MC-RAG System uses subgraph matching for multi-constraint queries

Standard Retrieval-Augmented Generation (RAG) systems struggle with complex multi-constraint queries, often producing violations, inconsistencies, or hallucinations. To address this, researchers from the CS & Information Retrieval community present MC-RAG (Multi-Constraint RAG), a structure-driven system that reimagines retrieval as a subgraph matching problem over a knowledge graph. The approach combines semantic and structural embeddings with path-level indexing, enabling interpretable, structure-aware, and constraint-consistent retrieval and generation. Unlike naive RAG pipelines, MC-RAG parses multiple constraints explicitly and matches them against graph structures, ensuring that generated answers adhere to all given conditions.

In a demonstration, participants can input medical or encyclopedic multi-constraint queries — e.g., "find drugs that treat hypertension and don't interact with aspirin" — and visualize the full pipeline: constraint parsing, structural matching, and final answer generation. The system outputs explanations for each step, offering transparency often missing in black-box RAG. The paper (arXiv:2607.10151) provides technical details on embedding integration and indexing. A demo video is available online. This work promises to make RAG more reliable for critical domains where multiple constraints must be satisfied simultaneously.