Supabase

/

gte-small

like 101

Follow

Supabase 65

For context, I want to use this model in two places and generate the same embeddings.
In my Supabase typescript edge function code I generate the embedding using these lines of code
const session = new Supabase.ai.Session("gte-small");
const output = await session.run(title, {
mean_pool: true,
normalize: true,
});

And in python I do the following

def average_pool(last_hidden_states: Tensor,
attention_mask: Tensor) -> Tensor:
last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]

tokenizer = AutoTokenizer.from_pretrained("Supabase/gte-small")
model = AutoModel.from_pretrained("Supabase/gte-small")

batch_dict = tokenizer(titles, max_length=512, padding=True, truncation=False, return_tensors='pt')
batch_dict = {k: v.to(device) for k, v in batch_dict.items()}
#generate embeddings
outputs = model(**batch_dict)
embeddings = average_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
embeddings = F.normalize(embeddings, p=2, dim=1)
#convert embeddings to list
embeddings_list = embeddings.detach().cpu().numpy().tolist()

Based on my understanding, these two method should generate the same embeddings but when I printed out the embeddings they are slightly different
for some examples I've tested. The cosine similarity of the embedding generated by the edge function and the one by the hugging face model in python turned out to be higher than 0.999 so they are very similar.

I was wondering what might cause this? Is there something significantly wrong with my code? And is the model used by supabase the same version as the most recent model here?