llm-grep

LLM plugin for matching text using semantic regular expressions. The high-level matching technique is loosely based on this paper. Matching is done in two passes: one using traditional regular expressions to narrow down candidate matches, and a second pass using an LLM to filter those candidates based on any semantic tags.

The pattern syntax is similar to traditional regular expressions (enclosed in {{ and }}), but adds semantic tags (enclosed in < and >) to indicate the type of concept being matched.

Installation

Install this plugin in the same environment as LLM.

llm install llm-grep

Usage

The plugin adds a new command, llm grep. This command has an interface similar to the GNU grep command, but extends it with semantic matching capabilities, using an LLM as a matching oracle.

Input can be a standard file or stdin. Simple examples you can try:

# Match lines from a file
llm grep -e '^{{.*}}<about birds>$' notes.txt

# Read from standard input
cat recipes.txt | llm grep -e '^{{.*}}<baking related>$'

# Enable color, only output matched content, and use a custom model and prompt:
llm grep --color always -o -e '{{[A-Za-z0-9]+}}<outdoor activities>' --model gpt-4\
    --prompt 'Answer yes or no. Query: {query} Text: {span}.' headlines.txt

# Slightly more specific capture (will not match names like 'sparrow')
llm grep -e '\b{{[A-Z][a-z]+(?:\s+[A-Z][a-z]+)?}}<bird species>\b' bird_log.txt

The default prompt used is:

Does the following text satisfy the semantic concept described by the query?

Query: {query}

Text: {span}

Your answer should include "yes" or "no".

Development

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

cd llm-grep
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

Future plans

• Currently, some of grep's functionality is not implemented (e.g. -r for recursive searching). Re-implementing these features is a losing battle, and will be deprioritized in favor of somehow hooking into (or wrapping around) existing grep implementations.