This example project demonstrates how to download paper datasets from readme.md, build a DEG graph with preset parameters, run search queries, measure Recall vs. QPS (Queries Per Second), and display the results plot interactively.

• audio (192D, 53k base vectors, L2 distance)
• enron (1369D, 94k base vectors, L2 distance)
• sift1m (128D, 1M base vectors, L2 distance)
• deep1m (96D, 1M base vectors, L2 distance)
• glove (100D, 1.18M base vectors, Angular / InnerProduct distance)

The deglib Python library contains C++ pybind11 bindings (deglib_cpp).

Since pyproject.toml references deglib locally (path = "../../python"), running uv sync automatically copies the C++ sources from cpp/, invokes CMake, and compiles/installs the latest deglib C++ bindings into the local environment:

# Install dependencies & compile latest deglib C++ bindings
uv sync

# If C++ code or pybind11 bindings were modified, force rebuilding the extension:
uv sync --reinstall-package deglib

Note: A C++ compiler (MSVC on Windows, GCC/Clang on Linux/macOS) and CMake must be available in your system PATH. setup.py handles copying C++ files, running CMake, and compiling the extension during the uv sync step.

uv run automatically uses the managed virtual environment.

uv run main.py sift1m

uv run main.py audio --max-base-vecs 1000

• dataset: Positional dataset name (sift1m, deep1m, glove, audio, enron, all). Default: audio.
• --graph-path: Save the generated .deg graph file to this path. Default: none (graph kept in RAM).
• --instruction: SIMD instruction set (auto, scalar, avx2, avx512). Default: auto.
• --force-rebuild: Force rebuilding graph files even if they already exist.
• --threads: Number of threads used for building the graph. Default: 1.
• --cache-dir: Persistent cache folder for datasets. Default: ~/.cache/deg_datasets.
• --no-show: Disable opening the interactive plot window (useful for headless CI environments).
• --max-base-vecs: Limit base vectors for fast debugging / testing.