fix: don't send default parameters; Params may vary by model

Madzionator · Madzionator · commit ac54712e0611 · 2026-03-18T11:24:05.000+01:00
diff --git a/MaIN.Core.IntegrationTests/BackendParamsTests.cs b/MaIN.Core.IntegrationTests/BackendParamsTests.cs
diff --git a/src/MaIN.Domain/Configuration/BackendInferenceParams/AnthropicInferenceParams.cs b/src/MaIN.Domain/Configuration/BackendInferenceParams/AnthropicInferenceParams.cs
@@ -8,9 +8,9 @@ public class AnthropicInferenceParams : IBackendInferenceParams
 {
     public BackendType Backend => BackendType.Anthropic;
 
-    public float Temperature { get; init; } = 1.0f;
-    public int MaxTokens { get; init; } = 4096;
-    public int TopK { get; init; } = -1;
-    public float TopP { get; init; } = 1.0f;
+    public float? Temperature { get; init; }
+    public int? MaxTokens { get; init; }
+    public int? TopK { get; init; }
+    public float? TopP { get; init; }
     public Grammar? Grammar { get; set; }
 }
diff --git a/src/MaIN.Domain/Configuration/BackendInferenceParams/DeepSeekInferenceParams.cs b/src/MaIN.Domain/Configuration/BackendInferenceParams/DeepSeekInferenceParams.cs
@@ -8,11 +8,11 @@ public class DeepSeekInferenceParams : IBackendInferenceParams
 {
     public BackendType Backend => BackendType.DeepSeek;
 
-    public float Temperature { get; init; } = 0.7f;
-    public int MaxTokens { get; init; } = 4096;
-    public float TopP { get; init; } = 1.0f;
-    public float FrequencyPenalty { get; init; }
-    public float PresencePenalty { get; init; }
+    public float? Temperature { get; init; }
+    public int? MaxTokens { get; init; }
+    public float? TopP { get; init; }
+    public float? FrequencyPenalty { get; init; }
+    public float? PresencePenalty { get; init; }
     public string? ResponseFormat { get; init; }
     public Grammar? Grammar { get; set; }
 }
diff --git a/src/MaIN.Domain/Configuration/BackendInferenceParams/GeminiInferenceParams.cs b/src/MaIN.Domain/Configuration/BackendInferenceParams/GeminiInferenceParams.cs
@@ -8,9 +8,9 @@ public class GeminiInferenceParams : IBackendInferenceParams
 {
     public BackendType Backend => BackendType.Gemini;
 
-    public float Temperature { get; init; } = 0.7f;
-    public int MaxTokens { get; init; } = 4096;
-    public float TopP { get; init; } = 0.95f;
+    public float? Temperature { get; init; }
+    public int? MaxTokens { get; init; }
+    public float? TopP { get; init; }
     public string[]? StopSequences { get; init; }
     public Grammar? Grammar { get; set; }
 }
diff --git a/src/MaIN.Domain/Configuration/BackendInferenceParams/GroqCloudInferenceParams.cs b/src/MaIN.Domain/Configuration/BackendInferenceParams/GroqCloudInferenceParams.cs
@@ -8,10 +8,10 @@ public class GroqCloudInferenceParams : IBackendInferenceParams
 {
     public BackendType Backend => BackendType.GroqCloud;
 
-    public float Temperature { get; init; } = 0.7f;
-    public int MaxTokens { get; init; } = 4096;
-    public float TopP { get; init; } = 1.0f;
-    public float FrequencyPenalty { get; init; }
+    public float? Temperature { get; init; }
+    public int? MaxTokens { get; init; }
+    public float? TopP { get; init; }
+    public float? FrequencyPenalty { get; init; }
     public string? ResponseFormat { get; init; }
     public Grammar? Grammar { get; set; }
 }
diff --git a/src/MaIN.Domain/Configuration/BackendInferenceParams/OllamaInferenceParams.cs b/src/MaIN.Domain/Configuration/BackendInferenceParams/OllamaInferenceParams.cs
@@ -8,11 +8,11 @@ public class OllamaInferenceParams : IBackendInferenceParams
 {
     public BackendType Backend => BackendType.Ollama;
 
