fix: fallback streaming

ee/hogai/assistant.py

@@ -1,5 +1,5 @@
 from collections.abc import Generator
-from typing import cast
+from typing import Any, Literal, TypedDict, TypeGuard, Union, cast
 
 from langchain_core.messages import AIMessageChunk
 from langfuse.callback import CallbackHandler
@@ -22,6 +22,33 @@
 )
 
 
+def is_state_update(update: list[Any]) -> TypeGuard[tuple[Literal["updates"], AssistantState]]:
+    """
+    Update of the state.
+    """
+    return len(update) == 2 and update[0] == "values"
+
+
+def is_value_update(update: list[Any]) -> TypeGuard[tuple[Literal["values"], dict[AssistantNodeName, Any]]]:
+    """
+    Transition between nodes.
+    """
+    return len(update) == 2 and update[0] == "updates"
+
+
+class LangGraphState(TypedDict):
+    langgraph_node: AssistantNodeName
+
+
+def is_message_update(
+    update: list[Any],
+) -> TypeGuard[tuple[Literal["messages"], tuple[Union[AssistantMessage, AIMessageChunk, Any], LangGraphState]]]:
+    """
+    Streaming of messages. Returns a partial state.
+    """
+    return len(update) == 2 and update[0] == "messages"
+
+
 class Assistant:
     _team: Team
     _graph: StateGraph
@@ -55,30 +82,40 @@ def _compile_graph(self):
 
     def stream(self, messages: list[AssistantMessage]) -> Generator[str, None, None]:
         assistant_graph = self._compile_graph()
-        generator = assistant_graph.stream(
-            {"messages": messages},
-            config={"recursion_limit": 24, "callbacks": [langfuse_handler]},
-            stream_mode="messages",
-        )
 
         chunks = AIMessageChunk(content="")
+        state: AssistantState = {"messages": messages, "intermediate_steps": None, "plan": None}
 
-        for message, state in generator:
-            if state["langgraph_node"] == AssistantNodeName.GENERATE_TRENDS:
-                if isinstance(message, AssistantMessage):
-                    yield FrontendAssistantMessage(
-                        type=message.type, content=message.content, payload=message.payload
-                    ).model_dump_json()
-                elif isinstance(message, AIMessageChunk):
-                    message = cast(AIMessageChunk, message)
-                    chunks += message  # type: ignore
-                    parsed_message = GenerateTrendsNode.parse_output(chunks.tool_calls[0]["args"])
-                    if parsed_message:
+        for update in assistant_graph.stream(
+            state,
+            config={"recursion_limit": 24, "callbacks": [langfuse_handler]},
+            stream_mode=["messages", "values", "updates"],
+        ):
+            if is_state_update(update):
+                _, new_state = update
+                state = new_state
+
+            elif is_value_update(update):
+                _, payload = update
+
+                # Reset chunks when schema validation fails.
+                if AssistantNodeName.GENERATE_TRENDS in payload:
+                    chunks = AIMessageChunk(content="")
+
+            elif is_message_update(update):
+                message, langraph_state = update[1]
+                if langraph_state["langgraph_node"] == AssistantNodeName.GENERATE_TRENDS:
+                    if isinstance(message, AssistantMessage):
                         yield FrontendAssistantMessage(
-                            type="ai",
-                            content=parsed_message.model_dump_json(),
-                            payload=VisualizationMessagePayload(plan=""),
+                            type=message.type, content=message.content, payload=message.payload
                         ).model_dump_json()
-            elif state["langgraph_node"] == AssistantNodeName.GENERATE_TRENDS_TOOLS:
-                # Reset tool output parser when encountered a validation error
-                chunks = AIMessageChunk(content="")
+                    elif isinstance(message, AIMessageChunk):
+                        message = cast(AIMessageChunk, message)
+                        chunks += message  # type: ignore
+                        parsed_message = GenerateTrendsNode.parse_output(chunks.tool_calls[0]["args"])
+                        if parsed_message:
+                            yield FrontendAssistantMessage(
+                                type="ai",
+                                content=parsed_message.model_dump_json(),
+                                payload=VisualizationMessagePayload(plan=""),
+                            ).model_dump_json()