Keep Every LangSmith Trace Without the 10 Retention Bill

LangSmith is a valuable tool for debugging live AI systems. However, storing every trace in its extended-retention tier can quickly become expensive. LangSmith offers a new solution that allows you to keep the best of both worlds: live debugging capabilities via LangSmith's 14-day retention, while also archiving verified traces to your organization's private S3 storage.

This way, you can query the retained Parquet data directly with DuckDB, without having to rely on LangSmith's extended-retention tier that costs 10 times more. The new langsmith-cli archive workflow separates live observability from long-term retention. Traces are captured in two passes - D+2 export and D+12 reconciliation - which deduplicate by run ID to ensure safety and convergence of the archive.

You can route projects to separate S3 buckets or prefixes for development and production data, and schedule the sync using cron, Kubernetes CronJob, or your existing scheduler. Queries of archived traces can be made using familiar commands that use DuckDB over Parquet in S3, without requiring a LangSmith API key. This means incident investigations, regression analysis, audits, and offline analytics can continue even after the base-tier trace has disappeared from the live LangSmith API.

By keeping base retention for day-to-day debugging, disabling automatic retention extension where unnecessary, exporting every trace to private S3 before the 14-day window closes, and using the new archive workflow, you can maintain LangSmith's live debugging experience while retaining your full historical dataset under your own control. This approach also makes long-term storage costs predictable.