Conversational analytics in Looker overview

Conversational analytics is a chat-with-your-data feature that is powered by Gemini for Google Cloud. Grounded in the Looker semantic modeling layer, conversational analytics empowers users in your organization to ask data-related questions in regular, natural (conversational) language for governed, trusted self-service BI. This approach accelerates adoption of analytics across your organization by providing enterprise-grade governance and security.

Conversational analytics is available in both Looker (Google Cloud core) instances and in Looker (original) instances.

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

Key features

Conversational analytics includes the following key features:

  • Query Explores in natural language: Converse in natural language with an individual Explore or with a data agent that is querying as many as five Explores.
  • Create a data agent to query Explores: You can customize an AI-powered Explore-querying agent by providing context and instructions that are specific to your data, which helps the data agent generate more accurate and contextually relevant responses. You can also share your data agents with other users or publish them to other applications, like Gemini Enterprise, so that they can ask questions with the same context. You can connect this type of data agent to as many as five Explores.
  • Query a dashboard by using a data agent Preview: Use conversational analytics to generate a dashboard data agent that can query the dashboard and its underlying models. Dashboard agents are available when the Enable Chat with Dashboard setting is enabled on the Gemini in Looker page in the Admin panel.
  • Advanced analytics: The Advanced Analytics feature within conversational analytics translates your natural language questions into Python code and executes that code. Compared to standard SQL-based queries, the Advanced Analytics feature's use of Python enables more complex analysis and visualizations.
  • Embedding: You can embed conversational analytics in a website or an application by using an HTML iframe tag just as you can with other Looker content types. Conversational analytics supports both private embedding, where users are authenticated with their Looker login, and signed embedding, where users are authenticated through your own application.
  • Engagement and token usage monitoring: Users with access to Looker System Activity content can learn more about user engagement with conversational analytics agents and conversations as well as monitor users' token usage. The Conversational Analytics dashboard is available in the System Activity section of the Admin panel. Information about token usage is available when the Conversational Analytics Observability preview feature is enabled.
  • Triggered agentic workflows Preview: Set up metric-based alerts and monitoring directly through natural language queries in standalone conversations with Explores or with Explore data agents. Agentic workflows are available when the Agentic Workflows setting is enabled on the Gemini in Looker page in the Admin panel.

Data agents versus Conversations

Chatting with a data agent offers several key advantages over having a standard conversation with a single Explore. While starting a conversation from an Explore allows for quick natural language querying of that specific data source, a data agent functions as a specialized standalone analyst that can be customized and shared across your organization.

Data agents provide the following advantage over conversations with Explores:

  • Chat with multiple Explores: In a conversation with an Explore, you're limited to querying one Explore at a time. However, a data agent can connect to as many as five distinct Explores, enabling users to perform cross-domain analysis and receive more comprehensive answers.
  • Chat with a Looker dashboard: When you chat with a dashboard, conversational analytics creates a dashboard data agent that you can use to query not only a dashboard but the dashboard's underlying Explores.
  • Authored context: You can provide your data agent with custom instructions that are not available in a standard Explore conversation, including the following resources:
    • Golden queries: You can provide the agent with pairs of natural language questions and verified Looker queries to anchor common business patterns and significantly reduce ambiguity for the model.
    • Business glossaries: You can define organization-specific jargon or acronyms directly within the agent's instructions.
    • Persona framework: You can assign a specific role or expertise to the agent, setting a consistent tone and professional judgment for the conversation.
  • Agent specialization: Rather than using the single generic interface of a conversation, you can create specialized agents for different business units, such as a Revenue Agent or an Ops Agent. This allows for a more guided analysis experience, leading users toward the specific fields and filters most relevant to their needs.
  • Collaboration and reuse: Explore conversations are typically limited to one user, while data agents can be shared with other members of your organization. Sharing ensures that multiple users can benefit from the same authored context and governance that was developed by an administrator or data expert.
  • Custom behavior: You can configure your agent to operate within strict requirements, such as using default filters — for example, "always default to the last 6 months if no timeframe is mentioned." These guardrails ensure that the agent operates within your organization's specific governance and security standards. You can also hide fields in Explores to prevent the data agent from using them in queries.

