We introduce GRAB, a constructor–encoder–bridge pipeline for table question answering. Our method lifts relational data into a heterogeneous graph, encodes it via message passing, and transfers the signals to an LLM through a small set of query-conditioned latent tokens. The LLM receives these tokens alongside the serialized table, and the two play different roles: the graph shows where the relevant evidence is, while the text keeps the exact cell values that the graph does not encode. The graph supplements the text rather than replacing it, and our ablations confirm that both are required. Crucially, the LLM remains strictly frozen to preserve its general reasoning capabilities; we train only the lightweight graph encoder and latent bridge (91M parameters), allowing the entire pipeline to be trained efficiently. Our pipeline significantly improves performance on relational Question Answering, with the largest gains in demanding multi-table settings, offering an efficient, principled way to connect relational deep learning with LLMs.