Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

fix(integrations): batch LangGraph history writes to avoid duplicates and lost context (#3528)

Hao Zhe9d0b71f76351
diff --git a/openviking/integrations/langchain/middleware.py b/openviking/integrations/langchain/middleware.py
index d3038b70b..d73bd0255 100644
--- a/openviking/integrations/langchain/middleware.py
+++ b/openviking/integrations/langchain/middleware.py
@@ -41,6 +41,7 @@ from openviking.integrations.langchain.retrievers import OpenVikingRetriever
logger = logging.getLogger(__name__)
+_BATCH_ADD_MESSAGES_LIMIT = 100
_SESSION_ID_ERROR = (
"OpenVikingContextMiddleware requires a LangGraph session id. Pass "
'config={"configurable": {"thread_id": "..."}}, set state["session_id"], '
@@ -197,27 +198,41 @@ class OpenVikingContextMiddleware(AgentMiddleware):
start = len(previous_signatures)
client = ensure_client(self._connection)
- added = 0
- pending_context_parts = list(self._pending_context_parts.pop(capture_key, []))
- for message in messages[start:]:
+ batch: list[dict[str, Any]] = []
+ chunks: list[tuple[list[dict[str, Any]], int, bool]] = []
+ batch_has_context = False
+ pending_context_parts = list(self._pending_context_parts.get(capture_key, []))
+ for message_index, message in enumerate(messages[start:], start=start):
if OPENVIKING_CONTEXT_MARKER in _message_content(message):
continue
payloads = langchain_message_to_openviking(message)
+ if batch and len(batch) + len(payloads) > _BATCH_ADD_MESSAGES_LIMIT:
+ chunks.append((batch, message_index, batch_has_context))
+ batch = []
+ batch_has_context = False
for payload in payloads:
if pending_context_parts and payload["role"] == "assistant":
payload["parts"].extend(pending_context_parts)
pending_context_parts = []
- call_openviking(
- client,
- "add_message",
- session_id=session_id,
- role=payload["role"],
- parts=payload["parts"],
- peer_id=peer_id,
- )
- added += 1
+ batch_has_context = True
+ if peer_id is not None:
+ payload["peer_id"] = peer_id
+ batch.append(payload)
+ if batch:
+ chunks.append((batch, len(messages), batch_has_context))
+ for chunk, signature_end, has_context in chunks:
+ call_openviking(
+ client,
+ "batch_add_messages",
+ session_id=session_id,
+ messages=chunk,
+ )
+ self._captured_signatures[capture_key] = current_signatures[:signature_end]
+ if has_context:
+ self._pending_context_parts.pop(capture_key, None)
+ self._pending_context_parts.pop(capture_key, None)
self._captured_signatures[capture_key] = current_signatures
- if added:
+ if batch:
apply_commit_policy(client, session_id, self.commit_policy)
return None
diff --git a/tests/unit/test_langchain_integration.py b/tests/unit/test_langchain_integration.py
index 939f36d48..433af97ab 100644
--- a/tests/unit/test_langchain_integration.py
+++ b/tests/unit/test_langchain_integration.py
@@ -1677,3 +1677,112 @@ def test_langgraph_middleware_clears_pending_context_on_duplicate_retry():
unrelated_assistant_parts = client.sessions["middleware-pending-cleanup"][-1]["parts"]
assert unrelated_assistant_parts == [{"type": "text", "text": "Unrelated turn stored."}]
+
+
+def test_langgraph_middleware_retries_failed_batch_with_pending_context():
+ class FailFirstBatchClient(InMemoryOpenVikingClient):
+ def __init__(self):
+ super().__init__({"viking://user/memories/profile.md": "Retained retry context."})
+ self.batch_attempts = 0
+
+ def batch_add_messages(self, session_id, messages, **kwargs):
+ self.batch_attempts += 1
+ if self.batch_attempts == 1:
+ raise RuntimeError("batch write failed")
+ return super().batch_add_messages(session_id, messages, **kwargs)
+
+ client = FailFirstBatchClient()
+ middleware = OpenVikingContextMiddleware(
+ client=client,
+ target_uri="viking://user/memories",
+ session_id_resolver=lambda state, runtime: "middleware-batch-retry",
+ )
+
+ class Request:
+ messages = [HumanMessage(content="What retry context?")]
