization already stores, and nothing else. A call without a stored body sends its record alone.
Chat calls send gen_ai.system_instructions, gen_ai.input.messages and gen_ai.output.messages in the format of the GenAI conventions. Embeddings, speech, transcription, image and video calls send the stored body as it is, in zabio.content.request and zabio.content.response: the text and a description of the media (type, count, size), never the media. Images, audio and video are never sent, only a note that the message had one. A body larger than 256 KB is left out and the span says why in zabio.content.omitted.
What is never sent
Section titled “What is never sent”- Prompts and answers, unless you switched them on as above.
- Images, audio, video or their addresses.
- The AI firewall’s internal scores.
- Web search queries and results.
- Names of API keys or projects, emails of members, IP addresses.
- Error messages from vendors.
Privacy
Section titled “Privacy”If your organization sets up a call data export, Zabio sends a record of each call (model, vendor, tokens, cost, timing, status, the routing line, the firewall’s tag and charge, and the ids of the key, project, session, conversation and response; every field is listed on the Call data export page) to the destination your organization configures. Prompts and answers are sent only when you switch on Include bodies and the call’s key has body logging on. Once sent, that data is held by the destination under its own terms. See Privacy.
Delivery
Section titled “Delivery”Records are sent in batches, usually within 20 seconds of the call. If your endpoint is unreachable or answers with an error, Zabio keeps the records and retries: after 5 seconds, 30 seconds, 2 minutes, 5 minutes, then every 15 minutes. A wrong key or address can be corrected in Settings and the waiting records are sent on the next try. A record that could not be delivered within 6 hours is dropped, and so are the oldest records beyond 100,000 waiting. If your tool rejects a batch as invalid, that batch is dropped.
A record can arrive twice after a retry; its trace and span ids are the same both times.
Switching the export off, or removing the destination, drops the records still waiting and deletes the stored headers.