Hierarchical Navigable Small Worlds
zilliz.com
Hierarchical Navigable Small Worlds
1–10 of 10 posts
Re: Hierarchical Navigable Small Worlds
#2Kinda wish they would use some more letters when writing the sample code. This terse style in general is fine but when the purpose of the code is to try to explain how something works, it's adding an extra layer of difficulty decoding what the heck 'nns, 'cv', 'd', 'e' and 'ef' is supposed to mean.
Like this ChatGPT-looking function
nns = [best]
visit = set(best) # set of visited nodes
candid = [best] # candidate nodes to insert into nearest neighbors
# find top-k nearest neighbors
while candid:
cv = heappop(candid)
if nns[-1][0] ef:
nns.pop()
nearest_neighbors = [best]
visited = set(best)
candidates = [best]
# find top-k nearest neighbors
while candidates:
candidate_vector = heappop(candidates)
if nearest_neighbors[-1][0] epsilon:
nearest_neighbors.pop()Re: Hierarchical Navigable Small Worlds
#3Kinda wish they would use some more letters when writing the sample code. This terse style in general is fine but when the purpose of the code is to try to explain how something works, it's adding an extra layer of difficulty decoding what the heck 'nns, 'cv', 'd', 'e' and 'ef' is supposed to mean. Like this ChatGPT-looking function nns = [best] visit = set(best) # set of visited nodes candid = [best] # candidate nod…
Good suggestion!
Actually the "ef" is not epsilon. It is a parameter of the HNSW index: https://github.com/nmslib/hnswlib/blob/master/ALGO_PARAMS.md...
Re: Hierarchical Navigable Small Worlds
#4Kinda wish they would use some more letters when writing the sample code. This terse style in general is fine but when the purpose of the code is to try to explain how something works, it's adding an extra layer of difficulty decoding what the heck 'nns, 'cv', 'd', 'e' and 'ef' is supposed to mean. Like this ChatGPT-looking function nns = [best] visit = set(best) # set of visited nodes candid = [best] # candidate nod…
Good suggestion! Actually the "ef" is not epsilon. It is a parameter of the HNSW index: https://github.com/nmslib/hnswlib/blob/master/ALGO_PARAMS.md...
Yeah it wasn't referenced or mentioned anywhere in the article so I had to guess what it was :-/ ... which I guess supports my point.
ef being "epsilon factor" was the least nonsensical thing I could come up with since it looked analogous to how you do like abs(a-b)<epsilon in floating point comparisons.
Re: Hierarchical Navigable Small Worlds
#5Earlier quoted context omitted.
Good suggestion! Actually the "ef" is not epsilon. It is a parameter of the HNSW index: https://github.com/nmslib/hnswlib/blob/master/ALGO_PARAMS.md...
Yeah it wasn't referenced or mentioned anywhere in the article so I had to guess what it was :-/ ... which I guess supports my point. ef being "epsilon factor" was the least nonsensical thing I could come up with since it looked analogous to how you do like abs(a-b)<epsilon in floating point comparisons.
Hm... adding links in code comment might be better.
Re: Hierarchical Navigable Small Worlds
#6Earlier quoted context omitted.
Yeah it wasn't referenced or mentioned anywhere in the article so I had to guess what it was :-/ ... which I guess supports my point. ef being "epsilon factor" was the least nonsensical thing I could come up with since it looked analogous to how you do like abs(a-b)<epsilon in floating point comparisons.
Hm... adding links in code comment might be better.
Probably a comment above the function explaining the incoming parameters, with a link to the specs if necessary, and then comments in the text and better variable names.
I think in this scenario when the primary audience of the code isn't an interpreter, but people who don't understand HNSWs, and who will probably only glance over the code once or twice, you really can't lay on the verbosity heavily enough.
Re: Hierarchical Navigable Small Worlds
#7https://github.com/instant-labs/instant-distance is a compact, fairly readable, pretty fast implementation of the paper in Rust (that I built).
We use it for searching aftermarket domains on Instant Domain Search.
Re: Hierarchical Navigable Small Worlds
#8Why have a cookie dialog if my only option is to "accept"? Not much of a dialog there.
Re: Hierarchical Navigable Small Worlds
#9Why have a cookie dialog if my only option is to "accept"? Not much of a dialog there.
probably required by some laws
Re: Hierarchical Navigable Small Worlds
#10Why have a cookie dialog if my only option is to "accept"? Not much of a dialog there.
probably required by some laws
Actually that’s a common misunderstanding of the ePrivacy directive (often bundled with the GDPR(, if you’re only providing an accept dialog you might as well save the effort and not show a cookie notice at all.
You need to obtain consent before you set any marketing cookies and effective consent can’t be obtained if your only option is “yeah track me”