Skip to main content
Deep search finds traces by meaning, not only exact text. Use it when you remember what happened in a conversation but do not know the exact words, IDs, tags, or attributes. Deep search page with semantic query, source filters, and related trace results

​
Search by meaning

Enter a natural-language query such as:
  • users asking for direct financial advice
  • assistant forgot to cite sources
  • tool call failed after retrieval
  • refund policy answer with low confidence
Adaline searches the embedded trace content and returns semantically related messages. Deep Search is best for exploratory investigation. Use table filters when you know exact metadata, and use Deep Search when the evidence is conceptual: a user intent, a bad answer shape, a tool failure pattern, or a policy concern.

​
Filter by source

Use source filters to focus the search:
SourceUse it for
User messageWhat users asked for.
Model responseWhat the assistant said.
ReasoningIntermediate model reasoning when your stack logs it.
Tool callTool invocation names and arguments.
Tool responseData returned by tools.
RetrievalSearch or RAG context.
OrchestrationRouting and workflow spans.
GuardrailSafety, policy, or moderation spans.

​
Write better queries

Good Deep Search queries describe the situation, not only keywords:
Less usefulMore useful
refunduser asks for a refund after a subscription renewal and assistant gives policy answer
bad answerassistant gives confident answer without citing retrieved context
tool failedtool call times out after retrieval and assistant apologizes without retrying
Use one query to find the pattern, then open several results before deciding whether it is a Behavior, evaluator, dataset, or prompt issue. When you open a result, Adaline can land you on the relevant span inside the trace viewer so you can inspect the surrounding request, not only the matched message.

​
Turn results into evidence

Open promising results in Traces, then:
  • Add useful examples to datasets.
  • Check whether a Behavior already captures the pattern.
  • Create or refine evaluators for the issue.
  • Start an Improve cycle when the evidence points to a prompt-level fix.
For trace-derived dataset rows, see Add spans to datasets.

​
Privacy and scope

Deep Search searches trace content that your application sent to Adaline. Treat results as production data. Avoid broad queries when a narrower time range, project, prompt, or source filter is enough.