-    public float Temperature { get; init; } = 0.8f;
-    public int MaxTokens { get; init; } = 4096;
-    public int TopK { get; init; } = 40;
-    public float TopP { get; init; } = 0.9f;
-    public int NumCtx { get; init; } = 2048;
-    public int NumGpu { get; init; } = 30;
+    public float? Temperature { get; init; }
+    public int? MaxTokens { get; init; }
+    public int? TopK { get; init; }
+    public float? TopP { get; init; }
+    public int? NumCtx { get; init; }
+    public int? NumGpu { get; init; }
     public Grammar? Grammar { get; set; }
 }
diff --git a/src/MaIN.Domain/Configuration/BackendInferenceParams/OpenAiInferenceParams.cs b/src/MaIN.Domain/Configuration/BackendInferenceParams/OpenAiInferenceParams.cs
@@ -8,11 +8,11 @@ public class OpenAiInferenceParams : IBackendInferenceParams
 {
     public BackendType Backend => BackendType.OpenAi;
 
-    public float Temperature { get; init; } = 0.7f;
-    public int MaxTokens { get; init; } = 4096;
-    public float TopP { get; init; } = 1.0f;
-    public float FrequencyPenalty { get; init; }
-    public float PresencePenalty { get; init; }
+    public float? Temperature { get; init; }
+    public int? MaxTokens { get; init; }
+    public float? TopP { get; init; }
+    public float? FrequencyPenalty { get; init; }
+    public float? PresencePenalty { get; init; }
     public string? ResponseFormat { get; init; }
     public Grammar? Grammar { get; set; }
 }
diff --git a/src/MaIN.Domain/Configuration/BackendInferenceParams/XaiInferenceParams.cs b/src/MaIN.Domain/Configuration/BackendInferenceParams/XaiInferenceParams.cs
@@ -8,10 +8,10 @@ public class XaiInferenceParams : IBackendInferenceParams
 {
     public BackendType Backend => BackendType.Xai;
 
-    public float Temperature { get; init; } = 0.7f;
-    public int MaxTokens { get; init; } = 4096;
-    public float TopP { get; init; } = 1.0f;
-    public float FrequencyPenalty { get; init; }
-    public float PresencePenalty { get; init; }
+    public float? Temperature { get; init; }
+    public int? MaxTokens { get; init; }
+    public float? TopP { get; init; }
+    public float? FrequencyPenalty { get; init; }
+    public float? PresencePenalty { get; init; }
     public Grammar? Grammar { get; set; }
 }
diff --git a/src/MaIN.Domain/Entities/Chat.cs b/src/MaIN.Domain/Entities/Chat.cs
@@ -40,7 +40,7 @@ public AIModel? ModelInstance
     public List<Message> Messages { get; set; } = [];
     public ChatType Type { get; set; } = ChatType.Conversation;
     public bool ImageGen { get; set; }
-    public IBackendInferenceParams BackendParams { get; set; } = new LocalInferenceParams();
+    public IBackendInferenceParams? BackendParams { get; set; }
     public LocalInferenceParams? LocalParams => BackendParams as LocalInferenceParams;
 
     public Grammar? InferenceGrammar
diff --git a/src/MaIN.Services/Services/ChatService.cs b/src/MaIN.Services/Services/ChatService.cs
@@ -1,4 +1,5 @@
 using MaIN.Domain.Configuration;
+using MaIN.Domain.Configuration.BackendInferenceParams;
 using MaIN.Domain.Entities;
 using MaIN.Domain.Exceptions.Chats;
 using MaIN.Domain.Models;
@@ -38,6 +39,7 @@ public async Task<ChatResult> Completions(
             chat.ImageGen = true;
         }
         chat.Backend ??= settings.BackendType;
+        chat.BackendParams ??= BackendParamsFactory.Create(chat.Backend.Value);
 
         chat.Messages.Where(x => x.Type == MessageType.NotSet).ToList()
             .ForEach(x => x.Type = chat.Backend != BackendType.Self ? MessageType.CloudLLM : MessageType.LocalLLM);
diff --git a/src/MaIN.Services/Services/LLMService/AnthropicService.cs b/src/MaIN.Services/Services/LLMService/AnthropicService.cs
@@ -524,9 +524,9 @@ private object BuildAnthropicRequestBody(Chat chat, List<ChatMessage> conversati
 
