Recursive misi construction

Construct and evaluate a recursive misi hierarchy in which the inner index over the sampled vocabulary is itself built using the misi method, particularly at sample sizes of at least 10^8 where conventional inner-index construction may exceed available memory.

Background

Misi is a combinator: it builds an index over the full database by using an approximate-nearest-neighbor index over a linear-size sample. At very large database scales, the sample and its inner index may themselves become too large for RAM. Recursive construction is proposed as the design’s natural response, but its feasibility depends on the recall and scalability of recursively built inner indexes and is not measured.

References

Everything unmeasured is labeled future work, including the two compositions the analysis motivates most strongly: a resident product-quantization screen before exact verification, and recursive construction (the inner index built by the same method).

— misi: a Metric Inverted Sample Index  (2608.27422 - Chavez, 27 Aug 2026) in Introduction, final paragraph; Section 6, “Limitations and future work”