fixed streaming

app.py

@@ -100,6 +100,7 @@ def retrieve_context(query, top_k=2):
 def stream_response(query):
     logger.info(f"Processing query: {query}")
     start_time = time.time()
+    first_token_logged = False  # Flag to log first token time only once
     
     # FAQ check first
     query_embedding = embedder.encode(query, convert_to_numpy=True).astype("float32")
@@ -132,16 +133,15 @@ def stream_response(query):
         text = chunk['choices'][0]['delta'].get('content', '')
         if text:
             buffer += text
+            if not first_token_logged and time.time() - start_time > 0:  # Log first token once
+                logger.info(f"First token time: {time.time() - start_time:.2f}s")
+                first_token_logged = True
             if buffer.endswith(" ") or buffer.endswith(".") or buffer.endswith("!"):
                 yield f"data: {buffer}\n\n"
                 buffer = ""
-        if time.time() - start_time > 1:  # Log first token
-            logger.info(f"First token time: {time.time() - start_time:.2f}s")
-            break
-    if buffer:
+    if buffer:  # Flush any remaining buffer
         yield f"data: {buffer}\n\n"
     yield "data: [DONE]\n\n"
-
 class QueryRequest(BaseModel):
     data: list