+ system_message = None
+
+ def override(self, **overrides):
+ new_request = Request()
+ new_request.messages = overrides.get("messages", self.messages)
+ new_request.system_message = overrides.get("system_message", self.system_message)
+ return new_request
+
+ middleware.wrap_model_call(Request(), lambda request: AIMessage(content="ok"))
+ state = {
+ "messages": [
+ HumanMessage(content="What retry context?"),
+ AIMessage(content="Retry context stored."),
+ ]
+ }
+
+ with pytest.raises(RuntimeError, match="batch write failed"):
+ middleware.after_agent(state, runtime=None)
+ middleware.after_agent(state, runtime=None)
+
+ assert client.batch_attempts == 2
+ assert len(client.sessions["middleware-batch-retry"]) == 2
+ assistant_parts = client.sessions["middleware-batch-retry"][1]["parts"]
+ assert sum(part["type"] == "context" for part in assistant_parts) == 1
+
+
+def test_langgraph_middleware_chunks_batches_at_server_limit():
+ class CappedBatchClient(InMemoryOpenVikingClient):
+ def __init__(self):
+ super().__init__()
+ self.batch_sizes = []
+
+ def batch_add_messages(self, session_id, messages, **kwargs):
+ if len(messages) > 100:
+ raise RuntimeError("server batch limit exceeded")
+ self.batch_sizes.append(len(messages))
+ return super().batch_add_messages(session_id, messages, **kwargs)
+
+ client = CappedBatchClient()
+ middleware = OpenVikingContextMiddleware(
+ client=client,
+ session_id_resolver=lambda state, runtime: "middleware-large-batch",
+ )
+ state = {"messages": [HumanMessage(content=f"Message {index}") for index in range(101)]}
+
+ middleware.after_agent(state, runtime=None)
+ middleware.after_agent(state, runtime=None)
+
+ assert client.batch_sizes == [100, 1]
+ assert len(client.sessions["middleware-large-batch"]) == 101
+
+
+def test_langgraph_middleware_retries_only_failed_batch_chunk():
+ class FailSecondBatchClient(InMemoryOpenVikingClient):
+ def __init__(self):
+ super().__init__()
+ self.batch_sizes = []
+
+ def batch_add_messages(self, session_id, messages, **kwargs):
+ self.batch_sizes.append(len(messages))
+ if len(self.batch_sizes) == 2:
+ raise RuntimeError("second batch failed")
+ return super().batch_add_messages(session_id, messages, **kwargs)
+
+ client = FailSecondBatchClient()
+ middleware = OpenVikingContextMiddleware(
+ client=client,
+ session_id_resolver=lambda state, runtime: "middleware-chunk-retry",
+ )
+ middleware._pending_context_parts[("middleware-chunk-retry", "")] = [
+ {"type": "context", "uri": "viking://resources/retry.md"}
+ ]
+ messages = [HumanMessage(content="First"), AIMessage(content="Context holder")]
+ messages.extend(HumanMessage(content=f"Message {index}") for index in range(98))
+ messages.append(AIMessage(content="Retry only this message"))
+ state = {"messages": messages}
+
+ with pytest.raises(RuntimeError, match="second batch failed"):
+ middleware.after_agent(state, runtime=None)
+ middleware.after_agent(state, runtime=None)
+ middleware.after_agent(state, runtime=None)
+
+ assert client.batch_sizes == [100, 1, 1]
+ stored = client.sessions["middleware-chunk-retry"]
+ assert len(stored) == 101
+ assert sum(part["type"] == "context" for message in stored for part in message["parts"]) == 1