Introducing Data Token Observability for Conversational Analytics in Looker
Conversational Analytics (CA) in Looker can transform how users interact with data through natural language processing and advanced model reasoning. To help manage this new landscape, token observability is now available in Preview. With Looker Release 26.12, we’ve introduced observability dashboards and Explores to help administrators track and audit consumption of the data tokens that power these AI interactions.
Understanding Data Tokens in Looker
Interactions with Conversational Analytics capabilities in Looker are measured in data tokens—the fundamental units of text and metadata processed by Large Language Models (LLMs). In English, one token is roughly 4 characters, or approximately three-quarters of a word.
All token usage is divided into two distinct categories:
-
Input Data Tokens: The information sent to the model for processing. This includes the user’s natural language question, the conversational history of the session, and the associated LookML schema metadata (fields, descriptions, and agent instructions) required to answer the question.
-
Output Data Tokens: The response generated by the model. This includes the model’s internal reasoning thoughts (visible in Thinking Mode), the generated SQL or Looker API queries, the text response, and the final visualization.
In Looker, input data tokens counts are typically much higher than output token counts. This is because Looker’s semantic layer passes rich metadata to the LLM to ensure accurate and consistent query generation. Based on the input context from Looker, Gemini constructs a Looker API call that is passed back to Looker, with Looker compiling and running a SQL query and returning data based on the underlying semantic definitions.
How to monitor token usage
Starting in Release 26.12, Looker administrators have out-of-the-box visibility into token consumption through System Activity dashboards and explores.
1. Token Usage with System Activity dashboard
Navigate to your System Activity folder and open the new Conversational Analytics dashboard. Here, you can review token observability under the Token Usage tab:
The Token usage tab shows the following data:
-
Total Estimated Input Tokens: The total number of input data tokens that are estimated to be used in Conversational Analytics
-
Total Estimated Output Tokens: The total number of output data tokens that are estimated to be used in Conversational Analytics
-
Daily Estimated Token Usage: A visualization showing the estimated daily use of input and output tokens over time
-
Top Agents by Token Usage: A visualization showing the input tokens that are sent by the agent and the output token responses
Token Usage with System Activity Dashboard
2. Token Usage with System Activity Explores
You can use the native System Activity Explores to build your own custom dashboards. Since Looker tracks telemetry at both the agent and user level, you can easily filter, group, and extract this data for up to a year.
Token Usage with System Activity Explores
Best practices to optimize your Token consumption
A key driver of input token usage in Looker is the Explore size. Consider a single LookML field. A typical LookML field definition that includes parameters like name, label, type, sql, and description averages between 150 and 200 characters. This translates to approximately 40-50 tokens per field that is present and unhidden within an Explore.
Because token consumption scales with the size of your LookML model, you can significantly reduce your footprint and maximize your included quota by following these optimizations:
-
Hide Unused Fields: Apply the hidden: yes parameter to any fields in your Explores that aren’t useful for conversational queries. This dramatically shrinks the LookML schema payload sent to the LLM.
-
Scope Your Explores: Instead of linking your agents to monolithic Explores with thousands of fields, connect your Data Agents to focused, domain-specific Explores. Use extends to build off existing LookML and curate the fields that the agent needs. In System Activity, users can see what fields are used in queries generated by agents and decide on fields to exclude.
-
Enrich Field Metadata Natively: Put descriptions and labels directly into your LookML. Centralizing business logic in LookML keeps downstream system prompts brief and lightweight.
Additional recommendations and best practices are outlined in the Looker Documentation.
Your monthly Token allocation and pricing
To provide maximum flexibility, Looker utilizes instance-based pooled quotas. This model offers several benefits:
-
Shared Efficiency: tokens are aggregated across all authenticated users on an instance rather than being assigned to individuals, enabling more flexibility in how these are utilized.
-
Simplified Tracking: management at an instance level makes governance and spend controls easier to govern and control.
-
Stable Budgeting: pooled usage will be more predictable and will ensure that all included quota is consumed prior to any overage charges being incurred.
For specific monthly token allocations by tier and current overage rates, please refer to the official Looker Pricing Page.
Known limitations
-
24-hour refresh lag: Token metrics presented in the System Activity dashboard are refreshed once every 24 hours. Because of this, current-day numbers are displayed as “estimated actuals” until the next daily refresh.
-
No historical Token data support: Token tracking begins only after the feature is enabled. Consequently, users cannot view token usage for prior Conversational Analytics conversations.
-
No Token count for Looker Agent conversations in Gemini Enterprise: The current feature does not track token usage for Looker agent conversations published in Gemini Enterprise. This capability will be added in the upcoming releases.
What’s coming next
-
The upcoming Looker 26.14 release will introduce token tracking at following granularity:
- Feature-level
- Conversation-level
- User-level
How to get started
We encourage all Looker Admins to access Token usage metrics for CA as soon as their instance is updated to 26.12. To enable this feature, navigate to Admin > Previews and turn on the toggle for Conversational Analytics Agent Token usage.
Enable the Conversational Analytics Agent Token Usage feature from Previews
If you have any questions about accessing your Conversational Analytics token usage, or would like to share your early feedback on token usage, let us know.