The derived index is ~19 MB for 5,678 chunks, almost all of it float32 vectors.
int8 with a per-vector scale is standard practice among the low-storage retrievers surveyed: ~4x smaller file, faster matmul, negligible ranking loss at this scale. Add a residual-norm tiebreaker for near-equal cosines.
Backlog: BACKLOG.md → Open → Efficiency.
The derived index is ~19 MB for 5,678 chunks, almost all of it
float32vectors.int8 with a per-vector scale is standard practice among the low-storage retrievers surveyed: ~4x smaller file, faster matmul, negligible ranking loss at this scale. Add a residual-norm tiebreaker for near-equal cosines.
Backlog:
BACKLOG.md→ Open → Efficiency.