diff options
author | polwex <polwex@sortug.com> | 2025-09-23 03:50:53 +0700 |
---|---|---|
committer | polwex <polwex@sortug.com> | 2025-09-23 03:50:53 +0700 |
commit | 57aaafdb137fe49930711f6ed5ccc83b3a119cd2 (patch) | |
tree | 1a7556927bed94377630d33dd29c3bf07d159619 /NOTES.md |
init
Diffstat (limited to 'NOTES.md')
-rw-r--r-- | NOTES.md | 22 |
1 files changed, 22 insertions, 0 deletions
diff --git a/NOTES.md b/NOTES.md new file mode 100644 index 0000000..85f571b --- /dev/null +++ b/NOTES.md @@ -0,0 +1,22 @@ +We install some stuff on Nix (as much as we can) and the rest with `uv pip`. + +That works well. Which means it installs. But the `venv`, which is in `.venv`, also has a python runtime there, and as we're using the nix store python runtime, it won't see the venv packages. So you gotta do `source .venv/bin/activate` which is pretty weird but it does work. + + +21-9-2025- Long convo with ChatGPT about torch and numpy and tensors and ndarrays. idgi but interesting, should revisit. +Apparently transformers need the with torch.no_grad(): thing all the time but sentence transformers don't. at all + + + + +## DIM creating the DB table is retarded damn you chatgpt +>> he says +TL;DR + +DIM = number of floats per embedding vector (fixed by the model). + +sqlite-vec needs it at table creation. + +You don’t restart the DB every run — you just make sure new embeddings match the stored DIM. + +Best practice: store model+dim in a meta table so you don’t accidentally mix dimensions. |