AI execution tracing with Langfuse
Send optional AI execution traces to LangFuse to review model calls, tool activity and failures alongside INSIDE usage records.
What it is
Tracing sends optional AI execution records to LangFuse. A trace can link the model's instructions and responses with tool calls, usage and errors. INSIDE also keeps local usage and diagnostic records, which are separate from external tracing.
Who uses it and why
Developers inspect traces to understand unexpected answers, failed tools or prompt changes. Supervisors can use trace findings when reviewing AI behaviour, with access to the external project arranged separately. Powerfront manages the tracing connection and account diagnostics.
Where to find it
Select Tracing in Settings › AI › Features for a supported feature. The Settings role with AI permission, or the Developer role, grants feature configuration access.
Review external records in the configured LangFuse project. Powerfront manages the connection in Settings › AI › Connections and the account's Debug option in Settings › AI › Accounts.
How it works
When enabled, tracing records the start and end of an AI execution. Model calls can include the prompt, model settings, output, status and available token usage. Tool records can include the tool name, arguments and returned content. External prompt identity can also be recorded when available.
INSIDE batches trace events before sending them. Tracing failures are handled so optional telemetry does not interrupt an otherwise running AI flow. The AI can therefore return an answer even when its external trace is incomplete.
Local usage records can identify the AI feature and model associated with a conversation or action. They are not a complete replacement for the external sequence of model and tool activity.
Configure it
Ask Powerfront to create and test a LangFuse connection for the correct project. Edit the AI feature, choose it under Tracing and save. Select None to stop external tracing for that feature.
Ask Powerfront to change Debug when additional account diagnostics are needed. It is a separate choice from Tracing. Arrange access and retention in LangFuse for the people who will review records.
Settings
| Setting | Where | Applies to | Default | Who can change it | What it does |
|---|---|---|---|---|---|
| Tracing | Settings › AI › Features | website | None | Settings role, Developer role | Selects the external tracing connection for this AI feature. |
| Debug | Settings › AI › Accounts | website | Off | Powerfront | Enables additional AI account diagnostics; Powerfront uses this separately from external tracing. |
Fields
| Field | Where | Required | What it is for |
|---|---|---|---|
| Host | Settings › AI › Connections | No | Selects the LangFuse service receiving trace events. |
| Public Key | Settings › AI › Connections | Yes | Supplies the public credential for the LangFuse project. |
| Secret Key | Settings › AI › Connections | Yes | Supplies the matching secret credential for the LangFuse project. |
Permissions
| Permission | Lets you |
|---|---|
| AI | With the Settings role, allows selection of Tracing on supported AI features. |
| Developer | Allows selection of Tracing on supported AI features. |
Limits and good to know
- Traces can contain conversation content, instructions, tool arguments and results. Consider those payloads when arranging external access and retention.
- Normal event flushing starts at one-second intervals; records need not appear immediately.
- The default queue holds up to 10,000 events. New events can be dropped when it is full.
- Events larger than three MiB are dropped. Failed deliveries can also leave gaps.
- Tracing is a diagnostic aid, not a guaranteed complete audit record.
- Privacy processing for AI content does not guarantee that every trace field or local diagnostic message is redacted.
Depends on and works with
View all connections as a list
This feature uses
- AI feature settings and change historyConfigured by
Each supported AI feature selects its own tracing connection.
- AI functions and required inputsEnhances
Traces can include tool calls, arguments and results.
- AI Experiments Lab for response testingEnhances
Experiment runs can link dataset results and scores with external traces.
- Shared AI service connectionsRequires
External tracing requires a configured LangFuse connection.
- External AI prompts and versionsEnhances
Managed prompts can supply effective prompt identity for trace records.
- Personal information masking for AIEnhances
Configured privacy processing can replace known personal values in supported trace inputs without guaranteeing complete redaction.
Used by
- AI Experiments Lab for response testingEnhances
A configured tracing service can receive experiment traces.
- INSIDE AiFlow test dashboard · managed by PowerfrontEnhances
Chat Analysis can use conversation and tracing evidence.
- Shared AI service connectionsEnhances
LangFuse connections receive optional AI traces.
- External AI prompts and versionsEnhances
Tracing can record the effective prompt name and version.
- AI guardrail policy connectionsRelated
Use optional execution traces to investigate recorded guardrail assessments and interventions.
FAQ and troubleshooting
Why is there no trace for an answer?
Check the feature's Tracing selection, the website and the external project. Ask Powerfront to test the connection. Queue pressure, oversized content or failed delivery can also leave a missing record.
Can tracing failures stop the AI answering?
Optional tracing failures are caught so the flow can continue. A missing trace by itself does not show that generation failed.
Why do usage figures and trace counts differ?
They come from separate recording paths. One execution can contain several model and tool events, while external delivery can lose events.
Does Debug enable LangFuse tracing?
No. Debug adds account diagnostics. Select a connection under the feature's Tracing control to enable external tracing.
Can I assume traces contain no personal information?
No. They can include inputs, outputs and tool content. Check your configured privacy processing and the external project's access and retention arrangements.