Expand Ollama context windows (num_ctx) and output limits

server/src/ollama.rs

@@ -23,6 +23,8 @@ struct GenerateRequest<'a> {
     stream: bool,
     #[serde(skip_serializing_if = "Option::is_none")]
     images: Option<Vec<&'a str>>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    options: Option<serde_json::Value>,
 }
 
 #[derive(Deserialize)]
@@ -50,12 +52,17 @@ impl OllamaClient {
         let reasoning_model =
             env::var("OLLAMA_REASONING_MODEL").unwrap_or_else(|_| "deepseek-r1:1.5b".to_string());
         let vision_model =
-            env::var("OLLAMA_VISION_MODEL").unwrap_or_else(|_| "qwen3-vl:2b".to_string());
+            env::var("OLLAMA_VISION_MODEL").unwrap_or_else(|_| "moondream".to_string());
         let embed_model = env::var("OLLAMA_EMBED_MODEL")
             .unwrap_or_else(|_| "nomic-embed-text:latest".to_string());
 
+        let timeout_sec = env::var("OLLAMA_TIMEOUT_SEC")
+            .unwrap_or_else(|_| "60".to_string())
+            .parse::<u64>()
+            .unwrap_or(60);
+
         let client = reqwest::Client::builder()
-            .timeout(Duration::from_secs(60))
+            .timeout(Duration::from_secs(timeout_sec))
             .build()
             .unwrap_or_default();
 
@@ -112,6 +119,10 @@ impl OllamaClient {
             system,
             stream: false,
             images: None,
+            options: Some(serde_json::json!({
+                "num_ctx": 32768, // Massive context window win
+                "num_predict": 4096 // Give reasoning models plenty of output room
+            })),
         };
 
         let res = self
@@ -151,6 +162,10 @@ impl OllamaClient {
             ),
             stream: false,
             images: Some(vec![image_base64]),
+            options: Some(serde_json::json!({
+                "num_ctx": 8192,
+                "num_predict": 1024
+            })),
         };
 
         let res = self