A Tripartite Perspective on GraphRAG

 arXiv:2504.19667v1 [cs.LG] 28 Apr 2025

 

 Large Language Models (LLMs) have shown remarkable capabilities across various domains, yet they struggle with knowledgeintensive tasks in areas that demand factual accuracy, such as in industrial automation and healthcare. Key limitations include their tendency to hallucinate, lack of source traceability (provenance), and challenges in timely knowledge updates. Retrieval Augmented Generation (RAG) techniques have attempted to address these issues by incorporating external knowledge, but they face their own limitations,.... Combining language models with knowledge graphs (GraphRAG) offers promising avenues for overcoming these deficits. However, a major challenge lies in creating such a knowledge graph in the first place. 

Construir KG usando LLM só empurra o problema para o KG

While language models (LLMs) have demonstrated impressive capabilities, they still have their limitations in knowledge-intensive tasks - especially in areas where factually correct information is essential ....

Factual mas CoaKG contém alegações contextualizadas

Another central problem is the lack of provenance, i.e. the lack of traceability of the information source, which makes it difficult to assess the trustworthiness of its output.

O KG precisa ter a proveniência já que não é a fonte primária

A major challenge of common RAG approaches is the selection of appropriate text chunks w.r.t a given query, chiefly because of the reliance on embeddingbased similarity for textual chunk selection, which can be unreliable or lack coverage due to fluctuating chunk sizes and topic diversity within queries. To this end, our approach provides a natural remedy. Applying a transformation to the constructed knowledge graph, we formulate LLM prompt creation as an unsupervised node classification problem.

Não seria necessário GQL na CoaKG Engine, se usar LLM para a consulta em linguagem natural. Mas o LLM seria capaz de deduzir contexto explícito como relações temporais e espaciais não representadas no grafo? A Algebra de Contradomínio poderia ser substituída por Prompt Engeneering?