Querying very slow on large model (1,44gb)

[sajari]

Querying very slow on large model (1,44gb)

[mish15]

No worries, glad it's helpful.

In terms of speed, that's a larger model, on par to Google's news one. From memory the current has to scan all vectors, i think @dhowden made the calc about as fast as it's going to go (?), so your best bet is likely to:

a) make sure you're not collecting junk terms. Use a threshold for term addition. Also make sure punctuation, casing, IDs, phone numbers, etc are not causing lots of extra terms you don't care about.
b) reduce the dimensionality if you can and then
c) possibly shard the model into clusters and only query the clusters with close center points to the input vector.

Usually a) is the issue. b) is a handy compromise, c) i've never had to do.

[barthr]

Thanks for the fast response @mish15 I did an extra cleaning on the data and will try it out to see if the results are a better. It's pretty time consuming to run the model each time because there are more than 200 millions messages (tweets). Currently I'm only interested in the cluster where location indicators are, I think clustering the model will help me omit the things I'm not interested in.

Thanks for the help!