Dense Vectors in KGs

This repository contains an experimental fork of Qlever that has built-in vector search.

Setup

Prerequisites:

• uv
• docker
• gcc-13/cmake/`Ninja

Install Dependencies (Ubuntu/Debian)

apt install -y libboost1.88-dev libboost-iostreams1.88-dev libboost-random1.88-dev libboost-program-options1.88-dev libboost-url1.88-dev libboost-container1.88-dev libopenblas-dev

Setup environment

git submodule init
git submodule update
git submodule sync
uv sync # assumes uv is installed
source .venv/bin/activate
cd faiss
bash ./install_with_uv.sh --cpu

Compiling qlever

Ensure that you have a recent version of gcc+ cmakeinstalled!

cd qlever
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DLOGLEVEL=INFO -DUSE_DENSE_TENSOR_INDEX=true -DQLEVER_USE_TENSOR_BLAS=true -DUSE_PARALLEL=true -D_NO_TIMING_TESTS=ON -DCOMPILER_SUPPORTS_MARCH_NATIVE=FALSE -GNinja ..
cmake --build . -j --target qlever-index qlever-server

Add the executable's path to your PATH:

echo "\nexport PATH = \"$PWD:\$PATH\"" >> ~/.zshrc

Generating the Benchmark Datasets

Please refer to the specific instructions.

Getting Neo4j

We also benchmark against Neo4j by converting the graph to a property graph using the neosemantics library. To set it up, download Neo4j, and install it in a subfolder. The database hook might need updating to the correct path in the class.

Running the benchmarks

Now that everything is (hopefully) set up, we can run the benchmarks:

cd benchmarks
python run_bsbm_bench_full.py
python run_dbpedia_bench_full.py

You can use --help to inspect possible options w.r.t paths and other possible options. Once the benchmarks have successfully completed, the figures can be generated using the notebook in benchmarks/eval_timings.ipynb.