A new arXiv preprint tackles a common enterprise problem: a question whose answer depends on data spread across several teams, each of which controls its own tables and will not hand them over. The authors propose an LLM coordinator that routes the question to the relevant owners, who query their own data and return results without exposing the underlying tables.
The paper contrasts two routing strategies. In fan-out, the coordinator sends the question to every owner simultaneously and collects all answers. In relay, it queries owners one after another, so that answers from earlier owners can shape the questions asked of later ones. The trade-off is between parallelism and context: fan-out is faster, while relay can use intermediate results to refine subsequent queries.
The abstract is truncated, so the preprint's findings and recommendations are not fully visible in the source. What is clear is the framing: routing decisions should be ownership-aware, treating each table as a private resource that can only be accessed through its owner's query interface.