Sessions, users, scores, and evals

A trace answers "what happened in this request". Two more questions need more than a trace. "What is this customer costing us, and what did this conversation look like end to end" needs traces grouped by user and by session. "Which answers were bad" needs a number attached to each trace after the fact, from a person, a rule, or another model. Langfuse calls the first two attributes and the third a score, and both are a few lines of code on Relay.

Wrap the work of a request in propagate_attributes and every observation created inside the block, including nested spans and drop-in generations, carries the user, the session, and any metadata you add. For Relay the user is the customer's id and the session is the conversation id, so a multi-turn chat shows up as one session with one trace per turn.