Scholarlink

Product configuration often requires navigating complex hierarchies, constraints, and interdependent features, making traditional interfaces difficult to use, particularly for non-expert users and in domains involving complex products with numerous interrelated properties. Large Language Models (LLMs) provide a promising alternative by enabling conversational interaction, allowing users to express their configuration goals in natural language. During such dialogues, LLMs can extract user intent, interpret constraints, and map requirements to relevant product properties, thus reducing complexity and simplifying the configuration process. However, despite their fluency, LLMs frequently produce hallucinations, outputs that appear plausible but violate domain constraints or factual correctness, particularly when operating without access to domain-specific knowledge. This limitation arises because LLMs rely primarily on their internal knowledge and lack access to domain-specific product knowledge. To address this challenge, we propose a multi-stage framework that augments LLMs with domain Knowledge Graphs (KGs) to ensure semantic validity and reduce hallucinations. Our approach integrates KGs at two critical stages: (1) Knowledge-Aware Inference, where KG facts are injected into prompts using Retrieval-Augmented Generation (RAG), guiding LLM outputs with verified domain-specific knowledge; and (2) Knowledge-Aware Validation, where generated outputs are post-processed and checked against KG constraints to ensure logical consistency. This integration transforms the LLM into a KG-aware assistant capable of intelligent interpretation and constraint-compliant generation. This framework demonstrates a reliable direction for applying LLMs in domain-sensitive applications, where factual accuracy and constraint compliance are essential. Our framework offers a solution that combines the generative flexibility of LLMs with the symbolic precision of KGs, supports the development of more reliable, user-friendly configuration systems while reducing interface complexity.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-stage Knowledge Graph-Augmented LLMs for Reliable Product Configuration

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Stefan Berlik,
  • Mohammad Seidpisheh

摘要

Product configuration often requires navigating complex hierarchies, constraints, and interdependent features, making traditional interfaces difficult to use, particularly for non-expert users and in domains involving complex products with numerous interrelated properties. Large Language Models (LLMs) provide a promising alternative by enabling conversational interaction, allowing users to express their configuration goals in natural language. During such dialogues, LLMs can extract user intent, interpret constraints, and map requirements to relevant product properties, thus reducing complexity and simplifying the configuration process. However, despite their fluency, LLMs frequently produce hallucinations, outputs that appear plausible but violate domain constraints or factual correctness, particularly when operating without access to domain-specific knowledge. This limitation arises because LLMs rely primarily on their internal knowledge and lack access to domain-specific product knowledge. To address this challenge, we propose a multi-stage framework that augments LLMs with domain Knowledge Graphs (KGs) to ensure semantic validity and reduce hallucinations. Our approach integrates KGs at two critical stages: (1) Knowledge-Aware Inference, where KG facts are injected into prompts using Retrieval-Augmented Generation (RAG), guiding LLM outputs with verified domain-specific knowledge; and (2) Knowledge-Aware Validation, where generated outputs are post-processed and checked against KG constraints to ensure logical consistency. This integration transforms the LLM into a KG-aware assistant capable of intelligent interpretation and constraint-compliant generation. This framework demonstrates a reliable direction for applying LLMs in domain-sensitive applications, where factual accuracy and constraint compliance are essential. Our framework offers a solution that combines the generative flexibility of LLMs with the symbolic precision of KGs, supports the development of more reliable, user-friendly configuration systems while reducing interface complexity.