Noveum Trace
Capture production AI behavior with explicit Python instrumentation and framework adapters.
The noveum-trace Python package records complete AI requests and their model, retrieval, tool, agent, and voice operations. It sends connected traces to Noveum for observability, ETL, evaluation, and release comparison.
Choose an integration
| Application | Approach |
|---|---|
| Custom Python or provider client | trace_operation and trace_llm_call context managers |
| LangChain or LangGraph | Runtime callback handler |
| CrewAI | Event listener |
| LiveKit | Session tracing and optional STT/TTS wrappers |
| Pipecat | Pipeline observer and task handlers |
| Another language | REST trace ingestion |
Only instrumented operations are captured. SDK initialization configures delivery but does not discover uninstrumented application code.
Capture for evaluation
Preserve model inputs and outputs, system prompts, usage, offered tool schemas, tool calls and results, retrieval queries and context, terminal response, errors, and stable request identifiers. Use NOVEUM_SERVICE_VERSION to identify the behavior that produced each trace.
Review privacy and redaction before sending production prompts, transcripts, audio, or tool payloads.