impact-index

A Python/Rust library for efficient sparse retrieval. Built on Rust with PyO3 bindings for high performance.

impact-index supports both neural IR models with pre-computed floating-point impact scores and traditional BM25 bag-of-words retrieval with performance competitive with Lucene/Pyserini.

Features

• BM25 bag-of-words indexing with stemming (Snowball) and stop words (17 languages)

• Block-Max MaxScore and BMW WAND search with early termination

• SIMD bitpacking compression with quantized impacts and block-max pruning

• One-liner compression: index.compress("/path/to/output")

• Posting list splitting by quantile for term impact decomposition

• BMP (Block-Max Pruning) for fast approximate search

• Document store with zstd compression and key-based retrieval

• Async support for non-blocking search and document retrieval

Installation

From PyPI:

pip install impact-index

From source (requires Rust toolchain):

pip install maturin
maturin develop --release

Contents

• User Guide
  • Building an Index
  • Searching
  • BM25 and Bag-of-Words Indexing
  • Compression and Transforms
  • BMP (Block-Max Pruning)
  • Document Store
• API Reference