Dec 8, 2025

I just implemented vectorization over the most costly part of the address queries. This requires the extra cost of collecting the output links into a vector first, so it's less efficient for small queries, but it cuts the large query times in half. Considering that the link collection is sequential, as is the sort/de-dup/serialization, that's a pretty large reduction in the query cost. 5.2 million now in 15 seconds. With modern cores that will be under 10 seconds easy - for the largest query possible.