LLM plugin providing access to Mistral models using the Mistral API

Install this plugin in the same environment as LLM:

llm install llm-mistral

First, obtain an API key for the Mistral API.

Configure the key using the llm keys set mistral command:

llm keys set mistral

<paste key here>

You can now access the Mistral hosted models. Run llm models for a list.

To run a prompt through mistral-tiny:

llm -m mistral-tiny 'A sassy name for a pet sasquatch'

To start an interactive chat session with mistral-small:

llm chat -m mistral-small

Chatting with mistral-small
Type 'exit' or 'quit' to exit
Type '!multi' to enter multiple lines, then '!end' to finish
> three proud names for a pet walrus
1. "Nanuq," the Inuit word for walrus, which symbolizes strength and resilience.
2. "Sir Tuskalot," a playful and regal name that highlights the walrus' distinctive tusks.
3. "Glacier," a name that reflects the walrus' icy Arctic habitat and majestic presence.

To use a system prompt with mistral-medium to explain some code:

cat example.py | llm -m mistral-medium -s 'explain this code'

The pixtral-12b model is capable of interpreting images. You can use that like this:

llm -m pixtral-12b 'describe this image' -a https://static.simonwillison.net/static/2024/earth.jpg

You can also pass filenames instead of URLs.

All three models accept the following options, using -o name value syntax:

• -o temperature 0.7: The sampling temperature, between 0 and 1. Higher increases randomness, lower values are more focused and deterministic.
• -o top_p 0.1: 0.1 means consider only tokens in the top 10% probability mass. Use this or temperature but not both.
• -o max_tokens 20: Maximum number of tokens to generate in the completion.
• -o safe_mode 1: Turns on safe mode, which adds a system prompt to add guardrails to the model output.
• -o random_seed 123: Set an integer random seed to generate deterministic results.

Mistral sometimes release new models.

To make those models available to an existing installation of llm-mistral run this command:

llm mistral refresh

This will fetch and cache the latest list of available models. They should then become available in the output of the llm models command.

The Mistral Embeddings API can be used to generate 1,024 dimensional embeddings for any text.

To embed a single string:

llm embed -m mistral-embed -c 'this is text'

This will return a JSON array of 1,024 floating point numbers.

The LLM documentation has more, including how to embed in bulk and store the results in a SQLite database.

See LLM now provides tools for working with embeddings and Embeddings: What they are and why they matter for more about embeddings.

To set up this plugin locally, first checkout the code. Then create a new virtual environment:

cd llm-mistral
python3 -m venv venv
source venv/bin/activate

Now install the dependencies and test dependencies:

llm install -e '.[test]'

To run the tests:

pytest