The following table summarizes the key differences in these types of conversations and data agents:

conversational analytics feature Key use cases Entry point Sharing capabilities Limitations
Explore conversations
  • Standard natural language querying of a single Looker Explore.
  • Organizing related questions about a specific dataset into discrete sessions.
  • Utilizes Fast and Thinking question modes.
  • From the Explore: Select Start a conversation.
  • From the conversational analytics page: Select the Conversations tab.
 Typically limited to one user (not shareable).
  • Limited to querying only one Explore at a time.
  • No custom authored context or instructions (for example, golden queries, personas, business glossaries).
  • No agent specialization for different business units of an organization.
  • The Start a conversation option isn't available when using the new Explore experience.
Explore data agents
  • Specialized, standalone data analytics agents that are designed for specific business units of an organization (for example, Revenue Agent, Ops Agent).
  • Cross-domain analysis that uses up to five connected Explores.
  • Authored context that includes custom instructions, golden queries, business glossaries, and persona frameworks to enforce guardrails and improve accuracy.
  • Advanced analytics that use Python code execution, if enabled.
  • Doesn't require explicit explore permissions.
 From the conversational analytics page: Select the Agents tab. Can be shared with other users in the organization by granting content access (View or Manager access; Edit). Can also be published to external applications like Gemini Enterprise (Preview).
  • Requires access_data and other permissions on the models underlying the Explores (sharing the agent does not grant access to the Explores).
  • Does not support golden queries containing pivots or custom fields.
Dashboard agents
  • Direct natural language querying of a specific Looker dashboard and its query-linked tiles.
  • Curating and customizing agent behavior on dashboards with custom instructions.
  • Access is managed strictly through dashboard-level permissions (Manage access; Edit on the dashboard) and model permissions.
 From the dashboard: Select Chat with this dashboard. Cannot be shared with other users.
  • Cannot be individually shared.
  • Don't appear on the conversational analytics page. Conversations with a dashboard agent are accessible only from the Recent conversations page, which is available from the dashboard menu.
  • Limited to the context of the dashboard and its query-linked tiles.
  • Customizing dashboard agent behavior with custom instructions isn't supported for LookML dashboards.
  • The Advanced Analytics feature isn't supported.

How does conversational analytics work?

Conversational analytics uses Gemini for Google Cloud to interpret natural language questions and provide answers in Looker based on your data. It uses the Looker semantic model — the LookML definitions of your data — as its source of truth to ensure that responses are accurate and consistent. conversational analytics can interpret your business definitions for metrics like "revenue" or "churn" because they are defined in LookML, using these definitions to deliver accurate and consistent questions.

To ground its responses in your specific data and business context, conversational analytics uses several techniques:

  • LookML schema: At the beginning of a request, conversational analytics fetches the schema from the Explores that are connected to it. Conversational analytics uses parameters from the LookML model in several ways:
    • Field identification: Metadata from the schema helps conversational analytics focus on relevant fields. This metadata includes LookML parameters such as name, label, description, type, and dimension_group. These parameters help conversational analytics to map terms in user questions to the correct fields. For example, description can provide business-specific terminology or context for a field. conversational analytics ignores any fields that use the hidden parameter.
    • Response formatting: Conversational analytics uses label parameters for user-friendly field naming and value_format parameters for formatting data in responses.
  • Query generation: Instead of querying your database directly, conversational analytics determines which fields, filters, sorts, and limits should be used in the query. Looker then composes and executes the query by using the underlying LookML model. This process is similar to how a user interacts with an Explore interface; conversational analytics doesn't need to understand complex join logic or field definitions because Looker handles query composition based on the LookML model. Query generation ensures that all queries adhere to the join logic, filtering, aggregation, and data permissions that are defined in your LookML model. To generate queries, conversational analytics must determine the correct values to use in filters. These values must be exact matches of values in the underlying data or more advanced filter expressions such as wildcards. To resolve discrepancies between what values the user includes in their natural language questions and the exact values that may be required by a filter, conversational analytics uses values that are defined with allowed_value in parameter fields and can use tools to check for specific values in fields:
    • Sample data: Returns up to 100 values from a field to help conversational analytics learn patterns or find an exact match for filter values.
    • Fuzzy search: Generates a set of search terms based on user input and checks for their presence in a dimension to find appropriate filter values. Both sample data and fuzzy search use Looker's suggestion API and are therefore influenced by LookML parameters such as suggestions, suggest_explore, and suggest_dimension.
  • Analysis: After Looker runs the queries, conversational analytics analyzes query results to answer user questions. Conversational analytics can analyze results in one or more of the following ways:
    • It uses built-in Gemini capabilities to interpret and summarize results.
    • It uses Python code execution through Advanced Analytics to perform further analysis on the results.
    • It creates visualizations based on query results.

