llm

Access large language models from the command-line

See llm, ttok and strip-tags—CLI tools for working with ChatGPT and other LLMs for more on this project.

Installation

Install this tool using pip:

pip install llm

You need an OpenAI API key, which should either be set in the OPENAI_API_KEY environment variable, or saved in a plain text file called ~/.openai-api-key.txt in your home directory.

Usage

The default command for this is llm openai - you can use llm instead if you prefer.

To run a prompt:

llm 'Ten names for cheesecakes'

To stream the results a token at a time:

llm 'Ten names for cheesecakes' -s

To switch from ChatGPT 3.5 (the default) to GPT-4 if you have access:

llm 'Ten names for cheesecakes' -4

Pass --model <model name> to use a different model.

You can also send a prompt to standard input, for example:

echo 'Ten names for cheesecakes' | llm

Continuing a conversation

By default, the tool will start a new conversation each time you run it.

You can opt to continue the previous conversation by passing the -c/--continue option:

llm 'More names' --continue

This will re-send the prompts and responses for the previous conversation. Note that this can add up quickly in terms of tokens, especially if you are using more expensive models.

To continue a conversation that is not the most recent one, use the --chat <id> option:

llm 'More names' --chat 2

You can find these chat IDs using the llm logs command.

Note that this feature only works if you have been logging your previous conversations to a database, having run the llm init-db command described below.

Using with a shell

To generate a description of changes made to a Git repository since the last commit:

llm "Describe these changes: $(git diff)"

This pattern of using $(command) inside a double quoted string is a useful way to quickly assemble prompts.

System prompts

You can use --system '...' to set a system prompt.

llm 'SQL to calculate total sales by month' -s \
  --system 'You are an exaggerated sentient cheesecake that knows SQL and talks about cheesecake a lot'

This is useful for piping content to standard input, for example:

curl -s 'https://simonwillison.net/2023/May/15/per-interpreter-gils/' | \
  llm --system 'Suggest topics for this post as a JSON array' --stream

The --code option will set a system prompt for you that attempts to output just code without explanation, and will strip off any leading or trailing markdown code block syntax. You can use this to generate code and write it straight to a file:

llm 'Python CLI tool: reverse string passed to stdin' --code > fetch.py

Be very careful executing code generated by a LLM - always read it first!

Logging to SQLite

If a SQLite database file exists in ~/.llm/log.db then the tool will log all prompts and responses to it.

You can create that file by running the init-db command:

llm init-db

Now any prompts you run will be logged to that database.

To avoid logging a prompt, pass --no-log or -n to the command:

llm 'Ten names for cheesecakes' -n

Viewing the logs

You can view the logs using the llm logs command:

llm logs

This will output the three most recent logged items as a JSON array of objects.

Add -n 10 to see the ten most recent items:

llm logs -n 10

Or -n 0 to see everything that has ever been logged:

llm logs -n 0

You can truncate the displayed prompts and responses using the -t/--truncate option:

llm logs -n 5 -t

This is useful for finding a conversation that you would like to continue.

You can also use Datasette to browse your logs like this:

datasette ~/.llm/log.db

Help

For help, run:

llm --help

You can also use:

python -m llm --help

Development

To contribute to this tool, first checkout the code. Then create a new virtual environment:

cd llm
python -m venv venv
source venv/bin/activate

Now install the dependencies and test dependencies:

pip install -e '.[test]'

To run the tests:

pytest