         if (anthParams != null)
         {
-            requestBody["temperature"] = anthParams.Temperature;
-            if (anthParams.TopP < 1.0f) requestBody["top_p"] = anthParams.TopP;
-            if (anthParams.TopK > 0) requestBody["top_k"] = anthParams.TopK;
+            if (anthParams.Temperature.HasValue) requestBody["temperature"] = anthParams.Temperature.Value;
+            if (anthParams.TopP.HasValue) requestBody["top_p"] = anthParams.TopP.Value;
+            if (anthParams.TopK.HasValue) requestBody["top_k"] = anthParams.TopK.Value;
         }
 
         var systemMessage = conversation.FirstOrDefault(m =>
@@ -711,9 +711,9 @@ private async Task ProcessStreamingChatAsync(
         };
         if (anthParams2 != null)
         {
-            requestBody["temperature"] = anthParams2.Temperature;
-            if (anthParams2.TopP < 1.0f) requestBody["top_p"] = anthParams2.TopP;
-            if (anthParams2.TopK > 0) requestBody["top_k"] = anthParams2.TopK;
+            if (anthParams2.Temperature.HasValue) requestBody["temperature"] = anthParams2.Temperature.Value;
+            if (anthParams2.TopP.HasValue) requestBody["top_p"] = anthParams2.TopP.Value;
+            if (anthParams2.TopK.HasValue) requestBody["top_k"] = anthParams2.TopK.Value;
         }
 
         var requestJson = JsonSerializer.Serialize(requestBody);
@@ -806,9 +806,9 @@ private async Task ProcessNonStreamingChatAsync(
         };
         if (anthParams3 != null)
         {
-            requestBody["temperature"] = anthParams3.Temperature;
-            if (anthParams3.TopP < 1.0f) requestBody["top_p"] = anthParams3.TopP;
-            if (anthParams3.TopK > 0) requestBody["top_k"] = anthParams3.TopK;
+            if (anthParams3.Temperature.HasValue) requestBody["temperature"] = anthParams3.Temperature.Value;
+            if (anthParams3.TopP.HasValue) requestBody["top_p"] = anthParams3.TopP.Value;
+            if (anthParams3.TopK.HasValue) requestBody["top_k"] = anthParams3.TopK.Value;
         }
 
         var requestJson = JsonSerializer.Serialize(requestBody);
diff --git a/src/MaIN.Services/Services/LLMService/DeepSeekService.cs b/src/MaIN.Services/Services/LLMService/DeepSeekService.cs
@@ -54,11 +54,11 @@ protected override void ValidateApiKey()
     protected override void ApplyBackendParams(Dictionary<string, object> requestBody, Chat chat)
     {
         if (chat.BackendParams is not DeepSeekInferenceParams p) return;
-        requestBody["temperature"] = p.Temperature;
-        requestBody["max_tokens"] = p.MaxTokens;
-        requestBody["top_p"] = p.TopP;
-        if (p.FrequencyPenalty != 0) requestBody["frequency_penalty"] = p.FrequencyPenalty;
-        if (p.PresencePenalty != 0) requestBody["presence_penalty"] = p.PresencePenalty;
+        if (p.Temperature.HasValue) requestBody["temperature"] = p.Temperature.Value;
+        if (p.MaxTokens.HasValue) requestBody["max_tokens"] = p.MaxTokens.Value;
+        if (p.TopP.HasValue) requestBody["top_p"] = p.TopP.Value;
+        if (p.FrequencyPenalty.HasValue) requestBody["frequency_penalty"] = p.FrequencyPenalty.Value;
+        if (p.PresencePenalty.HasValue) requestBody["presence_penalty"] = p.PresencePenalty.Value;
         if (p.ResponseFormat != null) requestBody["response_format"] = new { type = p.ResponseFormat };
     }
 
diff --git a/src/MaIN.Services/Services/LLMService/GeminiService.cs b/src/MaIN.Services/Services/LLMService/GeminiService.cs
@@ -78,9 +78,9 @@ protected override void ValidateApiKey()
     protected override void ApplyBackendParams(Dictionary<string, object> requestBody, Chat chat)
     {
         if (chat.BackendParams is not GeminiInferenceParams p) return;
-        requestBody["temperature"] = p.Temperature;
-        requestBody["max_tokens"] = p.MaxTokens;
-        requestBody["top_p"] = p.TopP;
+        if (p.Temperature.HasValue) requestBody["temperature"] = p.Temperature.Value;
+        if (p.MaxTokens.HasValue) requestBody["max_tokens"] = p.MaxTokens.Value;
+        if (p.TopP.HasValue) requestBody["top_p"] = p.TopP.Value;
         if (p.StopSequences is { Length: > 0 }) requestBody["stop"] = p.StopSequences;
     }
 
