TypeSafe Jev .NET SDK 0.7.0: Experimental Images and Bounding Boxes
The community TypeSafe Jev .NET SDK 0.7.0 adds experimental image input and a Box question type that returns a normalized bounding box for a target in an image. The release also brings nullable usage counts, stricter request validation and zstd compression. Developers experimenting with vision requests in C# should read on.
Key points
- Image input requires the TypeSafe-Danger-Zone-Images: true header, so enable it only on clients that need vision.
- For image references inside text, use TextWithAssets.Create rather than plain string interpolation so the SDK registers the asset.
- Box edges are normalized from zero to one; multiply vertical values by image height and horizontal values by width to get pixels.
- BoxAnswer.Confidence and Usage token counts are nullable; treat null as not provided, not zero.
- Box can return a rectangle even when the target is absent, so pair it with a separate Noul presence check.
- Only PNG and JPEG are accepted; bodies of 1,024 bytes or more are zstd-compressed by default unless Compression = false is set.
TypeSafe's JavaScript SDK v0.7.0 introduces experimental image input and a Box question type, and the community .NET port has received a matching 0.7.0 update. The version numbers align, but these are separate releases, and the .NET package is not officially supported by TypeSafe. Image input stays off until the client sends the TypeSafe-Danger-Zone-Images header.
A Box question returns normalized Top, Bottom, Left and Right edges between zero and one, which are fractions of image height and width rather than pixels. Confidence is nullable, so a missing value means not reported, not zero. Box can return a rectangle even when the target is absent, so pair it with a separate Noul presence check. Image requests go out as multipart form data, and the port compresses bodies of at least 1,024 bytes with zstd on .NET 11.
Real-world use
- A support ticket pipeline that locates damaged areas in submitted product photos and routes them to human review.
- A content moderation workflow that checks an uploaded image for a specific object using a Noul presence question alongside a Box location question.
- A prototype ASP.NET Core backend that experiments with vision requests and returns typed answers to the caller.
Why it matters
Vision calls with Jev can now be written in C# using Image and TextWithAssets, but the feature is experimental and Box results should be treated as location suggestions rather than proof of presence. Test these limits before relying on them in production.