Lovelaice framing
Trace
Definition
The full record of one AI interaction: input, retrieved context, model calls, tool calls, output. Without traces you can only see that something failed, never why.
A trace is the complete record of one AI interaction, end to end: the user input, the retrieved context, every model and tool call in the chain, and the final output. Traces are the artifact you use to debug production failures, and they're also the raw material for building a golden dataset out of real usage — real failures become real test cases.
Why it matters
Without traces, debugging AI is guesswork. With them, you can reconstruct exactly what the model saw and decided — and mine production for the failures worth adding to your eval set.
Related terms