diff --git a/src/MaIN.Services/Services/LLMService/GroqCloudService.cs b/src/MaIN.Services/Services/LLMService/GroqCloudService.cs
@@ -48,10 +48,10 @@ protected override void ValidateApiKey()
     protected override void ApplyBackendParams(Dictionary<string, object> requestBody, Chat chat)
     {
         if (chat.BackendParams is not GroqCloudInferenceParams p) return;
-        requestBody["temperature"] = p.Temperature;
-        requestBody["max_tokens"] = p.MaxTokens;
-        requestBody["top_p"] = p.TopP;
-        if (p.FrequencyPenalty != 0) requestBody["frequency_penalty"] = p.FrequencyPenalty;
+        if (p.Temperature.HasValue) requestBody["temperature"] = p.Temperature.Value;
+        if (p.MaxTokens.HasValue) requestBody["max_tokens"] = p.MaxTokens.Value;
+        if (p.TopP.HasValue) requestBody["top_p"] = p.TopP.Value;
+        if (p.FrequencyPenalty.HasValue) requestBody["frequency_penalty"] = p.FrequencyPenalty.Value;
         if (p.ResponseFormat != null) requestBody["response_format"] = new { type = p.ResponseFormat };
     }
 
diff --git a/src/MaIN.Services/Services/LLMService/OllamaService.cs b/src/MaIN.Services/Services/LLMService/OllamaService.cs
@@ -45,11 +45,17 @@ protected override void ValidateApiKey()
     protected override void ApplyBackendParams(Dictionary<string, object> requestBody, Chat chat)
     {
         if (chat.BackendParams is not OllamaInferenceParams p) return;
-        requestBody["temperature"] = p.Temperature;
-        requestBody["max_tokens"] = p.MaxTokens;
-        requestBody["top_p"] = p.TopP;
-        if (p.TopK > 0) requestBody["top_k"] = p.TopK;
-        if (p.NumCtx > 0) requestBody["options"] = new { num_ctx = p.NumCtx, num_gpu = p.NumGpu };
+        if (p.Temperature.HasValue) requestBody["temperature"] = p.Temperature.Value;
+        if (p.MaxTokens.HasValue) requestBody["max_tokens"] = p.MaxTokens.Value;
+        if (p.TopP.HasValue) requestBody["top_p"] = p.TopP.Value;
+        if (p.TopK.HasValue) requestBody["top_k"] = p.TopK.Value;
+        if (p.NumCtx.HasValue || p.NumGpu.HasValue)
+        {
+            var options = new Dictionary<string, object>();
+            if (p.NumCtx.HasValue) options["num_ctx"] = p.NumCtx.Value;
+            if (p.NumGpu.HasValue) options["num_gpu"] = p.NumGpu.Value;
+            requestBody["options"] = options;
+        }
     }
 
     public override async Task<ChatResult?> AskMemory(
diff --git a/src/MaIN.Services/Services/LLMService/OpenAiService.cs b/src/MaIN.Services/Services/LLMService/OpenAiService.cs
@@ -41,11 +41,11 @@ protected override void ValidateApiKey()
     protected override void ApplyBackendParams(Dictionary<string, object> requestBody, Chat chat)
     {
         if (chat.BackendParams is not OpenAiInferenceParams p) return;
-        requestBody["temperature"] = p.Temperature;
-        requestBody["max_tokens"] = p.MaxTokens;
-        requestBody["top_p"] = p.TopP;
-        if (p.FrequencyPenalty != 0) requestBody["frequency_penalty"] = p.FrequencyPenalty;
-        if (p.PresencePenalty != 0) requestBody["presence_penalty"] = p.PresencePenalty;
+        if (p.MaxTokens.HasValue) requestBody["max_completion_tokens"] = p.MaxTokens.Value;
+        if (p.Temperature.HasValue) requestBody["temperature"] = p.Temperature.Value;
+        if (p.TopP.HasValue) requestBody["top_p"] = p.TopP.Value;
+        if (p.FrequencyPenalty.HasValue) requestBody["frequency_penalty"] = p.FrequencyPenalty.Value;
+        if (p.PresencePenalty.HasValue) requestBody["presence_penalty"] = p.PresencePenalty.Value;
         if (p.ResponseFormat != null) requestBody["response_format"] = new { type = p.ResponseFormat };
     }
 
