Users can search for patterns such as a PII leak, user frustration, or an agent claiming to have completed work it did not complete. Queries can be combined with existing filters, including error status and tags.
The company says AI Search uses a decision model to assess filtered candidate trajectories rather than relying on vector similarity alone. SmithDB normalizes supported trace formats into session trajectories and removes repeated messages. LangSmith assigns each trajectory an 8K-token budget, prioritizing user messages and agent replies before other content; material beyond the budget can be trimmed or dropped.