For a genomics app the driver is privacy: raw genomic data is highly sensitive PHI, and keeping retrieval (and ideally embedding) on the device means the sequence never leaves it.
Secondary benefits: Works offline Lower per-query cost (no server inference for retrieval) Lower/no network latency for the retrieval step Costs: phones have limited CPU/GPU/memory, the on-device LLM (if any) is weaker than a frontier model, models are harder to update than server code, and you must budget battery and storage.
A common hybrid: retrieval and embeddings on-device, final generation through a controlled LLM endpoint that only ever sees retrieved snippets + the query.