Missing ann-benchmarks documentation?

[loisaidasam]

The README says:

like Annoy, but with much higher recall

voyager/README.md

Line 17 in c6d09cc

Think of Voyager like [Sparkey](https://github.com/spotify/sparkey), but for vector/embedding data; or like [Annoy](https://github.com/spotify/annoy), but with [much higher recall](http://ann-benchmarks.com/). It got its name because it searches through (embedding) space(s), much like [the Voyager interstellar probes](https://en.wikipedia.org/wiki/Voyager_program) launched by NASA in 1977.

but I don't see any references to Voyager on the ann benchmarks page. Am I missing something?

[loretoparisi]

I think they were mentioning the fact that in respect to Annoy, Voyager is using hnsw lib as ANN, hence the base benchmark shows that Voyager has a higher Recall / Queries per seconds than Annoy. You can clearly see this from here.

[loisaidasam]

Which "base benchmark" are you referring to?

Perhaps faiss's hnsw implementation? or perhaps hnswlib? Those don't seem like fair comparisons, since Voyager is an entirely separate implementation.

Also, I think "clearly" is being generous. The way that the README links to ann-benchmarks suggests that one would see "Voyager" on these charts.

[loretoparisi]

hnswlib

I think a reasonable reference (Base line) could be the bare hnswlib. That said, according to the announcement, Voyager has a customized hnswllib version (as you can see from the sources), so the tests should be done from scratch to be more accurate.

[ijanderso]

Closing as it has been added to ann-benchmarks: erikbern/ann-benchmarks#473