Seismic

Seismic is a fast and lightweight search engine for learned sparse embeddings, written in Rust with Python bindings. It indexes sparse vector collections and retrieves results in microseconds with near-exact accuracy.

Requirements

• Python >= 3.8
• Rust toolchain (only needed if installing from source for hardware-specific optimizations)

Installation

The easiest way to use Seismic is via its Python API, which can be installed in two different ways:

1. the easiest way is via pip as follows:

pip install pyseismic-lsr

2. via Rust compilation that allows deeper hardware optimizations as follows (requires a working Rust toolchain, installable via rustup):

RUSTFLAGS="-C target-cpu=native" pip install --no-binary :all: pyseismic-lsr

Check docs/PythonUsage.md for more details.

Quick Start

Given a collection as a jsonl file, you can quickly index it by running

from seismic import SeismicIndex

json_input_file = "" # Your data collection

index = SeismicIndex.build(json_input_file)
print("Number of documents:", index.len)
print("Avg number of non-zero components:", index.nnz / index.len)
print("Dimensionality of the vectors:", index.dim)

index.print_space_usage_byte()

and then exploit Seismic to retrieve your set of queries quickly

import numpy as np

MAX_TOKEN_LEN = 30

string_type  = f'U{MAX_TOKEN_LEN}'

query = {"a": 3.5, "certain": 3.5, "query": 0.4}
query_id = "0"
query_components = np.array(list(query.keys()), dtype=string_type)
query_values = np.array(list(query.values()), dtype=np.float32)

results = index.search(
    query_id=query_id,
    query_components=query_components,
    query_values=query_values,
    k=10,
    query_cut=3,
    heap_factor=0.8,
)

Each document in the jsonl file should be a JSON object with an id (integer), an optional content (string), and a vector (dictionary mapping tokens to scores, e.g., {"dog": 2.45}). See docs/RunExperiments.md for full format details.

Features

• Multiple index variants — Standard (SeismicIndex), compressed (SeismicIndexDotVByte), and large vocabulary (SeismicIndexLV) for collections with >65K unique tokens
• RAG-ready — Build the index with load_content=True and retrieve document texts alongside scores (example)
• Python & Rust APIs — Use from Python via pyseismic-lsr or integrate directly in Rust via cargo add seismic (docs)
• Parallel batch search — Multi-threaded query processing via batch_search

Examples

Interactive Jupyter notebooks are available in the examples/ folder:

• HandsOnSeismic.ipynb — Quick 2-minute overview of building and querying an index
• SeismicGuide.ipynb — Comprehensive guide covering all features: indexing, k-NN graphs, search, evaluation
• RAG.ipynb — Plug Seismic into a RAG pipeline with document content retrieval
• DotVByteIndex.ipynb — Memory-efficient compressed index variant
• LargeVocabulary.ipynb — Handling collections with large vocabularies (>65K tokens)

Comparison with the state-of-the-art

Comparison with Dynamic Superblock Pruning (DSP) using the splade-v3 encoding of the MS MARCO dataset.

Experiments performed in single-threaded mode on an Intel Core Ultra 7265K CPU, equipped with 124 GB of RAM.

Best Results

Seismic is an approximate algorithm designed for high-performance retrieval over learned sparse representations. We provide pre-optimized configurations for several common datasets, e.g., MsMarco. Check the detailed documentation in docs/BestResults.md and the available optimized configurations in experiments/best_configs.

Resources

Check out our docs folder for detailed guides:

• PythonUsage.md - How to use the Seismic Python API.
• RustUsage.md - How to use Seismic directly in Rust.
• Guidelines.md - Step-by-step guide to build your Seismic index with hyperparameter tuning tips.
• BestResults.md - A detailed guide on how to replicate results with optimized configurations.
• RunExperiments.md - How to run custom experiments, download datasets, and data format details.
• TomlInstructions.md - TOML configuration reference.

Bibliography

@inproceedings{bruch2024seismic,
  author    = {Bruch, Sebastian and Nardini, Franco Maria and Rulli, Cosimo and Venturini, Rossano},
  title     = {Efficient Inverted Indexes for Approximate Retrieval over Learned Sparse Representations},
  booktitle = {Proceedings of the 47th International {ACM} {SIGIR} {C}onference on Research and Development in Information Retrieval ({SIGIR})},
  pages     = {152--162},
  publisher = {{ACM}},
  year      = {2024},
  url       = {https://doi.org/10.1145/3626772.3657769},
  doi       = {10.1145/3626772.3657769}
}

@inproceedings{bruch2024pairing,
  author    = {Bruch, Sebastian and Nardini, Franco Maria and Rulli, Cosimo and Venturini, Rossano},
  title     = {Pairing Clustered Inverted Indexes with $\kappa$-NN Graphs for Fast Approximate Retrieval over Learned Sparse Representations},
  booktitle = {Proceedings of the 33rd International {ACM} {C}onference on {I}nformation and {K}nowledge {M}anagement ({CIKM})},
  pages     = {3642--3646},
  publisher = {{ACM}},
  year      = {2024},
  url       = {https://doi.org/10.1145/3627673.3679977},
  doi       = {10.1145/3627673.3679977}
}

@inproceedings{bruch2025investigating,
  author    = {Bruch, Sebastian and Nardini, Franco Maria and Rulli, Cosimo and Venturini, Rossano and Venuta, Leonardo},
  title     = {Investigating the Scalability of Approximate Sparse Retrieval Algorithms to Massive Datasets},
  booktitle = {Advances in Information Retrieval},
  pages     = {437--445},
  publisher = {Springer Nature Switzerland},
  year      = {2025},
  url       = {https://doi.org/10.1007/978-3-031-88714-7_43},
  doi       = {10.1007/978-3-031-88714-7_43}
}

@article{bruch2026forward,
  title={Forward Index Compression for Learned Sparse Retrieval},
  author={Bruch, Sebastian and Fontana, Martino and Nardini, Franco Maria and Rulli, Cosimo and Venturini, Rossano},
  journal={European Conference on Information Retrieval 2026 (to appear)},
  year={2026}
}

@article{bruch2025efficient,
  title={Efficient Sketching and Nearest Neighbor Search Algorithms for Sparse Vector Sets},
  author={Bruch, Sebastian and Nardini, Franco Maria and Rulli, Cosimo and Venturini, Rossano},
  journal={arXiv preprint arXiv:2509.24815},
  year={2025}
}