Conversation
eddyxu
force-pushed
the
lei/search_diskann
branch
3 times, most recently
from
7fb7e6b to
c7bcfc3
Compare
eddyxu
force-pushed
the
lei/search_diskann
branch
from
c7bcfc3 to
4bd25dd
Compare
| // L parameter | ||
| uint32 L = 5; | ||
|
|
||
| /// Entry points to the graph |
There was a problem hiding this comment.
these are the vertex id's?
There was a problem hiding this comment.
These are the IDs in the graph. Basically the "index" in the graph file. Not the "row id" in original dataset
| if index_type != "IVF_PQ": | ||
| if index_type not in ["IVF_PQ", "DISKANN"]: | ||
| raise NotImplementedError( | ||
| f"Only IVF_PQ index_type supported. Got {index_type}" |
There was a problem hiding this comment.
change the message to include DISKANN?
| ivf_params.num_partitions = PyAny::downcast::<PyInt>(n)?.extract()? | ||
| }; | ||
|
|
||
| if let Some(n) = kwargs.get_item("num_bits") { |
There was a problem hiding this comment.
i forgot do we actually support configuring this now?
There was a problem hiding this comment.
We dont support "throw error for unsupported params". This is just do the handing silently. I am fine to remove it if you feel strongly.
| } | ||
| "DISKANN" => { | ||
| let mut params = DiskANNParams::default(); | ||
| if let Some(kwargs) = kwargs { |
There was a problem hiding this comment.
what're the default values for these? prolly worth tuning it and set reasonable defaults for like high embedding dimensions and like up to 1M vectors ?
There was a problem hiding this comment.
The default values were from the diskann paper. Will adjust them if necessary when we benchmark diskann
2 participants
No description provided.