diff --git a/src/MaIN.Services/Services/LLMService/XaiService.cs b/src/MaIN.Services/Services/LLMService/XaiService.cs
@@ -47,11 +47,11 @@ protected override void ValidateApiKey()
     protected override void ApplyBackendParams(Dictionary<string, object> requestBody, Chat chat)
     {
         if (chat.BackendParams is not XaiInferenceParams p) return;
-        requestBody["temperature"] = p.Temperature;
-        requestBody["max_tokens"] = p.MaxTokens;
-        requestBody["top_p"] = p.TopP;
-        if (p.FrequencyPenalty != 0) requestBody["frequency_penalty"] = p.FrequencyPenalty;
-        if (p.PresencePenalty != 0) requestBody["presence_penalty"] = p.PresencePenalty;
+        if (p.Temperature.HasValue) requestBody["temperature"] = p.Temperature.Value;
+        if (p.MaxTokens.HasValue) requestBody["max_tokens"] = p.MaxTokens.Value;
+        if (p.TopP.HasValue) requestBody["top_p"] = p.TopP.Value;
+        if (p.FrequencyPenalty.HasValue) requestBody["frequency_penalty"] = p.FrequencyPenalty.Value;
+        if (p.PresencePenalty.HasValue) requestBody["presence_penalty"] = p.PresencePenalty.Value;
     }
 
     public override async Task<ChatResult?> AskMemory(

Original file line number	Diff line number	Diff line change
`@@ -8,9 +8,9 @@ public class AnthropicInferenceParams : IBackendInferenceParams`
`8`	`8`	`{`
`9`	`9`	`public BackendType Backend => BackendType.Anthropic;`
`10`	`10`
`11`		`- public float Temperature { get; init; } = 1.0f;`
`12`		`- public int MaxTokens { get; init; } = 4096;`
`13`		`- public int TopK { get; init; } = -1;`
`14`		`- public float TopP { get; init; } = 1.0f;`
	`11`	`+ public float? Temperature { get; init; }`
	`12`	`+ public int? MaxTokens { get; init; }`
	`13`	`+ public int? TopK { get; init; }`
	`14`	`+ public float? TopP { get; init; }`
`15`	`15`	`public Grammar? Grammar { get; set; }`
`16`	`16`	`}`
Original file line number	Diff line number	Diff line change
`@@ -8,9 +8,9 @@ public class GeminiInferenceParams : IBackendInferenceParams`
`8`	`8`	`{`
`9`	`9`	`public BackendType Backend => BackendType.Gemini;`
`10`	`10`
`11`		`- public float Temperature { get; init; } = 0.7f;`
`12`		`- public int MaxTokens { get; init; } = 4096;`
`13`		`- public float TopP { get; init; } = 0.95f;`
	`11`	`+ public float? Temperature { get; init; }`
	`12`	`+ public int? MaxTokens { get; init; }`
	`13`	`+ public float? TopP { get; init; }`
`14`	`14`	`public string[]? StopSequences { get; init; }`
`15`	`15`	`public Grammar? Grammar { get; set; }`
`16`	`16`	`}`
Original file line number	Diff line number	Diff line change
`@@ -8,10 +8,10 @@ public class GroqCloudInferenceParams : IBackendInferenceParams`
`8`	`8`	`{`
`9`	`9`	`public BackendType Backend => BackendType.GroqCloud;`
`10`	`10`
`11`		`- public float Temperature { get; init; } = 0.7f;`
`12`		`- public int MaxTokens { get; init; } = 4096;`
`13`		`- public float TopP { get; init; } = 1.0f;`
`14`		`- public float FrequencyPenalty { get; init; }`
	`11`	`+ public float? Temperature { get; init; }`
	`12`	`+ public int? MaxTokens { get; init; }`
	`13`	`+ public float? TopP { get; init; }`
	`14`	`+ public float? FrequencyPenalty { get; init; }`
`15`	`15`	`public string? ResponseFormat { get; init; }`
`16`	`16`	`public Grammar? Grammar { get; set; }`
`17`	`17`	`}`