Vers une génération augmentée de récupération basée sur des graphes de connaissances efficace pour les agents conversationnels

Conversational agents have become widespread in recent years. Today, they have transcended their initial purpose of simulating a conversation with a computer program and are now valuable tools for accessing information and carrying out various tasks, from customer service to personal assistance. With the rise of text-generative models and Large Language Models (LLMs), the capabilities of conversational agents have increased tenfold. However, they are now subject to hallucinations, producing false information. A popular technique to limit the risk of hallucinations is Retrieval Augmented Generation (RAG), which injects knowledge into a text generation process. Such injected knowledge can be drawn from Knowledge Graphs (KGs), which are structured machine-readable knowledge representations. Therefore, we explore Knowledge Graph-based Retrieval Augmented Generation (KG-RAG) to build trusted conversational agents. We demonstrate our approach on a real-world use case for citizen support by building conversational agents for disability management in cities. We first present a history of conversational agents, introducing the approaches implemented over the years and the evaluation techniques. We then define KGs and ontologies, and explore construction and evaluation techniques. As we could not find a directly exploitable KG, our first contribution introduces the Ontology Learning Applied Framework (OLAF). This modular system is built for automated and repeatable KG construction from unstructured text. OLAF integrates linguistic, statistical, and LLM-based techniques to generate Minimum Viable Ontologies for specific domains. Applied to real-world datasets, OLAF demonstrates robust performance through gold-standard evaluations and task-specific Competency Questions. We detail the construction process for a KG about disability management in a French city. We then propose an architecture for KG-RAG systems to enhance information retrieval by aligning user queries with KG structures through entity linking, graph queries, and LLM-based retrieval approaches. We demonstrate our architecture on different use cases, which we evaluate using criteria such as performance, human preference, and environmental impact. While user preferences advantage Text-RAG, KG-RAG's reduced computational footprint underscores its potential for sustainable AI practices. Finally, we identify the critical part of the architecture as the retriever. Therefore, we tackle the retrieval task in our architecture by exploring embeddings in various contexts, i.e. improving EL, retrieval, and providing a caching system. We also propose mechanisms for handling multi-turn conversations. This work establishes a comprehensive framework for KG-RAG systems, combining the semantic depth of KGs with the generative capabilities of LLMs to deliver accurate, contextual, and sustainable conversational agents. Contributions include OLAF for scalable KG construction, a robust KG-RAG pipeline, and embedding-based enhancements for retrieval and interaction quality. By addressing conversational agents' industrial challenges, such as scalability, retrieval precision, and conversational coherence, this research lays the foundation for deploying KG-RAG systems in diverse and specialised domains.