PhD "Knowledge Integration and Traceability in a GraphRAG-Based Question-Answering System"

about the role

Your role is to conduct PhD thesis on the "Knowledge Integration and Traceability in a GraphRAG-Based Question-Answering System"

Problem Statement
The rise of AI conversational agents, has transformed the way information is searched. Tools like Le Chat or ChatGPT have proven effective in information retrieval tasks. However, these tools face limitations when it comes to leveraging up-to-date and company-specific knowledge. This gap has led to an increasing demand for solutions capable of exploiting internal company databases.
Enterprise Knowledge Graphs (EKGs) are emerging as a strategic resource covering the technical and organizational domains of the company. These graphs have reached notable maturity and offer significant potential to enhance the precision and traceability of information retrieval systems. However, effectively integrating these graphs with large language models (LLMs) remains a major challenge.
Scientific Objective
The central problematic of this PhD thesis is to enhance the robustness, precision, traceability, and autonomy of an information retrieval system based on the synergy between a Large Language Model (LLM) and an Enterprise Knowledge Graph (EKG) using a GraphRAG architecture. This approach aims to overcome the current limitations of Generative AI tools (e.g., hallucination, mistrust, information obsolescence) by leveraging the specific and up-to-date knowledge of the enterprise.

One of the key challenges is the balanced injection of knowledge, avoiding the "lost in the middle" information problem. Additionally, user trust in the system is crucial for its adoption. Therefore, it is essential to design an operational mode that strengthens this trust while encouraging users to contribute to the enrichment of the knowledge graphs. Furthermore, the autonomy of LLM agents can be improved through a better understanding of user intentions and the orchestration of specialized models for specific tasks. The responses provided to users can become increasingly personalized as the system is used and the history of past interactions is taken into account.
The objectives of this thesis are crystallized in the following tasks :

• Refine Context Collection: Enhance the collection of contexts by an orchestrating agent and explore new re-ranking methods that leverage the knowledge graph.
• Develop mechanisms to manage the traceability and transparency of generated responses by providing sources in a mode that allows broader consultation and exploration of knowledge. Define a virtuous loop between user feedback collection, graph modification, and updating the information retrieval engine.
• Enhance User Intention Understanding: Improve the understanding of user intentions and define a complex action plan (data collection, service execution, e.g., API Orange Developer).
• Develop a Prototype Service: Design, develop, and deploy a prototype service that makes the enterprise's data more exploitable.