feat: improve prompt formatting and system metrics

- Update prompt formatting in BotOrchestrator to use clearer labels (SYSTEM/CONTEXT) with emphasis markers
- Remove unused token_ratio field from SystemMetrics struct
- Increase default context size (2048->4096) and prediction length (512->1024) in config
- Clean up metrics calculation by removing redundant token ratio computation

The changes improve readability of system prompts and simplify metrics collection while increasing default model capacity.

src/bot/mod.rs

@@ -393,10 +393,10 @@ impl BotOrchestrator {
 
         let mut prompt = String::new();
         if !system_prompt.is_empty() {
-            prompt.push_str(&format!("AI:{}\n", system_prompt));
+            prompt.push_str(&format!("SYSTEM: *** {} *** \n", system_prompt));
         }
         if !context_data.is_empty() {
-            prompt.push_str(&format!("CTX:{}\n", context_data));
+            prompt.push_str(&format!("CONTEXT: *** {} *** \n", context_data));
         }
         for (role, content) in &history {
             prompt.push_str(&format!("{}:{}\n", role, content));

src/nvidia/mod.rs

@@ -7,7 +7,6 @@ use sysinfo::{System};
 pub struct SystemMetrics {
     pub gpu_usage: Option<f32>,
     pub cpu_usage: f32,
-    pub token_ratio: f32,
 }
 
 /// Gets current system metrics
@@ -25,17 +24,10 @@ pub fn get_system_metrics(current_tokens: usize, max_tokens: usize) -> Result<Sy
         None
     };
 
-    // Calculate token ratio
-    let token_ratio = if max_tokens > 0 {
-        current_tokens as f32 / max_tokens as f32 * 100.0
-    } else {
-        0.0
-    };
 
     Ok(SystemMetrics {
         gpu_usage,
         cpu_usage,
-        token_ratio,
     })
 }
 

templates/default.gbai/default.gbot/config.csv

@@ -19,8 +19,8 @@ llm-server-host,0.0.0.0
 llm-server-port,8081
 llm-server-gpu-layers,0
 llm-server-n-moe,0
-llm-server-ctx-size,2048
+llm-server-ctx-size,4096
-llm-server-n-predict,512
+llm-server-n-predict,1024
 llm-server-parallel,6
 llm-server-cont-batching,true
 llm-server-mlock,false