DCTR: Dual-Constraint Subgraph Optimization for Knowledge Graph-based Retrieval-Augmented Generation

Large Language Models (LLMs) have been proven to be highly effective in various language modeling tasks. However, LLMs still suffer from intrinsic limitations when it comes to capturing factual and up-to-date data. This is becoming a bigger challenge in organizations that work with proprietary data, which is updated on daily bases, and cannot be accessed by public LLMs for legal reasons. This often requires re-training proprietary LLMs within the organization, or fine-tuning public LLMs on specific tasks using internal data, which is time-consuming and rather costly. To address these challenges, Retrieval Augmented Generation (RAG) approaches have been introduced, which retrieve relevant knowledge from available data stores, leading to higher accuracy, and easy reuse of pre-trained public LLMs. In this work we introduce a Knowledge Graph (KG)-enhanced retrieval augmented generation approach, tailored specifically for the e-Commerce domain. We use a relationship-rich inventory-based Knowledge Graph to identify the most relevant knowledge for the given input and task, using entity linking and KG embeddings, which is then injected in the LLM prompt. We combine the power of LLMs for natural language understanding and the power of KGs for quick and easy access to proprietary factual knowledge to generate high-quality results, omitting hallucinations and generic outputs. We evaluate our approach on three e-Commerce tasks, significantly outperforming baseline LLM models, in both zero-shot and instruction-tuned settings.