Scaled down and converted images before sending them to AI models (#1101)

app/MindWork AI Studio/Provider/Anthropic/ProviderAnthropic.cs

@@ -59,18 +59,16 @@ public sealed class ProviderAnthropic() : BaseProvider(LLMProviders.ANTHROPIC, n
                 Text = text,
             },
             
-            // Anthropic-specific image sub-content:
-            async attachment => new SubContentImage
+            // Anthropic-specific image sub-content, which takes the base64 and its type apart:
+            image => new SubContentImage
             {
                 Source = new SubContentBase64Image
                 {
-                    Data = await attachment.TryAsBase64(token: token) is (true, var base64Content)
-                        ? base64Content
-                        : string.Empty,
-                    
-                    MediaType = attachment.DetermineMimeType(),
+                    Data = image.Base64,
+                    MediaType = image.MimeType,
                 }
-            }
+            },
+            token
         );
 
         //

app/MindWork AI Studio/Provider/OpenAI/ProviderOpenAI.cs

@@ -113,15 +113,14 @@ public sealed class ProviderOpenAI() : BaseProvider(LLMProviders.OPEN_AI, new Ur
                                        {
                                            Text = text,
                                        },
-                                       async attachment => new SubContentImageUrlNested
+                                       image => new SubContentImageUrlNested
                                        {
                                            ImageUrl = new SubContentImageUrlData
                                            {
-                                               Url = await attachment.TryAsBase64(token: token) is (true, var base64Content)
-                                                   ? $"data:{attachment.DetermineMimeType()};base64,{base64Content}"
-                                                   : string.Empty,
+                                               Url = image.ToDataUrl(),
                                            },
-                                       });
+                                       },
+                                       token);
 
                                    return new ChatCompletionAPIRequest
                                    {
@@ -192,12 +191,11 @@ public sealed class ProviderOpenAI() : BaseProvider(LLMProviders.OPEN_AI, new Ur
             {
                 Text = text,
             },
-            async attachment => new SubContentInputImage
+            image => new SubContentInputImage
             {
-                ImageUrl = await attachment.TryAsBase64(token: token) is (true, var base64Content)
-                    ? $"data:{attachment.DetermineMimeType()};base64,{base64Content}"
-                    : string.Empty,
-            });
+                ImageUrl = image.ToDataUrl(),
+            },
+            token);
 
         var baseInput = new List<object> { systemPrompt };
         baseInput.AddRange(messages);

app/MindWork AI Studio/Provider/SelfHosted/HostExtensions.cs

@@ -36,6 +36,29 @@ public static class HostExtensions
         _ => "embeddings",
     };
     
+    /// <summary>
+    /// The largest image file, in bytes, which the engine accepts in a chat request, where we know of one.
+    /// </summary>
+    /// <remarks>
+    /// Like the limit of a cloud provider, see LLMProvidersExtensions.GetMaxImageSizeBytes, this
+    /// belongs to the API of the engine, not to the model. What we actually send is asked through
+    /// IImageSourceExtensions.MaxImageSizeBytesFor.
+    /// </remarks>
+    /// <param name="host">The engine.</param>
+    /// <returns>The size limit of one image file, or null when the engine states none we know of.</returns>
+    public static long? GetMaxImageSizeBytes(this Host host) => host switch
+    {
+        //
+        // llmman answers a chat request larger than 2 MiB with 413: its route /v1/chat/completions
+        // keeps the default limit of its web framework, read on 2026-10-22 at
+        // https://github.com/llmmanorg/llmman/blob/main/src/cmd/serve.rs. An image of 1 MiB takes
+        // about 1.33 MiB as base64, which leaves room for the text of the conversation:
+        //
+        Host.LLMMAN => 1_048_576,
+
+        _ => null,
+    };
+
     public static bool IsChatSupported(this Host host)
     {
         switch (host)