ELO-GRAG@EvalLLM 2026: Vers un système HybridRAG combinant graphes de connaissances et texte, évalué sur un corpus de défense en français

We describe ELO-GRAG, our system submitted to the EvalLLM 2026 RAG Challenge. The official corpus contains 375 documents, for a total of 10741 pages. We convert this corpus into annotated Markdown in order to extract the textual content while preserving the document structure, and page level provenance. For Task 1, we compare a RAG system combining dense and lexical retrieval, Microsoft GraphRAG, and two hybrid variants that use an automatically extracted RDF knowledge graph. For Task 2, we formulate source attribution as a controlled LLM-as-a-judge task, with or without additional verbalized triples from the knowledge graph. Preliminary results show that our vector RAG system is the strongest evidence retrieval method, while the best source attribution trade-off is obtained by a Gemma judge using only the textual content of candidate pages.