hae

A command line tool for semantic search over text input.

Example:

~> man ls | hae "how to show file sizes" -n 1 -hl
   The Long Format
     If the -l option is given, the following information is displayed for
     each file: file mode, number of links, owner name, group name, number of
     bytes in the file, abbreviated month, day-of-month file was last
     modified, hour file last modified, minute file last modified, and the
     pathname.

hae is best suited for use cases where your input text contains up to thousands of paragraphs. Try it with some text on the clipboard, an RSS feed converted to JSON, or with a cup of coffee ☕ and a nice book 📖 from Project Gutenberg!

Disclaimer: hae is experimental software intended primarily for interactive use, and should not be considered ready for production applications.

Features

• Highlights best-matching sentences, which makes it easy to quickly evaluate the results.
• Optional JSON input and output. Input is automatically interpreted as JSON if it is in the "JSON lines" format, one object per line, and each object contains a field content. If a title field is present, it will also be used for the search.

Installation

For Linux and Apple Silicon Macs you can download a prebuilt binary. You can also build from source as described below.

Build from source

The repo contains a build script that also downloads the SentenceTransformers embedding model and converts it to ONNX format using Python libraries. So you will need some tools installed.

• cmake (tested to work with 3.28.1)
• clang
• rust (tested to work with 1.75.0, required for Tokenizers dependency)
• wget (for fetching ONNX runtime and model files)
• xxd (for embedding model files in headers)
• jq (for preprocessing the tokenizer config file)
• Python (3.10)

The build steps are as follows:

• First fetch the git submodules: git submodule update --init --recursive --depth=1.
• Then run the build: ./build.sh $ARCH where $ARCH is one of the following: osx-arm64, linux-x64, osx-x86_64

The application and the ONNX runtime dynamic library required to run it will be found under ./dist.

Acknowledgements

• This project is only possible thanks to SentenceTransformers https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
• ONNX Runtime for efficient CPU inferencing https://onnxruntime.ai/
• HuggingFace Tokenizers via a C++ wrapper https://github.com/mlc-ai/tokenizers-cpp
• https://github.com/p-ranav/argparse for creating the CLI
• https://rapidjson.org/ for JSON I/O