Skip to main content
Deep search helps you find relevant production evidence when you do not know the exact trace name, tag, attribute, or keyword. Ask for the situation in plain language, and Adaline searches across logged trace and span content to find semantically related results. Use it when conventional filters are too literal: a user intent, answer shape, tool behavior, policy concern, or recurring issue may appear in many different words. Traces table with the Deep search entry button visible above filtered production logs

​
Search by meaning

Start from a phrase that describes what you want to find:
  • users asking about cancellation but keeping old data
  • cooking questions about substitutions or technique
  • assistant gives an answer without using the right tool
  • model response sounds confident but misses the user constraint
Deep search returns matched log content with the source type, related prompt or span context, and the matching text. Open a result to move back into the normal trace or span view and inspect the surrounding request. Deep search results for cooking questions with user and model matches across logged traces

​
Use source filters

Source filters help you search the right part of the log. For example, search User message when you care about what users asked, Model response when you care about what the assistant said, and Tool call or Tool response when the issue might be in tool usage. Deep search source menu with user message, model response, reasoning, tool call, tool response, retrieval, orchestration, and guardrail options Use Deep search when the evidence is conceptual. Use filters when the evidence is exact. Deep search is discovery, not the final decision. Treat results as leads, then open the source trace or span before adding a dataset row, changing an evaluator, starting Improve, or making a release decision.

​
Turn matches into action

After you find useful matches:
  1. Open several results and inspect the surrounding trace or span.
  2. Check whether the same pattern appears in Behaviors.
  3. Add representative spans to Datasets when the case should become regression coverage.
  4. Create or refine Evaluators when the pattern should be measured continuously.
  5. Start Improve when the evidence points to a prompt-level fix.

Filter, search, export logs

Use exact filters and exports after finding the right evidence.

Inspect a trace

Open a Deep search result and review the full trace context.

Understanding Behaviors

See whether the matched examples are part of a repeated pattern.

Use logs to improve prompts

Turn production evidence into datasets, evaluators, and Improve cycles.