LP-RAG: A Link Prediction-Based Framework for Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) strategies have empowered large language models (LLMs) through integration with external knowledge sources, thereby enabling more accurate, up-to-date, and contextually relevant outputs. Among these, graph-based RAG methods stand out as particularly prominent. These approaches aim to structure external knowledge into graphs and leverage relational reasoning to retrieve relevant information to the context at hand. However, existing approaches remain limited in their ability to exploit query-based semantic cues. In this paper, we propose LP-RAG, a link prediction-based framework for document RAG. Specifically, LP-RAG employs an LLM-prompted chunker and text encoders to construct a graph of similarity relationships among chunks, which is then augmented with chunk-conditioned synthetic queries that emulate potential questions for each chunk. This design enables the incorporation of chunk-specific semantic information for model training. In our framework, retrieval is cast as an inductive link prediction problem, where the goal is to predict chunk–query links. Notably, LP-RAG is model-agnostic and can incorporate any link prediction method (e.g., graph neural network–based predictors). To demonstrate its effectiveness, we evaluate LP-RAG across diverse settings/benchmarks. Results show that LP-RAG consistently outperforms existing RAG methods.