Traces & Observations

The Tracing page is the main debugging view. It lists every request your application made, with inputs, outputs, timing, cost and the nested steps that produced the result.

The page has two tabs: Traces (one row per trace) and Observations (one row per span, generation or event across all traces).

The trace table

Traces are listed newest first. Columns shown by default are Timestamp, Name, Input, Output, Observation Levels, Latency, Tokens, Total Cost, Environment, Tags, Metadata and one column per score. Session, User, Version, Release, Trace ID and detailed cost and token columns can be turned on from the column menu.

Search

The search box has two modes:

• IDs / Names (default): searches trace ID, trace name and user ID.
• Full Text: also searches input and output. Narrow the table with filters first for better performance.

Filter

• Date range: from 30 minutes to All time. The default is the last 24 hours.
• Env: one or more environments.
• Source: Gateway or SDK & OTel. Shown only when Budget AI is available in the project.
• Filters: build conditions on bookmarked, ID, name, timestamp, user ID, session ID, metadata, version, release, level, tags, token counts, level counts, scores, latency and cost, plus:
  • AI service: the agent name or the model provider.
  • Data classification: traces where AI Guardrails found PII, PCI or PHI.
  • Policy decision: the guardrail result, Pass, Blocked or Sanitized.

Combine these with the date range, a user ID filter and Full Text search to answer investigation questions in one place, for example "blocked requests from this user to this AI service last week that mentioned a project name". Data classification and policy decision come from AI Guardrails. A trace that guardrails have not checked has neither, so it does not match these filters. How far back you can search follows your data retention setting. For a guided workflow, see Investigations.

Open the full session

When a matching trace belongs to a session, select its session badge to open the whole multi-turn conversation in order. See Sessions.

Save a table view

Open the Table View drawer and choose Save Current Table View. A view stores filters, search, sort order, column order and column visibility. Saved views can be updated and shared as a link.

Batch actions

Select rows (or select every row that matches the filter) to:

• Delete Traces. Deletion runs in the background. Deleted traces can stay visible for up to 15 minutes.
• Add to Annotation Queue for human review or to collect evidence in an investigation queue. See Annotation.

Export

Choose Export above the table, then as CSV, as JSON or as JSONL. The export uses the current filters, search and sort order. It runs in the background. You get an email when it is ready, and the file is listed under Project Settings > Exports. You may be asked to confirm with MFA first. Exports include input and output, so they need the content permission as well as data export access. See Data Exports.

The Observations tab has the same search, filter and export controls. The AI service, data classification and policy decision filters are on the Traces tab.

Peek

Click a row to open the trace in a side panel without leaving the table. Use the arrows to move to the previous or next trace, or expand it to a full page.

The observation tree

Open a trace to see its observations as a tree on the left. Search the tree by type, title or ID.

Each observation has a level (Debug, Default, Warning or Error). Use Min Level in the settings menu to hide low-level noise.

Turn on Timeline to show the same observations as a time chart, so you can see what ran in parallel and where time was spent.

The settings menu also controls Show Graph (see Agent Graphs), Show Comments, Show Scores, Show Metrics and Color Code Metrics.

Use the download icon to save the trace and its observations as a JSON file.

Trace and observation details

Select the trace or any observation to see it on the right.

• Badges: session, user ID, environment, latency, cost (with a breakdown), token usage, release and version. Generations also show the model, time to first token, token counts and the linked prompt version.
• Preview tab: input, output and metadata. Switch between Formatted and JSON. Chat messages are rendered as a conversation, and attached media is shown inline (see Multimodal & Media).
• Scores tab: every score on the trace or observation.
• Gateway panel: for requests that went through Budget AI, the key, model group, deployment, provider model, provider, request ID and cost.

If your organization has turned off trace input and output, a notice replaces the content.

Opening a trace, in the side panel or as a full page, is recorded in the audit log as a content view.

Working with a trace

Next steps

• Sessions: group traces into conversations.
• Agent Graphs: see a LangGraph agent's path.
• Scores: measure trace quality.