• number of dimensions d  =  1 .. 500

• number of objects N  =  10⁴ .. 10¹⁰

• all values in query vector are specified (i.e. no partial-match)

• implementation runs on "standard" hardware/OS
  (e.g. don't assume arrays of parallel disks, huge memory, ...)

• independent of distance measure (?)

• 100% retrieval accuracy (?)

And, most importantly, the solution should ...

• use minimal resources:   i/o,   cpu-time,   space

• provide fast database updates (for some applications e.g. Magnifi)