Events

Events are structured payloads you attach to a span at a point in time. They are ideal for storing inputs, outputs, and intermediate artifacts without turning your span list into noise.

Common event types

• input: prompts, tool arguments, retrieval queries (sanitized)
• output: model text, tool results (truncated if large)
• metadata: tags that help filtering (release, environment, model)
• warning: partial failures, fallbacks, validation issues

Redaction and truncation

Treat events like logs: they can contain sensitive data if you’re not careful.

Recommended approach:

• redact secrets and PII before emitting events
• truncate large blobs (documents, long tool results)
• store references (ids/URLs) instead of full content when policy requires

Filtering-friendly structure

Prefer JSON objects with stable keys over unstructured strings. That lets you:

• filter by model, tool, env, release
• compare outputs across versions
• build eval datasets from consistent artifacts

Next steps

• Read Spans for structuring workflows.
• Read Security & privacy for data policies.