Skip to main content

AI Model Selection and Configuration

Buzzbin runs models from a centrally managed catalog. Your team does not manage that catalog, but it can choose which models an organization or repository uses for the main AI tasks.

What the model catalog is​

The model catalog is managed by a system administrator. It contains each model's display name, provider key, USD prices, context window, active status, and display tier. Inactive or archived models are hidden from new selections, while existing assignments remain intact as long as the model can still run.

Model slots​

Review configuration exposes two selectable model slots:

  • Review model: used for code review and review-adjacent work
  • Answer model: used for comment answers when the bot is mentioned

The slots are separate because code review is usually long-context and deeper, while comment answers are shorter and latency-sensitive. You can use a faster or cheaper model for answers while keeping a stronger model for review.

Organization and repository levels​

Models follow the same inheritance pattern as review settings:

  • An organization owner or admin chooses the default review and answer models.
  • Each repository can override either model.
  • If a repository has no override, it inherits the organization value.

If a capability is enabled but no effective model can be resolved, the run does not start. Buzzbin intentionally avoids hidden fallback behavior for model-consuming tasks.

What tiers mean​

Model tiers are a UX hint, not a billing or routing rule. They are commonly used like this:

  • Economy: suitable for short answers or cost-sensitive teams
  • Standard: balanced cost, quality, and latency
  • Premium: suitable for sensitive or high-risk repositories

Actual cost still comes from model price, token usage, and pricing rules. The tier alone does not determine price.

Agentic models and light models​

Code Review runs through the agentic path. The model may take multiple steps, inspect the diff, read files, and emit findings in Buzzbin's standard review format. This path requires a model with the needed agentic or tool-calling capabilities.

Light work such as title, description, or label generation is shorter and uses the platform's centrally selected light-generation model. Organizations and repositories do not choose a separate light model.

Existing runs are not repriced​

When a run starts, Buzzbin snapshots the selected model and prices onto that run. Later catalog, tier, or default changes do not reprice historical runs.

Practical starting point​

Choose the review model based on repository risk, and choose the answer model based on cost and speed. If a repository is noisy or expensive, also review its enabled categories and ignored paths before changing the model.

For related controls, see Configure Reviews.