On-Device Retrieval-Augmented Generation over Personal Clinical Records

Personal health assistants aim to answer questions about a patient's own clinical history by grounding responses in personal medical records. Retrieval-augmented generation (RAG) is well suited to this setting, but existing clinical question-answering benchmarks often focus on single documents, use multiple-choice formats that reduce the role of retrieval, or lack evidence annotations needed to distinguish retrieval failures from generation errors. This thesis addresses these limitations through a patient-level benchmark for RAG over personal clinical records. Built from real, de-identified MIMIC-III clinical text, each patient's history is represented as a collection of per-admission PDF documents and paired with open-ended questions, reference answers, and gold evidence chunks. The benchmark includes both single-document and multi-document questions requiring evidence across admissions, including longitudinal questions that capture changes in a patient's clinical history over time. The thesis also evaluates the feasibility of running the complete RAG pipeline directly on a consumer smartphone, where sensitive clinical data can remain on-device but retrieval and generation must share limited memory and computational resources. Lexical, neural, and hybrid retrievers are compared together with quantized generative models from different families and scales. Performance is assessed at both retrieval and generation level, measuring evidence retrieval, ranking quality, and final answer correctness. The results show that patient-level clinical question answering is feasible under mobile constraints and highlight effective evidence retrieval as a key factor for reliable generation, particularly for longitudinal and multi-document questions.