Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix(sdk): preserve agent answer references (#19365)

Wenchao An099e4580a05a
diff --git a/sdk/python/ragflow_sdk/modules/session.py b/sdk/python/ragflow_sdk/modules/session.py
index 9cbd5d9c34..ef3b7a5e2b 100644
--- a/sdk/python/ragflow_sdk/modules/session.py
+++ b/sdk/python/ragflow_sdk/modules/session.py
@@ -121,13 +121,22 @@ class Session(Base):
continue # Skip lines that are not valid JSON
event = json_data.get("event", None)
- if event and event != "message":
- continue
+ if event:
+ if self.__session_type == "agent" and event not in {"message", "message_end"}:
+ continue
+ if self.__session_type == "chat" and event != "message":
+ continue
- if (self.__session_type == "agent" and event == "message_end") or (self.__session_type == "chat" and json_data.get("data") is True):
+ if self.__session_type == "chat" and json_data.get("data") is True:
return
if self.__session_type == "agent":
- yield self._structure_answer(json_data)
+ message = self._structure_answer(json_data)
+ if event == "message_end":
+ message.content = ""
+ if message.reference:
+ yield message
+ continue
+ yield message
else:
yield self._structure_answer(json_data["data"])
else:
@@ -139,15 +148,23 @@ class Session(Base):
def _structure_answer(self, json_data):
answer = ""
+ event = None
if self.__session_type == "agent":
- answer = json_data["data"]["content"]
+ event = json_data.get("event")
+ json_data = json_data["data"]
+ answer = json_data.get("content", "")
elif self.__session_type == "chat":
answer = json_data["answer"]
reference = json_data.get("reference", {})
temp_dict = {"content": answer, "role": "assistant"}
if reference and "chunks" in reference:
chunks = reference["chunks"]
+ if isinstance(chunks, dict):
+ chunks = list(chunks.values())
temp_dict["reference"] = chunks
+ if self.__session_type == "agent":
+ reference_count = len(chunks) if isinstance(chunks, list) else 0
+ logger.debug("Session.ask parsed agent references session_id=%s event=%s reference_count=%s", self.id, event, reference_count)
message = Message(self.rag, temp_dict)
return message
diff --git a/test/testcases/test_sdk_api/test_session_management/test_create_session_with_chat_assistant.py b/test/testcases/test_sdk_api/test_session_management/test_create_session_with_chat_assistant.py
index 1e0ded1576..08fb241162 100644
--- a/test/testcases/test_sdk_api/test_session_management/test_create_session_with_chat_assistant.py
+++ b/test/testcases/test_sdk_api/test_session_management/test_create_session_with_chat_assistant.py
@@ -32,6 +32,14 @@ class _DummyStreamResponse:
yield line
+class _DummyJsonResponse:
+ def __init__(self, payload):
+ self._payload = payload
+
+ def json(self):
+ return self._payload
+
+
@pytest.mark.usefixtures("clear_session_with_chat_assistants")
class TestSessionWithChatAssistantCreate:
@pytest.mark.p3
@@ -156,3 +164,54 @@ def test_session_module_streaming_and_helper_paths_unit(monkeypatch):
assert calls[1][2]["openai-compatible"] is False
assert calls[1][2]["top_p"] == 0.8
assert calls[1][3] is True
+
+
+def test_agent_session_preserves_non_stream_reference_chunks(monkeypatch, caplog):
+ first_chunk = {"id": "chunk-1", "content": "first source"}
+ second_chunk = {"id": "chunk-2", "content": "second source"}
+ session = Session(None, {"id": "session-agent", "agent_id": "agent-1"})
+ response = _DummyJsonResponse(
+ {
+ "data": {
+ "event": "message_end",
+ "data": {
+ "content": "agent-answer",
+ "reference": {"chunks": {"500": first_chunk, "100": second_chunk}},
+ },
+ }
+ }
+ )
+ monkeypatch.setattr(session, "post", lambda *_args, **_kwargs: response)
+
+ with caplog.at_level("DEBUG", logger="ragflow_sdk.modules.session"):
+ messages = list(session.ask("hello agent", stream=False))
+
+ assert len(messages) == 1
+ assert messages[0].content == "agent-answer"
+ assert messages[0].reference == [first_chunk, second_chunk]
+ assert "session_id=session-agent event=message_end reference_count=2" in caplog.text
+ assert "first source" not in caplog.text
+ assert "second source" not in caplog.text
+
+
+def test_agent_session_streams_all_message_references_without_duplicate_content(monkeypatch):
+ first_chunk = {"id": "chunk-1", "content": "first source"}
+ second_chunk = {"id": "chunk-2", "content": "second source"}
+ session = Session(None, {"id": "session-agent", "agent_id": "agent-1"})
+ response = _DummyStreamResponse(
+ [
+ 'data: {"event":"message","data":{"content":"first answer"}}',
+ 'data: {"event":"message_end","data":{"content":"first answer","reference":{"chunks":{"123":{"id":"chunk-1","content":"first source"}}}}}',
+ 'data: {"event":"message","data":{"content":"second answer"}}',
+ 'data: {"event":"message_end","data":{"content":"second answer","reference":{"chunks":{"456":{"id":"chunk-2","content":"second source"}}}}}',
+ "data: [DONE]",
+ ]
+ )
+ monkeypatch.setattr(session, "post", lambda *_args, **_kwargs: response)
+
+ messages = list(session.ask("hello agent", stream=True))
+
+ assert "".join(message.content for message in messages) == "first answersecond answer"
+ assert [message.reference for message in messages if message.reference] == [[first_chunk], [second_chunk]]