Component breakdown

Ingestion/parser — read the VCF/report, extract variants + annotations +
  report sections.
Chunker — record-aware chunks (one variant+annotation, one report section)
  with citation metadata (source, gene, position).
Embedder — on-device model (MiniLM/bge-small via ONNX/Core ML), quantized.
Index — sqlite-vec vector table + a lexical index for exact codes.
Retriever — hybrid retrieve → fuse → (optional) re-rank → top-k with citations.
Generator — LLM client with grounding prompt, citations,
  abstain-on-low-coverage, and the outbound guard.
Guardrails/audit — outbound payload assertion, audit log, medical-boundary
  disclaimer.
Eval/telemetry — offline eval set + on-device quality signals
  (privacy-preserving).