Default LLM Model
Read and set the project's default evaluation model from the terminal with the aievals CLI.
Default LLM Model
The default LLM model is the model your project falls back to whenever an evaluator — or another feature — needs an LLM but the caller did not pick one. Set it once, and later evaluation runs use it automatically.
This is a singleton resource: one default per project. There is no list, no
create, and no ID argument. set covers both create and update, because there
is only one instance to write.
Read the current default
aievals default-llm-model get --format jsonIf no default has been set for the project, the call returns 404 rather than
an empty result.
Set the default
--provider, --adapter, and --model are all required.
aievals default-llm-model set --provider OpenAI --adapter openai --model gpt-4oWith generation settings:
aievals default-llm-model set --provider Anthropic --adapter anthropic --model claude-sonnet-5 \
--model-params '{"temperature":0.2,"max_tokens":1024}'| Flag | Notes |
|---|---|
--provider | Name of a configured LLM connection in this project (required) |
--adapter | Protocol adapter (required) — see below |
--model | Model identifier, e.g. gpt-4o (required) |
--model-params | Generation settings as a JSON object |
--provider vs --adapter. --provider names an LLM connection you have
already configured in this project; its API key is resolved server-side.
--adapter is the protocol used to talk to it.
Adapter values
anthropic, openai, azure, bedrock, google-vertex-ai,
google-ai-studio, groq, mistral, fireworks, databricks
The CLI validates --adapter locally, so a typo fails immediately with the
valid values rather than as a server error.
Run aievals provider list to see the LLM connections configured in your
project, and use the exact name field for --provider.
There is no CLI verb to clear the default. The underlying operation is a delete — the only reset available — and exposing it under a friendlier-sounding name would ship a destructive operation while hiding what it does.
Clearing is still possible from the dashboard, and through the SDK's clear /
delete method. It is the CLI that deliberately omits it.