pgvector & pg-storm integration

[CC-Hsu]

Hi, Team,

It is now really popular on LLM and embedding vector feature support for databases, especially the post popular pgvector extensoin for PGSQL.

However, this extension is not going to be GPU-accelerated.

I think it would be great if PG-Storm can support pgvector integration and acceleration (maybe walk together).

Specifically, I think PG-Storm can toward in-depth GPU-accelerated support on

1. vector Datatype support
2. Function support, such as distance functions, comparison operators
3. Index support (if possible), including searching index, (re)building index

I feel it would be popular if such feature is available.

Best Regards,

[kaigai]

I don't know whether "pgvector integration" is a suitable direction, or not.
However, I have worked on non-index based vector search with PG-Strom, and probably it makes sense for the recent demands.
We already list up items to be developed.

• [idea] ベクトル検索機能 #698 vector search functionality (GPU version of distance function based on float2/4/8[] array, or harhaps vector type)
• [idea] GpuSort + Limit #738 GpuSort + Limit (A feature to pick up the nearest N items from the query results)
• [idea] Large-tables GpuJoin #752 Large-tables GpuJoin -> it also allows to preserve GpuScan results on GPU memory to reuse other task's input like GpuJoin and potentialy GpuSort.

I plan to implement these features by end of this summer.
If you have real-life vector search workloads to test these workloads, it is welcome for us.