Customization Options

Tracing Enabled#

Controls whether the SDK should send traces.

When disabled:

• new sessions can become no-op sessions
• spans/events/errors are ignored by the SDK or backend
• application code can keep calling the SDK safely

Use this when:

• debugging noisy local environments
• disabling tracing for a sensitive customer
• reducing observability overhead

Capture Prompt/Output Content#

Controls whether TraceLLM can store prompt and output text.

Default: off.

When off:

• input and output content is stripped or stored as null
• metadata, timing, token counts, and status can still be captured

Turn it on only when:

• you need prompt/output debugging
• your data handling policy allows it
• redaction settings are enabled as needed

Capture Inputs#

Controls whether input fields may be stored.

This is narrower than full content capture. It is useful when teams want prompt/request visibility but not generated output storage.

Capture Outputs#

Controls whether output fields may be stored.

This is useful when debugging answer quality or response formatting.

Errors#

Controls structured error capture.

When enabled, TraceLLM records:

• error name
• error message
• error type
• optional stack, depending on content capture

Keep this enabled for most environments. It is high signal and usually low data volume.

Token Usage#

Controls token usage capture.

Token usage fields:

• inputTokens
• outputTokens
• totalTokens
• optional estimatedCostUsd

Use this for:

• cost visibility
• model comparison
• quota management
• usage-based billing later

Metadata#

Controls arbitrary attributes on sessions, spans, events, and errors.

Good metadata examples:

{
  "provider": "openai",
  "model": "gpt-4.1-mini",
  "customerTier": "pro",
  "workflow": "support-answer"
}

Avoid putting secrets or raw user content into metadata.

Sampling Rate#

Controls how often the SDK traces workflows.

Values:

• 1: trace every sampled operation
• 0.5: trace about half
• 0: trace none

Use sampling when:

• traffic is high
• you want production visibility without storing everything
• you want to reduce cost or storage

Redact Emails#

Replaces email-like strings in captured content.

Example:

alice@example.com -> [redacted-email]

Redact API Keys#

Replaces API-key-like strings in captured content.

Example:

sk_abc123... -> [redacted-key]

Ignored Span Kinds#

Lets the project ignore whole categories of spans.

Examples:

• keep only llm spans for a minimal view
• ignore future retrieval spans in a high-volume RAG system
• ignore future tool spans for noisy internal operations

Tool and retrieval span kinds are supported by the data model and SDK API, but their dedicated UI controls are intentionally hidden until the demo app emits real tool/retrieval spans.

Ignored Tools#

Reserved for filtering specific tool names.

Example future usage:

{
  "ignoredTools": ["healthcheck", "cache.get"]
}

This is useful when some tools are noisy and not useful for LLM debugging.