Store vectors in: sqlite-vec — a SQLite extension with vector columns + KNN ObjectBox a hand-rolled flat index — for small N For a genomic report you index thousands of chunks, not millions of reads, so even a brute-force flat scan can be fast enough; ANN (HNSW) matters once N is large.
Persist the index in the app's local database so it survives restarts.
Memory budget: keep the working index in RAM if it fits (thousands of 384-dim int8 vectors is a few MB); otherwise memory-map it.