LLM_MODEL env var for model selection as per-terminal, per-session default

[mcint]

Current models can be specified with llm [prompt] [-m [model-name/alias]] [-|...]. This requires specifying a model on each invocation, requiring the user to add more content manually to each invocation, instead of providing a name or alias before each invocation of the prompt or chat commands.

llm models default can be set globally (or per llm install 😟) but not per terminal session.

[simonw]

This is a great idea!

[simonw]

I'm using LLM_MODEL for default model and LLM_EMBEDDING_MODEL for default embedding model.

[simonw]

Tested the LLM_EMBEDDING_MODEL one manually like this:

llm embed -c 'This is some content' |  jq '.[0:4]'

[
  0.036857523024082184,
  0.030655818060040474,
  -0.023410698398947716,
  0.020912380889058113
]

export LLM_EMBEDDING_MODEL=text-embedding-3-large-1024
llm embed -c 'This is some content' |  jq '.[0:4]'    

[
  0.010029823519289494,
  -0.017291441559791565,
  -0.03479279577732086,
  0.0032044267281889915
]

[simonw]

Tested LLM_MODEL manually like this:

LLM_MODEL=openai/gpt-4.1-nano llm hi
LLM_MODEL=gemini-2.5-pro-exp-03-25 llm hi
llm logs -n 2

2025-04-20T03:36:31 conversation: 01js8m9t79mkchw3hzaqwf9bz7 id: 01js8m9t7bsc0harv9x4esx3sz

Model: openai/gpt-4.1-nano

Prompt

hi

Response

Hello! How can I assist you today?

2025-04-20T03:37:04 conversation: 01js8mawrcbsx9hdz8xjac57t4 id: 01js8mawrj06k76ak3bstns4vd

Model: gemini-2.5-pro-exp-03-25

Prompt

hi

Response

Hello there! How can I help you today?

[simonw]

And for llm chat:

LLM_MODEL=gemini-2.5-pro-exp-03-25 llm chat
Chatting with gemini-2.5-pro-exp-03-25
Type 'exit' or 'quit' to exit
Type '!multi' to enter multiple lines, then '!end' to finish
> your name?
I am a large language model, trained by Google.

[simonw]

And llm embed-multi:

LLM_EMBEDDING_MODEL=text-embedding-3-large-1024 \
  echo '{"id": "one", "content": "This is the first item"}' | llm embed-multi env-var-test -
llm collections | grep env-var-test

Outputs:

env-var-test: text-embedding-3-large-1024

[simonw]

Updated docs:

• https://llm.datasette.io/en/latest/usage.html#executing-a-prompt
• https://llm.datasette.io/en/latest/embeddings/cli.html#llm-embed