language_model: Fix `ImageSize` storing pre-downscale dimensions (#54357)

Daniel Strobusch created 1 week ago

When images are resized to meet provider size constraints (Anthropic's
1568px limit or the 5MB encoded-PNG cap), the stored ImageSize was still
recording the original width/height rather than the final post-downscale
dimensions. This caused incorrect token estimation via estimate_tokens()
since it uses width * height / 750.

Use processed_image.dimensions() after all downscale passes so that
ImageSize reflects the actual image sent to the provider.

Release Notes:

- Fixed an issue where token estimation would be incorrect in case where
the thread contained downscaled images.

Change summary

crates/language_model/src/request.rs | 16 ++++++++++++++--
1 file changed, 14 insertions(+), 2 deletions(-)

Detailed changes

crates/language_model/src/request.rs

@@ -118,10 +118,12 @@ impl LanguageModelImageExt for LanguageModelImage {
             // SAFETY: The base64 encoder should not produce non-UTF8.
             let source = unsafe { String::from_utf8_unchecked(base64_image) };
 
+            let (final_width, final_height) = processed_image.dimensions();
+
             Some(LanguageModelImage {
                 size: Some(ImageSize {
-                    width: width as i32,
-                    height: height as i32,
+                    width: final_width as i32,
+                    height: final_height as i32,
                 }),
                 source: source.into(),
             })
@@ -227,5 +229,15 @@ mod tests {
             w,
             h
         );
+
+        let size = lm_image.size.expect("ImageSize should be present");
+        assert_eq!(
+            size.width, w as i32,
+            "ImageSize.width should match the encoded PNG width after downscaling"
+        );
+        assert_eq!(
+            size.height, h as i32,
+            "ImageSize.height should match the encoded PNG height after downscaling"
+        );
     }
 }