From 335b4f498c0c606d07da87e96283940527aed4b4 Mon Sep 17 00:00:00 2001
From: kcar <kcar@kevlarai.com>
Date: Wed, 20 May 2026 20:46:52 +0000
Subject: [PATCH] fix: vad control, session metadata, health endpoint

---
 hybrid_server.py | 32 ++++++++++++++++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)

diff --git a/hybrid_server.py b/hybrid_server.py
index a679630..a6085ff 100644
--- a/hybrid_server.py
+++ b/hybrid_server.py
@@ -91,7 +91,25 @@ class HybridWhisperServer:
     def setup_routes(self):
         @self.app.route('/health', methods=['GET'])
         def health_check():
-            return jsonify({'status': 'healthy', 'service': 'WhisperLive Hybrid Server'})
+            # Get GPU memory from nvidia-smi (GPU 1)
+            import subprocess
+            try:
+                gpu_mem = float(subprocess.check_output(
+                    'nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 1', shell=True
+                ).decode().strip()) / 1024.0
+            except Exception:
+                gpu_mem = 0.0
+            
+            # Get active WS connections
+            active = len(self.whisper_server.clients) if hasattr(self.whisper_server, 'clients') else 0
+            
+            return jsonify({
+                'status': 'healthy',
+                'service': 'WhisperLive Hybrid Server',
+                'model_loaded': self.shared_transcriber is not None,
+                'gpu_memory_used_gb': round(gpu_mem, 1),
+                'active_connections': active
+            })
         
         @self.app.route('/', methods=['GET'])
         def serve_test_form():
@@ -841,6 +859,7 @@ print(transcription.text)</code></pre>
                 language = request.form.get('language', None)
                 task = request.form.get('task', 'transcribe')  # 'transcribe' or 'translate'
                 model_size = request.form.get('model', 'base')
+                use_vad = request.form.get('use_vad', 'true').lower() == 'true'
                 
                 # For now, we'll use the shared transcriber regardless of the requested model size
                 # In the future, we could create different transcriber instances for different models
@@ -865,7 +884,8 @@ print(transcription.text)</code></pre>
                         segments, info = self.shared_transcriber.transcribe(
                             temp_path,
                             language=language,
-                            task=task
+                            task=task,
+                            vad_filter=use_vad
                         )
                     else:
                         # For other backends, use the server's transcriber
@@ -1104,6 +1124,14 @@ print(transcription.text)</code></pre>
                         if data == "END_OF_AUDIO":
                             internal.send_binary(b"END_OF_AUDIO")
                         else:
+                            # Pass through session_metadata if present in config message
+                            try:
+                                msg = json.loads(data)
+                                session_metadata = msg.get('session_metadata')
+                                if session_metadata:
+                                    logger.info(f"Session metadata received: session_id={session_metadata.get('session_id')}, teacher_id={session_metadata.get('teacher_id')}")
+                            except (json.JSONDecodeError, AttributeError):
+                                pass
                             internal.send(data)
 
             except Exception as e: