Customizing LLM Usage by Project

This guide shows Brim Admins or Project Owners how to control which LLMs each project uses and how many tokens it can spend. Tailoring this by project helps you balance accuracy, speed, and cost across different workflows.

Overview

Brim lets you:

• Choose which models each project uses
• Set token usage limits per project to control cost

These settings are configured for each project in Settings → Manage Project. To add models, set defaults for new projects, or choose which projects can use a model, see Admin LLM Configuration.

You need Manage Project or Administrator permissions to change these settings.

Assigning Models to a Project

Each project has its own LLM Model and Advanced LLM Model. New projects start with the models set in New Project Defaults on the LLM Config page, and you can change them at any time.

This allows you to:

• Use faster, lower-cost models for routine work
• Reserve advanced models for complex abstraction
• Control costs and speed per project

To change a project's models:

1. Go to Settings → Manage Project.
2. Select the project.
3. Choose the models below and save your changes.

• LLM Model: The default model used for most generations.
• Advanced LLM Model: The higher-capability model used for complex tasks.

Only models that are active and available to this project appear in the dropdowns. If a model is missing, an Admin can check its project access on the LLM Config page.

Setting Token Limits

You can cap how many tokens a project uses in Settings → Manage Project with the Maximum Tokens per Generation field.

• The default is 1 billion tokens.
• Leave the field blank for no limit.
• When a project reaches its limit, all token usage for that project is blocked until the limit is raised.

Note: If a generation is running when the limit is reached, it stops. Labels generated before the stop are saved. See Configuring Token Usage by Project for what users will see.

Store Result Trace

Result Trace shows you how Brim produced each value it generated. When this is turned on, you can see every LLM call that ran when you produced your results. This helps you understand why a value came out the way it did. To learn more see How to Use Result Trace.

Best Practices

• Match the model to the work: lower-cost models for routine projects, advanced models only for the hardest abstraction tasks.
• Set conservative token limits to prevent runaway usage.
• Check token usage before large generations so a long run isn't stopped partway through.