LLM plugin for semantically sorting lines. Ranking techniques are based on this paper.

Install this plugin in the same environment as LLM.

llm install llm-sort

The plugin adds a new command, llm sort. This command has an interface similar to the GNU sort command, but instead of sorting lines based on alphabetical order, it input lines based on some semantic ranking criteria.

Lines can be provided via files or standard input. For example:

# Sort lines from a file using the "sorting" method (default)
llm sort --query "Which of these names is more suitable for a pet monkey?" names.txt
# Read from standard input
cat titles.txt | llm sort --query "Which book should I read to cook better?"
# Use a different method, limit the output to the top 5 lines, and specify a custom model and prompt:
llm sort --query "Rank these slogans" --method sliding --top-k 5 --model gpt-4 \
  --prompt 'Decide which line is more compelling. Answer with "Line A" or "Line B" Query: {query} Lines: {docA} {docB}.' quotes.txt

Note: The prompt template variables are {query}, {docA}, and {docB}.

The default prompt used is:

Given the query: {query}

Compare the following two lines:

Line A: {docA}

Line B: {docB}

Which line is more relevant to the query? Please answer with "Line A" or "Line B".

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

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

Now install the dependencies and test dependencies:

pip install -e '.[test]'

To run the tests:

python -m pytest