refactx 0.1.0

Refactx package

ReFactX: Scalable Reasoning with Reliable Facts via Constrained Generation

Riccardo Pozzi, Matteo Palmonari, Andrea Coletta, Luigi Bellomarini, Jens Lehmann, Sahar Vahdati

The implementation corresponding to the ISWC 2025 paper is available at the ISWC2025 branch.

A preprint that has not undergone peer review is available at https://arxiv.org/abs/2508.16983.

We present ReFactX, a scalable method that enables LLMs to access external knowledge without depending on retrievers or auxiliary models. Our approach uses constrained generation with a pre-built prefix-tree index. Triples from Wikidata are verbalized in 800 million textual facts, tokenized, and indexed in a prefix tree for efficient access. During inference, to acquire external knowledge, the LLM generates facts with constrained generation which allows only sequences of tokens that form an existing fact.

This repository contains the source code for using ReFactX and reproducing our work accepted at ISWC 2025.

Setup

• install the requirements pip install -r requirements.txt
• prepare the .env file: cp env-sample.txt .env, then edit .env (can be skipped if using the simple index in the try_refactx notebook)

Try ReFactX

For quickly trying ReFactX with an in-memory prefix tree (derived from a 31k-facts knowledge base) use the notebook try_refactx.ipynb.

Wikidata Prefix Tree

Refer to PrefixTree.md for creating the Wikidata prefix tree we used in our work.

Experiments

To reproduce our experiments use the eval.py script replacing INDEX, MODEL, and DATASET according to your needs (each of them is a python file to import).

python eval.py --index INDEX --model MODEL --dataset DATASET

Throughtput

For the throughput experiment run

python throughput.py --model MODEL --index INDEX --max-tokens 4001 --output out.json [--unconstrained-generation]

Cite

@InProceedings{10.1007/978-3-032-09527-5_16,
      author="Pozzi, Riccardo
      and Palmonari, Matteo
      and Coletta, Andrea
      and Bellomarini, Luigi
      and Lehmann, Jens
      and Vahdati, Sahar",
      editor="Garijo, Daniel
      and Kirrane, Sabrina
      and Salatino, Angelo
      and Shimizu, Cogan
      and Acosta, Maribel
      and Nuzzolese, Andrea Giovanni
      and Ferrada, Sebasti{\'a}n
      and Soulard, Thibaut
      and Kozaki, Kouji
      and Takeda, Hideaki
      and Gentile, Anna Lisa",
      title="ReFactX: Scalable Reasoning with Reliable Facts via Constrained Generation",
      booktitle="The Semantic Web -- ISWC 2025",
      year="2026",
      publisher="Springer Nature Switzerland",
      address="Cham",
      pages="290--308",
      isbn="978-3-032-09527-5",
      doi="10.1007/978-3-032-09527-5_16",
      url="https://doi.org/10.1007/978-3-032-09527-5_16"
}