fix(ollama): improve model detection and ensure robust token/context limits

- Use case-insensitive matching for model names and routing
- Default max_tokens/num_predict to 8192 for all Ollama models to prevent truncation
- Increase default context window and add more large-context model families
- Ensure DeepSeek routing handles Ollama-hosted variants correctly

internal/providers/ollama.go

@@ -117,11 +117,20 @@ func BuildOllamaBody(request *models.UnifiedRequest, messagesJSON []interface{},
 	}
 
 	options := make(map[string]interface{})
+	modelLower := strings.ToLower(request.Model)
 
-	// Context window size (32k for modern models to avoid truncation)
-	if strings.Contains(request.Model, "gemma") || strings.Contains(request.Model, "llama") || strings.Contains(request.Model, "mistral") || strings.Contains(request.Model, "qwen") {
-		options["num_ctx"] = 32768
+	// Context window size (default 8k for all, 32k+ for modern large-context models)
+	ctxSize := 8192
+	if strings.Contains(modelLower, "llama3") || 
+	   strings.Contains(modelLower, "mistral") || 
+	   strings.Contains(modelLower, "mixtral") ||
+	   strings.Contains(modelLower, "qwen") || 
+	   strings.Contains(modelLower, "deepseek") ||
+	   strings.Contains(modelLower, "command-r") ||
+	   strings.Contains(modelLower, "phi3") {
+		ctxSize = 32768
 	}
+	options["num_ctx"] = ctxSize
 
 	if request.Temperature != nil {
 		body["temperature"] = *request.Temperature
@@ -131,9 +140,10 @@ func BuildOllamaBody(request *models.UnifiedRequest, messagesJSON []interface{},
 	if request.MaxTokens != nil {
 		body["max_tokens"] = *request.MaxTokens
 		options["num_predict"] = *request.MaxTokens
-	} else if strings.Contains(request.Model, "gemma") || strings.Contains(request.Model, "llama") || strings.Contains(request.Model, "mistral") || strings.Contains(request.Model, "qwen") {
-		// Default to 8192 for common Ollama models if not specified, 
-		// as Ollama's compatibility layer sometimes defaults to 128
+	} else {
+		// Default to 8192 for all Ollama models if not specified, 
+		// as Ollama's compatibility layer defaults to 128 if neither 
+		// max_tokens nor num_predict are provided.
 		body["max_tokens"] = 8192
 		options["num_predict"] = 8192
 	}