Converse with Looker data

Conversational Analytics, powered by Gemini for Google Cloud, lets you investigate your data by asking questions in regular, natural (conversational) language through an intuitive chat interface.

This page explains how to use the Conversational Analytics interface in Looker (Google Cloud core) and Looker (original) instances to perform the following tasks:

Learn how and when Gemini for Google Cloud uses your data.

You can access Conversational Analytics in Looker in the following ways:

Start a conversation with an Explore or a data agent

Sets of questions that you ask about a dataset are organized by conversation. There are multiple ways to "have a conversation" using Conversational Analytics. You can ask questions about data in a single Explore or ask a Conversational Analytics data agent questions about as many as five Explores at once. Splitting work into multiple conversations can be useful for organizing lines of inquiry. To create a new conversation, follow these steps:

  1. Navigate to the Conversations page.

  2. Choose one of the following options to start your conversation:

    • Explores: To start a conversation based on up to five Looker Explore, select the Explore panel. The project name is listed beside the Explore name.

    • Agents: Data agents are customized with context and instructions that are specific to your data. To start a conversation with an existing data agent, select the Agents tab, and then select a data agent. You can start a conversation with a data agent that you already created or that another user has shared with you. To create a new data agent, select New agent.

  3. By default the conversation is called "Untitled." After you ask your first question in the conversation, Conversational Analytics automatically generates a conversation title that is based on your question and response. To change the generated name, click the title at the top of the conversation page and enter a new conversation name. To save your changes, click elsewhere on the page, or press return (Mac) or Enter (PC).

Once you have created a conversation, you can ask questions about the data in the Ask a question field within the conversation. You can return to the conversation from the Recent Conversations section.

Start a conversation from within a Looker Explore

You can also start a conversation directly with a Looker Explore. To start a conversation, navigate to the Explore and select Start a conversation.

Ask questions about Looker data

When you're beginning a new conversation, Conversational Analytics suggests some starting questions to ask. The questions don't need to be in a specific format or use a specific syntax. However, they do need to relate to the Explore that you've selected.

Type your question in natural language into the Ask a question field. Select a question mode and click send. After you submit your query, you can cancel Conversational Analytics' response by clicking Stop response. Conversational Analytics stops running the query and displays the following message: The query was cancelled.

Type your question in natural language into the Ask a question field. Select a question mode and click send. After you submit your query, you can cancel Conversational Analytic's response by clicking Stop response. Conversational Analytics stops running the query and displays the following message: The query was cancelled.

For more guidance on the types of questions that you can ask, see Limitations on questions.

Select a question mode

When you ask a question, you can select the question mode from a drop-down menu that contains the options Fast and Thinking. The UI describes Fast as being intended for answering quickly and Thinking as being intended for solving complex problems. Conversational Analytics defaults to Thinking question mode. Conversational Analytics will maintain the same question mode throughout any multi-turn conversations, unless you manually change it.

Fast mode

When you ask a question in Fast mode, Conversational Analytics attempts to directly map your natural language query to the LookML parameters that are defined in the LookML models that underlie your conversation's Explores. Conversational Analytics is able to respond quickly because it relies on the governed definitions of your LookML, and it doesn't use or display any kind of reasoning.

For example, a query like "What was our total revenue last month?" can be quickly translated into a query that selects the total_revenue measure and filters for the previous month.

Select Fast mode if your query is asking for specific facts or predefined metrics from your data.

Thinking mode

Thinking mode is intended for more complex analytical requests that require analysis beyond direct LookML lookups. In this mode, the agent "plans" its approach, deciding which tools to use and how to combine results. This mode allows for solving multi-step problems and performing advanced data science tasks that may not be possible with a single SQL-based query.

Select Thinking mode when asking why about your data, when you're comparing trends, or when you're making more complex analytical requests that may require multiple steps. This mode is also especially helpful when testing an agent to understand how it's using the underlying LookML of its data sources.

How Conversational Analytics processes your questions

Conversational Analytics may rephrase your question after you've submitted a query, and the rephrased question will be displayed in the conversation window following your original question. For example, Conversational Analytics might rephrase the question "What is the mean of user ages?" to "What is the average user age?"

While Conversational Analytics runs your query, you can observe its reasoning and thought process.

Multi-turn conversations

Conversational Analytics will take previous questions and answers into account as you continue the conversation. You can take previous answers and build on them by further refining results or changing the visualization type.

Conversation metadata

When you converse with an Explore or a data agent, the collapsible Data panel shows the name of the Looker Explore that is being used by the conversation. The Data panel also provides the following options:

  • View fields: When chatting with an Explore, you can view the Explore in a new browser window by clicking View fields.
  • Edit agent: When chatting with a data agent, you can edit the details about the data agent by clicking Edit agent.
  • New conversation: Start a new conversation with the Looker Explore that the current conversation is using.

Manage queries within a conversation

When you converse with data, you can manage the conversation by stopping an active query response while it is running or by deleting the most recent question and its response.

Delete the most recent question

To delete the most recent question and its response, follow these steps:

  1. Hold your cursor over the most recent question, and then click Delete message.
  2. In the Permanently delete message? dialog, click Delete to permanently delete the question and its response.

Understand query results and calculations

When you ask questions about your data in Conversational Analytics, the response might include a visualization, a data table, or other details, depending on your specific query and the connected data. To open the query results as an Explore, click Open in Explore within the query results.

