✨(agent) add fake streaming when the model does not support it

To have a better UX, we simulate a streamed return. The default
value is quite fast to not block the worker too long...

src/backend/chat/clients/pydantic_ai.py

@@ -102,6 +102,7 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
         # Feature flags
         self._is_document_upload_enabled = is_feature_enabled(self.user, "document_upload")
         self._is_web_search_enabled = is_feature_enabled(self.user, "web_search")
+        self._fake_streaming_delay = settings.FAKE_STREAMING_DELAY
 
     @property
     def _stop_cache_key(self):
@@ -490,14 +491,20 @@ class AIAgentService:  # pylint: disable=too-many-instance-attributes
                         # A user prompt node => The user has provided input
                         pass
 
-                    elif Agent.is_model_request_node(node):
+                    elif Agent.is_model_request_node(node):  # pylint: disable=too-many-nested-blocks
                         # A model request node => agent is asking the model to generate a response
                         if not self._support_streaming:
                             result = await node.run(run.ctx)
                             logger.debug("node.run result: %s", result)
                             for part in result.model_response.parts:
                                 if isinstance(part, TextPart):
-                                    yield events_v4.TextPart(text=part.content)
+                                    if self._fake_streaming_delay:
+                                        for i in range(0, len(part.content), 4):
+                                            await self._agent_stop_streaming()
+                                            yield events_v4.TextPart(text=part.content[i : i + 4])
+                                            time.sleep(self._fake_streaming_delay)
+                                    else:
+                                        yield events_v4.TextPart(text=part.content)
                                 elif isinstance(part, ToolCallPart):
                                     yield events_v4.ToolCallPart(
                                         tool_call_id=part.tool_call_id,

src/backend/chat/tests/views/chat/conversations/test_conversation.py

@@ -1098,16 +1098,19 @@ def test_post_conversation_model_selection_new(api_client, mock_openai_stream, s
 
 
 @freeze_time("2025-07-25T10:36:35.297675Z")
+@pytest.mark.parametrize("stream_delay", [None, 0.0001])
 @respx.mock
 def test_post_conversation_data_protocol_no_stream(
     api_client,
     mock_openai_no_stream,
     mock_uuid4,
     settings,
+    stream_delay,
 ):
     """
     Test posting messages to a conversation using the 'data' protocol without streaming.
     """
+    settings.FAKE_STREAMING_DELAY = stream_delay
     settings.LLM_CONFIGURATIONS = {
         "default-model": LLModel(
             hrid="model-1",
@@ -1149,12 +1152,35 @@ def test_post_conversation_data_protocol_no_stream(
     assert response.get("x-vercel-ai-data-stream") == "v1"
     assert response.streaming
 
-    # Wait for the streaming content to be fully received
+    # Wait for the content to be fully received
     response_content = b"".join(response.streaming_content).decode("utf-8")
-    assert response_content == (
-        '0:"The sky appears blue due to a phenomenon called Rayleigh scattering."\n'
-        'd:{"finishReason":"stop","usage":{"promptTokens":0,"completionTokens":135}}\n'
-    )
+
+    if stream_delay:
+        assert response_content == (
+            '0:"The "\n'
+            '0:"sky "\n'
+            '0:"appe"\n'
+            '0:"ars "\n'
+            '0:"blue"\n'
+            '0:" due"\n'
+            '0:" to "\n'
+            '0:"a ph"\n'
+            '0:"enom"\n'
+            '0:"enon"\n'
+            '0:" cal"\n'
+            '0:"led "\n'
+            '0:"Rayl"\n'
+            '0:"eigh"\n'
+            '0:" sca"\n'
+            '0:"tter"\n'
+            '0:"ing."\n'
+            'd:{"finishReason":"stop","usage":{"promptTokens":0,"completionTokens":135}}\n'
+        )
+    else:
+        assert response_content == (
+            '0:"The sky appears blue due to a phenomenon called Rayleigh scattering."\n'
+            'd:{"finishReason":"stop","usage":{"promptTokens":0,"completionTokens":135}}\n'
+        )
 
     assert mock_openai_no_stream.called
 

src/backend/conversations/settings.py

@@ -471,6 +471,11 @@ class Base(Configuration):
     LLM_ROUTING_MODEL_HRID = values.Value(
         "default-routing-model", environ_name="LLM_ROUTING_MODEL_HRID", environ_prefix=None
     )
+    FAKE_STREAMING_DELAY = values.FloatValue(
+        default=0.0025,
+        environ_name="FAKE_STREAMING_DELAY",
+        environ_prefix=None,  # 25ms
+    )
 
     # These settings are default values used in the default LLM_CONFIGURATIONS
     # They allow a deployment with only one model without a specific configuration file