Skip to main content
Sometimes you need to find traces based on what they mean, not just what they say. Deep search uses AI to understand the semantic meaning of your query and finds relevant traces even when they don’t contain your exact keywords. Deep search helps you:
  • Find concepts: Search for “frustrated users” even when they never use that word.
  • Discover patterns: Identify traces with similar issues or behaviors.
  • Ask questions: Query like “where did the agent give up?” instead of keyword matching.
  • Surface edge cases: Find unusual interactions you didn’t anticipate.
  • Analyze sentiment: Locate traces with specific emotional tones.
Unlike keyword search or SQL filters, deep search understands meaning and context. To use deep search, you must enable the deep search feature flag and ensure OpenAI is configured. Deep search relies on gpt-4o-mini and gpt-4o for semantic evaluation.
For self-hosted deployments, deep search requires version v1.1.23 or later.
  1. Enable the feature flag.
    1. Go to Settings > Feature flags.
    2. Find Deep search.
    3. Toggle it on.
  2. Check OpenAI configuration.
    1. Go to Settings > AI providers.
    2. Verify OpenAI is configured with an API key.
    3. If not configured, add your OpenAI API key.

​
Run a deep search query

Deep search runs in the Loop panel on the Logs page.
  1. Go to the Logs page in your project.
  2. Click Loop .
  3. In the Loop panel, click Deep search.
  4. Enter a natural language query.
Deep search evaluates a sample and returns up to 20 most relevant traces. For comprehensive analysis, combine with filters. Up to 20 relevant traces stream into the Loop panel in real-time.

​
Understand results

Deep search results stream into Loop. They are ordered by semantic relevance, not chronologically. The most relevant traces appear first. Each result shows:
  • Input: The user request or prompt.
  • Output: The response.
  • Quote: The specific text excerpt that matched your query.
Click any result to view full trace details with metadata. Deep search evaluates a sample and returns up to 20 most relevant traces. For comprehensive analysis, combine with filters.

​
Apply results as filter

Convert deep search results to a table filter:
  1. Review the traces in Loop.
  2. Click Apply as table filter. The logs table filters to just the found traces.
  3. Optionally, save this view for later.

​
Combine with other filters

Deep search works alongside other filtering:
  1. Apply SQL filters or tag filters first to narrow the search space.
  2. Run deep search on the filtered results.
For example, to find production traces with frustrated users:
  • Filter to metadata.environment = "production".
  • Deep search for “frustrated users”.

​
Query examples

Show me examples where users are struggling
This finds traces where:
  • Users asked for clarification multiple times
  • The interaction seemed confused or frustrated
  • The application failed to understand the request
Find conversations with frustrated users
This finds traces with:
  • Negative sentiment in messages
  • Complaints or criticism
  • Repeated failed attempts
Highlight cases where customers are happy
This finds traces where:
  • Users expressed satisfaction or thanks
  • Interactions completed successfully
  • Positive sentiment throughout
Where did the agent refuse to answer?
This finds traces where:
  • The application declined to respond
  • Safety or policy limitations were hit
  • Requests were out of scope
Show me traces where things went wrong
This finds problematic traces even if they don’t contain “error”:
  • Incomplete responses
  • Wrong information provided
  • User confusion or dissatisfaction
Find unusual or unexpected interactions
This surfaces traces that:
  • Don’t fit common patterns
  • Show unexpected user behavior
  • Reveal uncommon use cases

​
Query tips

❌ "Problems"
✓ "Show me traces where users couldn't complete their task"
Specific queries return more relevant results.
❌ "Bad"
✓ "Find traces where the response was factually incorrect or misleading"
Descriptive queries help the AI understand your intent.
❌ "Errors"
✓ "Where did the application fail to understand the user's question?"
Behavior-focused queries find semantic matches beyond keywords.
✓ "Which conversations had back-and-forth clarification?"
✓ "Where did users express confusion?"
✓ "When did the agent repeat itself?"
Natural questions often work better than keywords.

​
Deep search vs SQL

Choose the right tool for your task:
Use caseToolExample
Exact field matchesSQLmetadata.user_id = "user_123"
Numeric thresholdsSQLlatency > 2000
Semantic patternsDeep search”frustrated users”
Concept discoveryDeep search”where things went wrong”
Complex conditionsSQLcost > 0.10 AND error IS NULL
Sentiment analysisDeep search”happy customers”