By leveraging the Looker semantic model, conversational analytics provides access to data from various platforms — such as BigQuery, AlloyDB, Redshift, Snowflake, and Databricks — without needing to understand the underlying data complexity, and it ensures that all responses are consistent and governed.

How do conversational analytics data agents work?

A conversational analytics data agent bases its responses on two primary inputs: your instance's LookML schema, which is defined by a Looker developer, and agent instructions, which you write when you create or edit the agent.

From your query, a data agent must determine which LookML fields to select and which filters, sorts, or limits to apply. To do this accurately, it maps the natural language in your query to its own agent instructions and the data's LookML schema in the following ways:

  1. Mapping semantic terms: Users often use business jargon in their questions. The agent uses your agent instructions and the LookML field metadata to interpret your query. For example, for the query "How much new business did we do?" a data agent would be able to map "new business" to a measure that calculates monthly recurring revenue. For the query "Who are our top customers?" a data agent would be able to map "top" to a usage count and "customers" to a dimension called Customer Name.
  2. Mapping field values: The agent looks for specific data points — for example, "California" or "Slim-fit Jeans" — using specialized tools to sample the data or perform fuzzy searches. For example, if a user asks for "jeans," the agent might trigger a fuzzy search across the Product Name field to find the exact string matches in your database. If a user asks for "NY," is that the City field or a State field? The agent may need to sample the data to see which field contains "NY" or ask a clarifying question.
  3. Refinement with query examples: You can provide specific question-and-answer examples, called "golden queries", in data agent instructions to improve accuracy for common or critical queries.

Looker then uses the LookML definitions of these fields and other logic that's defined in the Explore — including field definitions, access grants or user attributes, user-defined dashboard filters, or complex Liquid or join logic — to compose the query that's sent to the database. Since the agent isn't writing the whole SQL query, it doesn't need to "understand" the data and can operate more accurately and deterministically.

To learn more about how your agent uses LookML, see the Best practices for configuring conversational analytics in Looker documentation page.

List of documentation

Understand the compliance capabilities of Gemini in Looker features

Conversational analytics is not yet included in FedRAMP High or FedRAMP Medium authorization boundaries. Before you enable the Gemini in Looker setting for your Looker instance, discuss with your authorizing body whether Gemini for Google Cloud's compliance offerings meet the needs of your organization.

For Looker (Google Cloud core) instances, each Assured Workloads control package that becomes available will add Gemini in Looker features to the default offerings as that package's change requirements and processes are met. Conversational analytics in Looker respects the compliance capabilities of the associated Looker (Google Cloud core) instance, with the following exception:

Data Residency (DRZ) support, specifically for data-at-rest, is available to all Looker customers. All data-at-rest that is associated with conversational analytics resides strictly within the Looker instance and is confined to a single region. Data in transit and machine learning (ML) are processed using a global service.

Compliance in the European Union

Now available for European Union (EU) based customers is a localized processing option that enables data processing within the EU jurisdiction. This option allows Looker's conversational analytics traffic to be routed through an EU Multi-regional Endpoint, which ensures EU customer data is processed within the EU.

Localized processing applies to all conversational analytics in Looker features (excluding dashboard data agents) and all Looker API conversational analytics endpoints. It doesn't apply to the Conversational Analytics API.

This option is available through an allowlist. To request enrollment in the Looker conversational analytics EU ML processing allowlist, complete the Looker ML Processing EU Request form. Note that capacity is limited; enrollment is granted on a functional "as-available" basis and may impact service throughput. You will receive a confirmation email when you have been enrolled.

Provide feedback

You can provide feedback to Google about individual responses in conversational analytics by selecting one of the following options:

  • thumb_up Good response: Indicate that the response was helpful.
  • thumb_down Bad response: Indicate that the response was not helpful.