← notes

Experiment 042

IVF is the premise, not a component (scope entry)

Metadata: 042-ivf-cell-scope.json. No perf numbers — this is an architecture/scoping entry, written to close questions/IVF.md (nothing was built for IVF, so there's no existing entry to point to).

The question

Is IVF needed in this engine? If this node is already one IVF cell, what's the point of IVF here — isn't it not required?

The answer

IVF-the-router is not required here, and we deliberately don't build it — but IVF is the premise that justifies this entire engine. Two things to keep separate:

  1. IVF-the-router is out of scope (and not built). The coarse quantizer that decides which cells a query visits lives a layer above us. If you assume "this node is already an IVF cell," routing is somebody else's job; nothing in our scan needs it to exist at runtime.

  2. IVF is why the within-cell scan is worth optimizing at all. Without IVF, a query must scan all N vectors → the ~100 ms brute-force number from 001. IVF partitions N into ~√N cells and routes each query to only nprobe of them, which is what makes a "cell" small enough that an exact within-cell scan is fast enough to matter. Our fast scan is precisely the kernel that runs nprobe times per query because IVF narrowed the candidate set. So IVF is the reason our work exists — a premise, not a component we own.

In-scope seam questions (logged, not resolved)

These are NOT "building IVF" — they're "how the cell scan should behave given it lives inside an IVF." Future entries if we revisit:

  1. Cell sizing. Cell size sets our scan length (N per cell): too big → scan dominates; too small → recall loss + more cells for the router. The sweet spot is a within-cell concern that our scan throughput directly informs.
  2. Residual encoding. IVF cells often store residuals (vector − centroid), which quantize better. Does the rotation+binary funnel (026/029) gain recall on residuals vs raw vectors? A real quantization-on-the-cell question.
  3. Per-cell top-C / cross-cell merge. A query hits nprobe cells; each returns top-C and the router merges. What C per cell, and does multi-cell merging change the optimal funnel width?

Conclusion

questions/IVF.md is removed: the "is IVF required?" question is answered (no, the router isn't built here; IVF is the premise). The remaining IVF interaction questions are genuinely in-scope but unbuilt, and are parked above rather than in a separate question file.