use float32 in embeddings #133

sashabaranov · 2023-03-08T10:07:20Z

robrohan · 2024-07-18T02:43:21Z

embeddings.go

@@ -103,7 +103,7 @@ var stringToEnum = map[string]EmbeddingModel{
 // then their vector representations should also be similar.
 type Embedding struct {
 	Object    string    `json:"object"`
-	Embedding []float64 `json:"embedding"`
+	Embedding []float32 `json:"embedding"`


We are trying to track down a bug, and I think this change causes problems when you use Python to store embeddings and golang to try to do similarity scores against them. Python seems to be using float64 (or something with even more precision), and when we try to use those values with the embeddings from this library we get wildly different results then when we do the same code using Python only. I am not sure if the endpoint returns 64, but it would seem down casting to 32 bit make more sense then losing precision you can't get back.

use float32 in embeddings

4a87cf2

sashabaranov merged commit c46ebb2 into master Mar 8, 2023

sashabaranov deleted the float32-embeddings branch March 8, 2023 10:08

robrohan reviewed Jul 18, 2024

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

use float32 in embeddings #133

use float32 in embeddings #133

sashabaranov commented Mar 8, 2023 •

edited

Loading

robrohan Jul 18, 2024

use float32 in embeddings #133

use float32 in embeddings #133

Conversation

sashabaranov commented Mar 8, 2023 • edited Loading

robrohan Jul 18, 2024

Choose a reason for hiding this comment

sashabaranov commented Mar 8, 2023 •

edited

Loading