A collaborative Multi-Agent LLM Approach for Knowledge Graph Curation and query from multimodal data sources

Ankur Krishna, surya ardham, Chetan Malhotra, Anagha Prashant Shinde

Keywords: Large Language Models (LLMs), Knowledge Graph Curation, Retrieval-Augmented Generation (RAG), Multi-Agent Strategy, Information Retrieval

Abstract: Retrieval-Augmented Generation (RAG) systems have demonstrated considerable effectiveness in querying private, short, unstructured data; however, they often encounter challenges in delivering accurate factual answers when working with larger corpora, frequently lacking context and failing to establish domain relationships. In this paper, we introduce a novel collaborative multi agent Retrieval-Augmented Generation (CoMaKG-RAG) framework designed to enhance the capabilities of large language models (LLMs) in complex information retrieval scenarios involving multimodal data sources.Our framework comprises a pool of customized collaborative agents, including a query generator agent, a domain model generator agent, a domain model populator agent, a knowledge graph curator agent, and a knowledge graph query agent, each tailored through a developed customization model and historical domain questions. The query generator formulates relevant queries related to text and image chunks within documents, while the domain model generator constructs a structured domain model based on these queries. The domain model populator agent enriches the model by integrating additional text and image fragments, and the knowledge graph generator assembles a comprehensive unified knowledge graph using Neo4j.Each agent interacts with one another, evaluates outputs, and provides feedback to enhance the overall process. Ultimately, user queries are transformed into cipher queries using the knowledge graph query agent, processed by a unified knowledge graph engine, and converted back into natural language responses. This approach enhances information retrieval from multimodal sources by mitigating hallucinations, generic responses, incomplete responses, and factual inaccuracies. We evaluated our method against the publicly available technical report "Operations & Maintenance Best Practices" and state-of-the-art knowledge graph generation and query software, Neo4j Graph Builder. Our results demonstrate that our method identifies a substantially higher number of entities and uncovers unique, contextually significant relationships, surpassing the performance of the graph builder in both the quantity and quality of extracted information. The proposed agentic graph RAG system was evaluated on both factual and descriptive queries and was able to provide accurate responses for both text and image-based questions, whereas the Neo4j graph performed sub optimally.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3637