mendhak/llm-cli-helper

CLI helper tool to lookup commands based on a description

A basic CLI lookup tool. Describe a bash command and it outputs sample line(s) by quering LLMs. It can make use of OpenAI (GPT 3.5) or Llama.cpp models.

Usage

Try a few commands

$ ? how much disk space 

df -h

$ ? show top processes by CPU usage

top -o %CPU

There is a history, so the next question can be a follow up. Example:

$ ? find .pickle files in this directory

find . -type f -name "*.pickle"

$ ? delete them

find . -type f -name "*.pickle" -delete

Another example, I didn't like the first output so asked for nc instead.

$ ? check if port 443 on example.com is open

echo | telnet example.com 443

$ ? using nc

nc -zv example.com 443

Set up

Set up dependencies

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Make a copy of the .env.sample file and call it .env

OpenAI setup (simple)

Get an API key from your OpenAI account. Place it in the .env file

OPENAI_API_KEY="......................................."

There is a small cost associated with OpenAI calls, so it's a good idea to set monthly limits on usage.

Create an alias

The application is best used as an alias called ?. Add it to ~/.bashrc like so:

# add alias
echo alias ?="\"$(pwd)/.venv/bin/python3 $(realpath openai.clihelper.py)\"" >> ~/.bashrc
# reload bash
exec bash

Now start using ?

Llama.cpp models (uses GPU)

Llama.cpp is a fast way of running local LLMs on your own computer. It is very fast with GPUs which will be my focus. It is free to use.

Install dependencies

First, ensure that CUDA Toolkit is installed. After installing Cuda, add it to your path and reload bash:

echo 'export PATH="/usr/local/cuda/bin:$PATH"' >> ~/.bashrc
exec bash
# test that it worked: 
nvcc --version

Next install the cmake and Python dependencies, and build one specific package with GPU support.

sudo apt install make cmake
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python --force --no-cache

Get some models

Because Llama is open, there are many Llama models you can choose from. Llama.cpp requires models to be in the GGML format. Here are some I tested with:

TheBloke/Llama-2-7B-Chat-GGML - llama-2-7b-chat.ggmlv3.q4_0.bin
TheBloke/StableBeluga-7B-GGML - stablebeluga-7b.ggmlv3.q4_0.bin

Download these then set the path to the model in the .env file. Example:

 LLAMA_MODEL_PATH="./models/7B/stablebeluga-7b.ggmlv3.q4_0.bin"

Create an alias

The application is best used as an alias called ?. Add it to ~/.bashrc like so:

# add alias
echo alias ?="\"$(pwd)/.venv/bin/python3 $(realpath llamacpp.clihelper.py)\"" >> ~/.bashrc
# reload bash
exec bash

Now start using ?

Implementation notes

This was made using langchain, a library that helps make calls to large language models (LLMs) and process its output.

In this case I did a 'few shot', which is a way of showing the LLM a few examples of questions and the kind of answers to generate.

I chose the gpt-3.5-turbo model which is the cheapest on OpenAI currently.