RAG4RE
A research implementation of retrieval-augmented generation for relation extraction, providing an end-to-end pipeline that runs with and without retrieval across several standard relation-extraction datasets.
The pipeline embeds sentences, builds a similarity index for retrieval, runs a generation step with either a plain or a retrieval-augmented prompt against a chosen language model, and then evaluates the extracted relations against a benchmark dataset, all driven from one command-line entry point and a single configuration file. It accompanies a published paper of the same name in the Semantic Web journal, covering the TACRED, TACREV, Re-TACRED and SemEval datasets.
Because TACRED is a licensed dataset available only through the Linguistic Data Consortium, the repository does not redistribute prompts or raw outputs that would expose its original text, and the related TACREV and Re-TACRED datasets are built from TACRED using separate published tools. The experiments were run on four NVIDIA GeForce GTX 1080 Ti GPUs.