athithyakrishnaa/fedrag

FedRAG is a research system that lets multiple hospitals answer medical questions together without ever sharing their patients' records. Each hospital keeps its documents on its own server, and that data never moves. Instead, the system converts documents into a numerical representation locally and then adds a layer of mathematical noise before sending anything to a shared coordinator. This noise scrambles the representation enough that an outsider cannot reconstruct the original text, yet the coordinator can still use the noisy data to find relevant passages. The coordinator combines results from all participating hospitals using a two-step search: one pass looks for exact keyword matches, another pass looks for meaning-level similarity. The two lists are blended together and then a reranking step reorders the final candidates. Once the best passages are selected, a text-generation model reads them and produces a short answer, typically yes, no, or maybe, for biomedical yes/no questions drawn from the PubMedQA dataset. The project tests several configurations to find where privacy and usefulness meet. At very high noise levels, the search quality collapses entirely. At a specific middle setting (labeled epsilon equals 256 in the results), the system finds the right passage 91 percent of the time and still adds enough noise that an attacker trying to identify which documents belong to which hospital gets results barely better than random guessing. The README includes tables showing these tradeoffs at each noise level tested. Scalability tests show the system runs at roughly the same speed whether there are 5, 10, or 20 hospital nodes, and latency stays under 60 milliseconds per query. Building the shared index takes about 9 minutes regardless of how many nodes participate, suggesting the architecture scales without a major time penalty. This is a student or researcher project, presented as a technical paper with reproducible experiments. It is aimed at people studying privacy-preserving machine learning or healthcare AI, not at a general technical audience seeking a ready-to-deploy product. The code supports swapping out the answer-generation model, and the README documents which model scored best on balanced class coverage across the yes/no/maybe categories.