Skip to main content
Perform the steps in this article to configure your LLM usage for billing directly through Google Vertex AI using the customer key option.
If you are using the Glean Key option, this document does not apply. With the Glean Key option, Glean transparently manages the configuration and provisioning of LLM resources.

Enable access to models in Vertex AI

Navigate to the Vertex AI Model Garden and ensure you have enabled access to the following foundation models from the GCP project where Glean is running.
Model nameHow Glean uses the model
Gemini 3 Pro PreviewThinking mode - Agentic reasoning model used for Assistant and autonomous agents
Gemini 2.5 FlashFast mode - Agentic reasoning model used for Assistant and autonomous agents
Gemini 2.5 ProLarge model used for other, more complex tasks in Glean Assistant
Gemini 2.5 FlashSmall model used for simpler tasks such as followup question generation

Request additional quota from Vertex AI

You will need to submit a standard GCP quota request, which is measured in Requests Per Minute (RPM) and Tokens Per Minute (TPM). Use the filter “base_model:” for the model names in the table below and “region:” for the region your GCP project is running in. Please be aware that quota is not a guarantee of capacity; it is Google’s way of ensuring fair use of shared resources. Your requests might not be served during peak times. For guaranteed capacity, you should contact your Google account team about purchasing Provisioned Throughput. An example with Claude Sonnet 4.5 and US-East5 is shown below. An example of a GCP quota request for Vertex AI, showing fields for service, name, type, dimensions, and value. The dimensions specify the region and base model, and the values show requested tokens per minute and requests per minute.

Capacity requirements

  • Gemini 3.0 Pro (Thinking mode): Glean Assistant uses an average of 35.6k full input, 11.2k cached input, and 2k output tokens per query. This is equivalent to about $0.09 per query based on current Gemini 3.0 Preview Pricing.
  • Gemini 2.5 Flash (Fast mode): Glean Assistant uses an average of 12.5k full input, 5.2k cached inputs, and 418 output tokens per query. This is equivalent to about $0.005 per query based on current Gemini 2.5 Flash Pricing.
These averages were determined by running a large, representative sample of queries. To estimate your weekly Glean Assistant LLM costs, you can multiply your weekly query volume by 0.09forThinkingmodeand0.09 for Thinking mode and 0.005 for Fast mode. Note that actual token usage can vary based on query complexity and document size. To estimate throughput requirements (TPM), find your deployment’s query-per-minute (QPM) rate at a desired percentile (like p90), and then multiply that by the average tokens per query. The following table provides example TPM conversions assuming 0.004 QPM per Daily Active User (DAU), based on historical customer data. TPM per Glean DAU
UsersTPM
500125,000
1000245,000
2500615,000
50001,225,000
100002,450,000
200004,895,000
It is highly recommended to estimate capacity using your deployment’s actual QPM, as QPM per DAU can vary significantly across customers.

Select the model in Glean Workspace

  1. Navigate to Admin Console > Platform > LLM.
  2. Click on Add LLM.
  3. Choose Vertex AI.
  4. For the agentic model, select Gemini 3 Pro Preview for Thinking mode and Gemini 2.5 Flash for Fast mode.
  5. Select Gemini 2.5 Flash for the small model.
  6. Select Gemini 2.5 Pro for the large model.
  7. Click Validate to confirm that Glean can use the models.
  8. After validation, click Save.
To use these models with Glean Assistant, Agentic Engine features must be enabled. Until these features are activated, Glean Assistant will continue to use your previously configured large and small models. You do not need to change your large and small models at this time. Glean will use Application Default Credentials to call the models, so no extra authentication is needed.

FAQ

All data is encrypted in transit between your Glean instance and the Vertex AI service, which operates in the same GCP region as your Glean instance.Please refer to the Vertex AI Generative AI and Data Governance guide. Key points include:
  • Foundation Model Training: Google Cloud does not use Customer Data to train its Foundation Models by default. This means your prompts, responses, and any Adapter Model training data are not used for training Foundation Models.
  • Prediction: Inputs and outputs processed during Prediction are considered Customer Data. Google never logs this Customer Data unless a customer explicitly opts in to allow caching.

Architecture diagram

The diagram below illustrates the process flow: A user’s query is processed through query planning, tool selection, query execution, and finally answer generation, utilizing the Glean Planner, Glean Index & Knowledge Graph, and Google Vertex AI. A system architecture diagram illustrating a user query being processed through a tool selection and query planning module, query execution, and answer generation, utilizing a Glean Planner, Glean Index & Knowledge Graph, and Google Vertex AI.