In addition to this query response, Conversational Analytics provides the following options for understanding query results and calculations:

Determine how your query was interpreted

If you use Thinking mode to ask your question, you can see how Conversational Analytics reasoned through your query. To see its reasoning, expand the Show reasoning option. To hide its reasoning, click Hide reasoning.

Conversational Analytics analyzes each query and thinks about how to respond, using the keywords from your query to infer the relevant dimensions, measures, and other parameters from the semantic layer of the conversation's associated datasets and interpreting from your query what aggregations may need to be performed. When you expand Show reasoning, Conversational Analytics displays a plain text explanation of the steps that it took to interpret your query. The explanation also includes the duration that Conversational Analytics thought about the query.

Following its reasoning, Conversational Analytics generates a response, which may include a request for clarification about your query.

Determine how an answer was calculated

To see how Conversational Analytics arrived at an answer or created a visualization, click How was this calculated? within the query results.

When you click How was this calculated?, Conversational Analytics displays a Text section. The Text section provides a plain text explanation of the steps that were taken by Conversational Analytics to arrive at the given answer. This explanation includes the raw field names that were used, the calculations that were done, the filters that were applied, the sort order, and other details.

If your Looker admin has enabled the Code Interpreter by turning on the Advanced analytics option for Conversational Analytics data agents, the Code tab displays the additional generated Python code for any advanced queries.

Manage conversations

Conversations are listed by title in the Recent section. You can change the names of conversations, delete conversations, or restore them from the trash folder.

Delete a conversation

To move a conversation to the trash, open the conversation and click Move to trash.

Restore or permanently delete a conversation

To restore or permanently delete a conversation from the trash, follow these steps:

  1. Within Conversational Analytics, select Trash in the left navigation panel to view the list of conversations that have been moved to the trash.
  2. In the Trash section, click the name of the conversation that you want to restore or permanently delete.

  3. In the Are you sure? dialog, select one of the following options:

    • Cancel: Cancels the action.
    • Restore: Restores the conversation. The conversation can be accessed from the Recent section of the left navigation menu within Conversational Analytics.
    • Delete forever: Permanently deletes the conversation.

Known limitations

Conversational Analytics has the following known limitations.

Limitations on visualizations

Conversational Analytics leverages Vega-lite for conversation chart generation. The following Vega chart types are fully supported:

  • Line chart (one or more series)
  • Area chart
  • Bar chart (horizontal, vertical, stacked)
  • Scatter plot (one or more groups)
  • Pie chart

The following Vega chart types are supported, but you may encounter unexpected behavior when rendering them:

  • Maps
  • Heatmaps
  • Charts with tooltips

Chart types that exist outside the Vega catalog are not supported. Any charts that are not specified in this section are considered unsupported.

Limitations on data sources

Conversational Analytics has the following data source limitations:

  • For Looker data, Conversational Analytics can return a maximum of 5,000 rows per query.
  • Conversational Analytics cannot set the value of a filter-only field that is defined using the LookML parameter or filter parameters.

Limitations on questions

Conversational Analytics supports questions that can be answered by a single visualization, for example:

  • Metric trends over time
  • Breakdown or distribution of a metric by dimension
  • Unique values for one or more dimensions
  • Single metric values
  • The top dimension values by metric

Conversational Analytics doesn't yet support questions that can only be answered with the following types of complicated visualizations:

  • Prediction and forecasting
  • Advanced statistical analysis, including correlation and anomaly detection

More advanced questions, such as forecasting, can be answered when the Code Interpreter is enabled.

Sample conversation

The following sample conversation shows how a user can interact with Conversational Analytics in a natural, back-and-forth way. In this example, the user asks the following question: "Can you plot monthly sales of hot drinks versus smoothies for 2023, and highlight the top selling month for each type of drink?" Conversational Analytics responds by generating a line graph that displays the monthly sales of hot drinks and smoothies for 2023, highlighting July as the month with the highest sales for both categories.

Conversational Analytics chat that includes a line graph of monthly sales of hot drinks and smoothies in 2023, with July highlighted. Conversational Analytics chat that includes a line graph of monthly sales of hot drinks and smoothies in 2023, with July highlighted.s

As this sample conversation illustrates, Conversational Analytics interprets natural language requests, including multi-part questions that use common terms like "sales" and "hot drinks," without requiring users to specify exact database field names (like Total monthly drink sales) or define filter conditions (like type of beverage = hot). Conversational Analytics describes its key findings, explains its reasoning, and provides an answer that includes text and, where appropriate, a chart. To encourage deeper analysis, Conversational Analytics may also suggest follow-up questions.

  • Conversational Analytics in Looker overview: The landing page for Conversational Analytics with a list of key features links to all Conversational Analytics documentation.

  • Create and manage data agents: With data agents, you can customize the AI-powered data querying agent by providing context and instructions that are specific to your data, which helps Conversational Analytics generate more accurate and contextually relevant responses.

  • Best practices for configuring Conversational Analytics in Looker: Strategies and best practices to help Looker administrators and LookML developers successfully configure and optimize Conversational Analytics.

  • Enable advanced analytics with the Code Interpreter: The Code Interpreter within Conversational Analytics translates your natural language questions into Python code and executes that code. Compared to standard SQL-based queries, the Code Interpreter's use of Python enables more complex analysis and visualizations.