Hi there,
I have no idea when this bug sneaked up, since I haven't really updated my version for 6 months, since you originally implemented the vector_full_scan_stream function. However, it's ordering things completely bananas now. You can see in the screenshot an example. This is the result of executing the following SQL;
select vss.distance, ts.id, ts.created, ts.type, ts.pushed, ts.uri, ts.prompt, ts.completion, ts.filename, ts.cached, ts.meta, ts.embeddings as embedding_vss
from ml_training_snippets ts
inner join vector_full_scan_stream('ml_training_snippets', 'embeddings', vector_as_f32(@embedding)) as vss on vss.rowid = ts.id
where 1 and type = @type order by vss.distance limit 10
I suspect if you start from the point of the implementation of the vector_full_scan_stream function, and track forward, you'll find something that broke the vector_full_scan_stream function ...
The only thing I did was to update to the latest binaries ...

Hi there,
I have no idea when this bug sneaked up, since I haven't really updated my version for 6 months, since you originally implemented the
vector_full_scan_streamfunction. However, it's ordering things completely bananas now. You can see in the screenshot an example. This is the result of executing the following SQL;I suspect if you start from the point of the implementation of the
vector_full_scan_streamfunction, and track forward, you'll find something that broke thevector_full_scan_streamfunction ...The only thing I did was to update to the latest binaries ...