Tracing
See what happened during a call — the transcript, tool calls, and (optionally) full traces in your own Langfuse project.
Call Transcript#
Every call run has a Call Transcript panel showing the conversation as it happened, message by message. When the agent calls a tool during the conversation, it appears inline as an expandable card showing the function name, its arguments, the result, and how long the model took to decide to call it.
Open a call from Agent Runs to see this panel alongside the recording, starting details, and collected details (extracted variables, call disposition, tags).
Full traces via Langfuse#
For a complete, provider-level trace of every LLM call, prompt, and tool invocation, Lemu can send trace data to your own Langfuse project. This is configured under Telemetry in account settings.
On this deployment, the account settings page that hosts the Telemetry/Langfuse form is served by Stack Auth, which is disabled while the app runs in local-auth mode — so this page currently isn't reachable. The setting exists in the product; whether to re-enable that settings page is a call for the Lemu team, not something to route around here.
Setup (once reachable):
- Sign up at Langfuse and create API credentials for a project.
- In Lemu, go to account settings → Telemetry and add your Langfuse host, public key, and secret key.
- Save — new calls will start sending traces to your Langfuse project going forward.
Tips for debugging calls#
- Debug unexpected responses — expand a tool-call card to see exactly what arguments the model sent and what came back.
- Check transcription accuracy — review the transcript for places the agent misheard the caller.
- Confirm tool usage — the transcript shows whether a tool was actually invoked, which helps when a workflow node has multiple tools attached.