Introducing Verified Queries for Looker Data Agents

Looker Tips & Resources Technical Tips & Tricks
,
1

Conversational Analytics enables business users to ask questions in plain English and get instant visualizations and data summaries. But as data leaders and developers know, business metrics can be nuanced. When a stakeholder asks, “What was our ARR in EMEA last quarter?”, close enough isn’t good enough. You need 100% precision, consistent filter logic, and trust in every single answer.

Today, we are introducing Verified Queries for Looker Data Agents.

Verified queries allow agent creators to pair specific natural language questions with pre-configured Looker Explore queries inside the agent context. This ensures that high-priority, complex or nuanced business questions return authorized, accurate exact results.

Key Benefits

  • Deterministic Accuracy: When a user’s prompt matches a verified query semantically, the agent executes the exact pre-approved Looker Explore query. This eliminates ambiguity and delivers precision for critical business metrics.

  • Continuous Learning: The agent learns from your verified queries, referencing their structure, dimension choices, and filter patterns to improve its overall query logic across future, related questions.

  • Streamlined Configuration: Native integration within the Agent Builder UI enables you to configure verified queries quickly by entering sample user questions and either pasting an existing Explore URL or constructing the query directly in the interface.

Configuring Verified Queries in the Agent Builder

You can set up verified queries inside the Looker Data Agent configuration page:

  1. Navigate to your Agent: Open your Explore Data Agent in Looker and enter the Edit Agent mode.

  2. Add a Verified Query: Click + Add verified query.

  3. Define Question & Answer:

    • Question: Enter the natural language question your stakeholders are likely to ask.

    • Answer (Option A): Paste an existing Looker Explore URL containing the exact fields, filters, sorts, and visualization you want.

    • Answer (Option B): Click Preview with an empty answer field to launch the native Explore builder, craft your query from scratch, and save it on the spot!

Configure Verified Queries

Best Practices for Agent Creators

To maximize the performance and accuracy of your data agents, keep these field-tested recommendations in mind:

Focus on Complex and High-Stakes Questions

Don’t spend time creating verified queries for simple, single-metric requests like “What were total sales yesterday?” Looker’s semantic model handles those standard aggregations. Instead, reserve verified queries for complex logic, such as:

  • Multi-condition date boundaries (e.g., Trailing Twelve Months vs. YTD).

  • Specialized business filters (e.g., excluding internal test accounts or specific transaction types).

  • Queries requiring specific dimensions to disambiguate similar metrics (e.g., Booked ARR vs. Recognized Revenue).

Keep Queries Under 50

We recommend keeping total verified queries under 50 per agent to optimize agent performance. Focus on high-value, recurring business questions rather than exhaustive coverage.

Close the Loop with the User Feedback System Activity Dashboard

Agent design is an iterative process. You don’t have to guess where your users encounter friction:

  • Looker includes a dedicated Responses Feedback System Activity Dashboard tracking thumbs-up and thumbs-down reactions from conversational sessions.

  • Pro Tip: Periodically filter the dashboard for queries with negative feedback (:-1:). Inspect the natural language prompt that caused the confusion, verify the intended Explore answer, and add it directly as a new verified query in your agent. This creates an automated improvement cycle driven by real-world usage.

Get Started Today

Verified queries bridge the gap between conversational AI and enterprise BI governance, giving your stakeholders the freedom of natural language with the precision of verified LookML queries.

Check out the documentation to start configuring your first verified queries!

If you have questions or feedback, give us your thoughts here in the Cloud Community or share how your team is putting conversational agents to work.

3 Likes
2

Exactly what conversational analytics needed! Pairing Looker Explore URLs directly with natural language prompts solves the trust gap for high-stakes business questions.

Love the proactive tip on using the Responses Feedback dashboard to curate verified queries over time. Can’t wait to roll this out to our stakeholders! :tada:

1 Like
3

@Vaibhavi_Sonavane there any benefits to using verified queries versus defining these in the agent instructions text box?

  1. The documentation refers to entering a ‘natural language question’ in the singular, and the text box for entering the question is small. Do we infer from this that we’re only supposed to enter one natural language question per query? This would seem pretty limited, as different users refer to the same thing in multiple ways (I typically include about 7-8 natural language variations for each Looker query).

  2. Are there any performance benefits to using verified queries versus defining these in the agent instructions (i.e. does it reduce token consumption)? Another limitation is that there is no opportunity to add any context around each verified query. For example, I’ve had success with explicitly telling the agent what it is supposed to infer from a golden query (e.g. “Note: this template applies to ANY entity (e.g. country, genre, supplier). You should dynamically swap the dimension view_name.field_name in the fields list with the entity requested by the user. Keep all other measures and filters identical.”). But additional text consumes tokens / eats into the context window, so I would maybe accept that loss if you said it was a more performant way of feeding the agent additional context.

4

Verified queries could be a useful way to improve consistency when working with Looker Data Agents. It would be interesting to see how queries are validated, maintained, and updated as underlying data models or business definitions change.

5

@Vaibhavi_Sonavane thanks for sharing. In your post you explicitly invited questions. Can you answer mine please (posted previously) :

Are there any benefits to using verified queries versus defining these in the agent instructions text box?

  1. The documentation refers to entering a ‘natural language question’ in the singular, and the text box for entering the question is small. Do we infer from this that we’re only supposed to enter one natural language question per query? This would seem pretty limited, as different users refer to the same thing in multiple ways (I typically include about 7-8 natural language variations for each Looker query).

  2. Are there any performance benefits to using verified queries versus defining these in the agent instructions (i.e. does it reduce token consumption)? Another limitation is that there is no opportunity to add any context around each verified query. For example, I’ve had success with explicitly telling the agent what it is supposed to infer from a golden query (e.g. “Note: this template applies to ANY entity (e.g. country, genre, supplier). You should dynamically swap the dimension view_name.field_name in the fields list with the entity requested by the user. Keep all other measures and filters identical.”). But additional text consumes tokens / eats into the context window, so I would maybe accept that loss if you said it was a more performant way of feeding the agent additional context.