Lesson 5.3

The novelty debate

TurboQuant and RaBitQ share a striking amount of machinery. Whether one is genuinely new or a restatement of the other is an open, public argument, and this project answers it the only honest way: by measuring, not by taking sides.

concepts research 8 min read

Two methods, suspiciously alike

Lay the two methods side by side and the resemblance is hard to miss. Both rotate the vector first. Both then quantize coordinates very coarsely, RaBitQ to one bit, TurboQuant with a per-coordinate quantizer plus a 1-bit correction. Both lean on the same high-dimensional concentration to make a crude code accurate. When RaBitQ came first and TurboQuant arrived with a louder splash, its first author, Jianyang Gao, publicly questioned whether TurboQuant is a real advance or largely a restatement of his earlier work. As of now the question is unresolved.

Compare the curves (shapes illustrative)

recall compression →

PQ at ratio 16×  ·  recall@10 0.91

The curves above are illustrative, drawn to show how a comparison reads, not to declare a winner. The real shapes come only from running both methods on the same data, the same way.

A benchmark beats an argument

Arguments about novelty are settled by claims; questions about which method is better for search are settled by measurement. That is exactly where this project's v2 goes: implement TurboQuant, implement Extended RaBitQ, run both against the v1 product-quantization baseline on the same datasets, the same recall and compression axes, and read the curves. Where one wins, say so and show it. Where they cross, say that too. A reproducible plot is worth more than a strong opinion.

That is also the honest end of this course. You have built the baseline and the ruler by hand; the frontier is now a fair fight you are equipped to referee.

Key takeaways

1

TurboQuant and RaBitQ share rotation, very coarse quantization, and high-dimensional concentration.

2

Whether TurboQuant is genuinely novel is a live, public debate, not a settled fact.

3

The resolution this project trusts is a reproducible head-to-head benchmark, not an argument.