noveum-trace

Capture AI application executions, find the requests that matter, and inspect the model, tool, retrieval, voice, and workflow spans your integration records.

noveum-trace is Noveum's observability product for AI applications. It records one end-to-end execution as a trace and the instrumented operations inside it as spans. A span can represent a model call, tool invocation, retrieval, agent step, voice operation, or another unit of work that your integration emits.

What noveum-trace helps you answer

Which request failed, timed out, or became slow?
Which model, tool, retrieval, or agent step caused the result?
Did behavior change after a deployment or model update?
Which production examples should become evaluation cases?

noveum-trace can only display information that your application records. A trace does not automatically contain prompts, responses, tool payloads, audio, token usage, or user and session identifiers unless the integration captures those fields.

Start with one complete trace

Install noveum-trace and configure the endpoint, API key, project, environment, and immutable release with service_version.

Wrap one complete request with a context manager or create explicit spans. Do not use a separate trace for every internal step.

Record model, retrieval, tool, agent, or voice spans below the root trace when those operations are useful for diagnosis.

Attach stable user_id, session_id, service_version, and request identifiers when they are available and safe to record.

Open Noveum, find the trace, and confirm that its parent-child order, status, duration, and important attributes match the application execution.

Use Quick SDK setup for the smallest working integration or Production configuration for batching, sampling, privacy, and deployment settings.

Choose the right investigation view

View	Scope	Use it when
All Traces	Organization	You know a name, identifier, project, service, or time range and want a flat table across the organization.
Explorer	Organization	You want to browse the Traces and Datasets directory, narrow by environment and project, and inspect a result in a three-pane workspace.
Project Traces	One project	You already know the project and want its trace table without a redundant Project filter.
Runs	One project dataset	You want to review mapped calls or workflows with their latest available evaluation results and source trace.

All Traces and Project Traces use the same flat table. Selecting a row opens a resizable, non-modal detail sheet. Explorer is the separate three-pane workspace with a directory tree, result list, and detail pane.

Read a trace from outside in

Confirm the trace name, status, start time, duration, project, and environment.
Open Spans and follow the execution flow from the root.
Select the first operation whose status, duration, input, or output differs from the expected path.
Use Overview for trace-level counts and timing, then Metadata for service, SDK, tags, and custom attributes.
Compare the trace with another release or convert representative traffic into a dataset when the behavior needs a repeatable check.

See Trace concepts, Span concepts, and The closed-loop workflow for the next step.

Keep the evidence useful and safe

Use stable project and environment values so filters remain meaningful across deployments.
Emit service_version with each release so before-and-after comparisons are possible.
Avoid secrets and unnecessary personal data in attributes, prompts, tool payloads, and recordings.
Keep trace boundaries aligned with the end-to-end operation you need to debug.
Validate instrumentation in a non-production environment before increasing capture volume.

noveum-trace preserves recorded evidence. It does not evaluate quality by itself. Use NovaEval when you need repeatable scorers and NovaPilot when you want evidence-backed recommendations for your team to investigate.