checkpoint 2

app.py

@@ -514,7 +514,6 @@ async def ws_voice(ws: WebSocket):
                     full_text += token
                     if not await _safe_text(ws, {"type": "llm_token", "token": token, "turn": my_turn}):
                         break
-                    await tts_streamer.add_token(token)
             except asyncio.CancelledError:
                 raise
             except Exception as exc:
@@ -524,6 +523,9 @@ async def ws_voice(ws: WebSocket):
                 # recover if it missed any streamed tokens.
                 if full_text:
                     await _safe_text(ws, {"type": "llm_full", "text": _normalize_ai_text(full_text), "turn": my_turn})
+                    # Voice synthesis uses the completed response so TTS gets
+                    # full sentence context instead of fragmentary token chunks.
+                    await tts_streamer.add_token(full_text)
                 await tts_streamer.flush()
 
         async def run_tts_framed():

core/backend.py

@@ -1430,7 +1430,7 @@ Important booking rules:
 - If the user has already confirmed the details, book immediately.
 
 Important update rules:
-- First ask what the user wants to change (date/doctor/time/phone/email).
+- First ask what the user wants to change (date/doctor/time).
 - Ask whether to keep the rest of the existing appointment unchanged.
 - If multiple appointments exist for a phone number, ask for the doctor name to select the correct one.
 - Email is REQUIRED to update. If the existing record has no email, ask for it.
@@ -1664,16 +1664,17 @@ class AIBackend:
         async for chunk, _meta in self.graph.astream(
             initial_state, config=config, stream_mode="messages"
         ):
-            # Only yield text content from AI messages.
-            # Exclude ToolMessage (tool execution results) — they contain
-            # raw JSON that should not be streamed directly to the user.
+            # Only yield assistant text. Gemini may return structured content
+            # as a list of text parts, so flatten it before streaming.
+            # Exclude ToolMessage (tool execution results) — they contain raw
+            # JSON that should not be streamed directly to the user.
             if (
                 isinstance(chunk, (AIMessage, AIMessageChunk))
                 and not isinstance(chunk, ToolMessage)
-                and isinstance(chunk.content, str)
-                and chunk.content
             ):
-                yield chunk.content
+                text = _message_text(chunk.content)
+                if text:
+                    yield text
 
         # Auto-summarise in background when history grows long
         try:

services/webrtc_pipeline.py

@@ -216,20 +216,24 @@ class _TurnPipeline:
 
     async def _run_llm(self, user_id: str, transcript: str) -> None:
         """Stream LLM tokens → TTS streamer (concurrent with audio delivery)."""
+        full_text = ""
         try:
             stream = await self._ai.main(user_id, transcript)
             async for token in stream:
                 if self._cancelled or not token:
                     break
+                full_text += token
                 if self._on_token:
                     self._on_token(token)
                 self._send_ctrl({"type": "llm_token", "token": token})
-                await self._streamer.add_token(token)
         except asyncio.CancelledError:
             raise
         except Exception as exc:
             print(f"[Pipeline] LLM error: {exc}")
         finally:
+            # Feed the completed response to TTS for more reliable synthesis.
+            if full_text and not self._cancelled:
+                await self._streamer.add_token(full_text)
             await self._streamer.flush()
 
     async def _run_tts_delivery(self) -> None: