Reading a trace in the LangSmith UI
The concepts are half the domain. The other half is knowing where things are, because Monitor is the most interface heavy part of LCAE and you sit the exam with smith.langchain.com open.
The three screens that matter
The project list. Traces are organised into tracing projects. One project is usually one app or one environment. Whatever you set as the project name at instrumentation time is the bucket everything lands in, which is why a misconfigured project name looks exactly like missing traces.
The runs table. Open a project and you get a filterable table of runs, one row each, with columns for name, status, latency, token counts and cost. This is your search surface. Filter on status to isolate errors, on time range to isolate an incident window, on metadata to isolate one user or one deployment.
The trace detail view. Click a row and you get the run tree from the last lesson rendered visually, with a per run breakdown beside it. Selecting any node in the tree swaps the detail pane to that run: its inputs, its outputs, its metadata, its feedback, its latency and token usage.
The workflow, in the order you actually do it
Filter the table down to a population. Open one representative case. Walk the tree until the output stops being what you wanted. The run where the output first goes wrong is the run that is broken, and everything downstream of it is just faithfully processing garbage.
That last sentence is the whole skill. Most debugging failures are people fixing the last run in the trace, which is only ever the run that reported the problem.
What good instrumentation buys you here
Every filter you wish existed is a piece of metadata somebody did or did not attach at trace time. That is the next lesson, and it is the one that turns this table from a log into a tool.
Try it yourself
Runs table versus trace detail
Two different screens, two different jobs. If you blur them together you will waste time on the day hunting in the wrong one.
What question is the runs table good at answering, and what question is the trace detail view good at answering? Give one example each.
Finding every failure in the last hour
Production is misbehaving and you want every failed run from the last hour. Where do you go first?
Where evaluator output lands
An online evaluator scored a run. You want to see that score. Where is it?
Navigate before you memorise
You get a provisioned LangSmith org roughly two hours before the exam, and smith.langchain.com stays open to you during it.
You can get from the org home to a specific run's metadata in under thirty seconds without guessing, and you know that the fastest way to answer a "where would you find X" question is to go and find X.