Merge pull request #36 from tokk-nv/feat/reasoning-user-prompt

feat: append reasoning prompt to user message instead of system prompt

Author: tokk-nv

src/multi_modal_ai_studio/backends/llm/openai.py

@@ -481,17 +481,22 @@ async def generate_stream(
         if self.config.minimal_output:
             suffix = " Answer with only a number or minimal tokens. No reasoning or explanation."
             sys_prompt = (sys_prompt or "") + suffix
-        elif getattr(self.config, "enable_reasoning", False):
-            reasoning_fmt = getattr(self.config, "reasoning_prompt", "") or ""
-            if reasoning_fmt:
-                sys_prompt = (sys_prompt or "") + reasoning_fmt
         if sys_prompt:
             messages.append({"role": "system", "content": sys_prompt})
 
         # Add history
         if history:
             messages.extend(history)
 
+        # Append reasoning prompt to user message when enabled
+        if (
+            not self.config.minimal_output
+            and getattr(self.config, "enable_reasoning", False)
+        ):
+            reasoning_fmt = getattr(self.config, "reasoning_prompt", "") or ""
+            if reasoning_fmt:
+                prompt = prompt + reasoning_fmt
+
         if image_data_urls and len(image_data_urls) > 0:
             user_content = self._build_vision_content(
                 image_data_urls, prompt, speech_duration,

src/multi_modal_ai_studio/config/schema.py

@@ -161,7 +161,7 @@ class LLMConfig:
     vision_buffer_fps: float = 3.0  # Ring buffer capture rate (fps)
     vision_api_format: Literal["openai", "tensorrt_edge"] = "openai"
     vision_video_encode: bool = False  # True for models with video input (e.g. Cosmos-Reason)
-    enable_reasoning: bool = False  # True to allow <think> chain-of-thought; filtered from TTS
+    enable_reasoning: bool = False  # Append reasoning prompt to every user message; <think> output filtered from TTS
     reasoning_prompt: str = (
         "\n\nThink step-by-step inside <think> tags, then write your final answer "
         "immediately after </think>. The final answer MUST be exactly one descriptive "

src/multi_modal_ai_studio/webui/static/app.js

@@ -1443,14 +1443,14 @@ function renderLLMConfig(config, readonly = false) {
                 <label class="checkbox-label">
                     <input type="checkbox" ${disabled} id="llm-enable-reasoning" ${config.enable_reasoning ? 'checked' : ''}
                            onchange="updateConfig('llm', 'enable_reasoning', this.checked); var grp = document.getElementById('reasoning-prompt-group'); grp.style.display = this.checked ? 'block' : 'none'; if (this.checked &amp;&amp; !currentConfig.llm.reasoning_prompt) { var ta = document.getElementById('llm-reasoning-prompt'); if (ta) updateConfig('llm', 'reasoning_prompt', ta.value); }">
-                    Enable reasoning (chain-of-thought) through system prompt
+                    Enable reasoning (chain-of-thought) through user prompt
                 </label>
-                ${!readonly ? '<span class="input-hint">Appends the reasoning prompt below to the system prompt. Reasoning text is stripped before TTS.</span>' : ''}
+                ${!readonly ? '<span class="input-hint">Appends the reasoning prompt below to every user message. Reasoning text is stripped before TTS.</span>' : ''}
                 <div id="reasoning-prompt-group" style="display: ${config.enable_reasoning ? 'block' : 'none'}; margin-top: 8px;">
                     <label>Reasoning Prompt</label>
                     <textarea id="llm-reasoning-prompt" ${disabled} rows="4" style="font-family: var(--font-mono); font-size: 0.85rem;" placeholder="e.g. Think step-by-step in <think>...</think> then your final answer."
                               onchange="updateConfig('llm', 'reasoning_prompt', this.value)">${escapeHtml(config.reasoning_prompt ?? defaultConfig.llm.reasoning_prompt ?? '')}</textarea>
-                    ${!readonly ? '<span class="input-hint">Appended to system prompt when reasoning is enabled. Customise for your model\'s format.</span>' : ''}
+                    ${!readonly ? '<span class="input-hint">Appended to every user message when reasoning is enabled. Customise for your model\'s format.</span>' : ''}
                 </div>
             </div>