Skip to content
MChatzakisPublic

About

[SIGMOD 2026] DARTH and DARTH+: Vector Search with Declarative Recall and Probabilistic Lower Bounds

Topics

Resources

Stars

32 stars

Watchers

0 watching

Forks

Repository files navigation

DARTH logo

DARTH / DARTH+

Vector Search with Declarative Recall and Probabilistic Lower Bounds

Manos Chatzakis, Yannis Papakonstantinou, Themis Palpanas

⚠️ The code recently went under refactoring. Please contact us if you notice issues ⚠️

DARTH and DARTH+ enable declarative recall for approximate nearest-neighbor search: specify a target recall, and a learned predictor estimates when to stop each query. Their adaptive early termination adjusts to query difficulty and reduces unnecessary search work. DARTH+ extends DARTH with more efficient training data collection and query-dimension-aware prediction features. Its lower-bound variant can also provide optional probabilistic lower bounds on recall.

How to Run

Build the pipeline from the repository root:

cmake -S . -B build-native -DCMAKE_BUILD_TYPE=Release \
  -DFAISS_ENABLE_GPU=OFF -DFAISS_ENABLE_PYTHON=OFF \
  -DLIGHTGBM_ROOT="$HOME/lightgbm-install"
cmake --build build-native --target darth_pipeline -j 4

Adjust LIGHTGBM_ROOT for your installation. If the runtime loader cannot find LightGBM, add its library directory to LD_LIBRARY_PATH:

export LD_LIBRARY_PATH="$HOME/lightgbm-install/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"

For example, run DARTH+ on SIFT100M with HNSW:

build-native/darth-runners/darth_pipeline \
  --dataset SIFT100M --dataset-dir-prefix /path/to/processed-data \
  --index hnsw --method darth-plus --k 50 \
  --index-filepath /path/to/indexes/sift-hnsw.index \
  --threads 12 --training-threads 12

Datasets

The pipeline reads .fvecs vectors and .ivecs exact-neighbor labels. For example, SIFT100M expects:

DATASET_DIRECTORY/SIFT100M/
  base.100M.fvecs
  learn.1M.fvecs
  validation.10K.fvecs
  query.10K.fvecs
  learn.groundtruth.1M.k1000.ivecs
  validation.groundtruth.10K.k1000.ivecs

See VectorDataLoader.cpp for more information.

Dataset sources: SIFT and GIST, DEEP and Text2Image, and GloVe.

References

If you use DARTH or DARTH+, please cite:

@article{chatzakis2025darth,
  title={Darth: Declarative recall through early termination for approximate nearest neighbor search},
  author={Chatzakis, Manos and Papakonstantinou, Yannis and Palpanas, Themis},
  journal={Proceedings of the ACM on Management of Data},
  volume={3},
  number={4},
  pages={1--26},
  year={2025},
  publisher={ACM, New York, NY, USA}
}

@article{chatzakis2026darth+,
  title={DARTH+: Approximate Nearest Neighbor Search with Declarative Recall and Quality Guarantees},
  author={Chatzakis, Manos and Papakonstantinou, Yannis and Palpanas, Themis},
  journal={HAL Preprint},
  year={2026}
}

@article{chatzakis2026declarativephd,
  title={Declarative Recall for Approximate Vector Search},
  author={Chatzakis, Manos},
  journal={Proceedings of the VLDB Endowment. ISSN},
  volume={2150},
  pages={8097}
}

@inproceedings{chatzakis2026darth+workshop,
  title={DARTH+: Vector Search with Declarative Recall and Guarantees},
  author={Chatzakis, Manos and Papakonstantinou, Yannis and Palpanas, Themis},
  booktitle={The 2nd Workshop on Vector Databases},
  year={2026},
}

About

This repository includes FAISS code, which remains subject to its original license. For questions, contact the authors. The DARTH logo was created by Eva Chamilaki.

About

[SIGMOD 2026] DARTH and DARTH+: Vector Search with Declarative Recall and Probabilistic Lower Bounds

Topics

Resources

Stars

32 stars

Watchers

0 watching

Forks

Contributors

Languages