Trevor McCulloch

@rmuir @uschindler Would it be sufficient to add a factory function to `VectorizationProvider` for this like `GroupVIntUtil.Decoder createGroupVIntDecoder(MemorySegment segment)`? I anticipated that his might be necessary so I already have...

@benwtrent Segments before 9.9 would perform a graph search. In 9c3679bf14b0d75cd58e1fda3ca4b8a76aa033b9 I back ported your change from #12806 to 9.2, 9.4, and 9.5 since it wasn't too much effort.

@benwtrent I see a couple of possibilities to internalize this in `KnnVectorsReader`: * add a `public abstract void exhaustiveSearch(field, target, knnCollector, acceptDocs)` call that internalizes the exact search logic that...

A "shadow field" sounds like a view. It would be nice to be able to build an index on top of a view, but wouldn't this introduce dependencies between fields...

Lucene does not write/merge fields in a well defined order, and with a view you really need that field to be written/merged "last" so that you can view the contents...

If you used direct io for everything you would want to introduce an explicit disk cache somewhere, even with prefetching I don't think performance would meet expectations for a lot...

I wonder if it would be better to push this into `RandomVectorScorer.bulkScore`? The necessary prefetch calls would happen at roughly the same point the bulkScore call happens. The scorer could...

Yeah, virtual threads for `madvise` is not great -- I think at that point you'd prefer to copy the vector back onto the heap and let the virtual threads parallelize...

I am able to reproduce so taking a look

I also experience failure with SINGLE_BIT_QUERY_NIBBLE but the quantization is lossier so I'm focusing on PACKED_NIBBLE. This also reproduces on `main` so I've been working there. Agreed that the exactly...