Skip to content

The agent engineering platform.

core, openai[patch]: prefer provider-assigned IDs when aggregating message chunks (#31080)

ccurme26ad239669ca
diff --git a/libs/core/langchain_core/language_models/chat_models.py b/libs/core/langchain_core/language_models/chat_models.py
index f0493497c1..712271778a 100644
--- a/libs/core/langchain_core/language_models/chat_models.py
+++ b/libs/core/langchain_core/language_models/chat_models.py
@@ -58,6 +58,7 @@ from langchain_core.messages import (
is_data_content_block,
message_chunk_to_message,
)
+from langchain_core.messages.ai import _LC_ID_PREFIX
from langchain_core.outputs import (
ChatGeneration,
ChatGenerationChunk,
@@ -493,7 +494,7 @@ class BaseChatModel(BaseLanguageModel[BaseMessage], ABC):
input_messages = _normalize_messages(messages)
for chunk in self._stream(input_messages, stop=stop, **kwargs):
if chunk.message.id is None:
- chunk.message.id = f"run-{run_manager.run_id}"
+ chunk.message.id = f"{_LC_ID_PREFIX}-{run_manager.run_id}"
chunk.message.response_metadata = _gen_info_and_msg_metadata(chunk)
run_manager.on_llm_new_token(
cast("str", chunk.message.content), chunk=chunk
@@ -583,7 +584,7 @@ class BaseChatModel(BaseLanguageModel[BaseMessage], ABC):
**kwargs,
):
if chunk.message.id is None:
- chunk.message.id = f"run-{run_manager.run_id}"
+ chunk.message.id = f"{_LC_ID_PREFIX}-{run_manager.run_id}"
chunk.message.response_metadata = _gen_info_and_msg_metadata(chunk)
await run_manager.on_llm_new_token(
cast("str", chunk.message.content), chunk=chunk
@@ -1001,7 +1002,7 @@ class BaseChatModel(BaseLanguageModel[BaseMessage], ABC):
chunk.message.response_metadata = _gen_info_and_msg_metadata(chunk)
if run_manager:
if chunk.message.id is None:
- chunk.message.id = f"run-{run_manager.run_id}"
+ chunk.message.id = f"{_LC_ID_PREFIX}-{run_manager.run_id}"
run_manager.on_llm_new_token(
cast("str", chunk.message.content), chunk=chunk
)
@@ -1017,7 +1018,7 @@ class BaseChatModel(BaseLanguageModel[BaseMessage], ABC):
# Add response metadata to each generation
for idx, generation in enumerate(result.generations):
if run_manager and generation.message.id is None:
- generation.message.id = f"run-{run_manager.run_id}-{idx}"
+ generation.message.id = f"{_LC_ID_PREFIX}-{run_manager.run_id}-{idx}"
generation.message.response_metadata = _gen_info_and_msg_metadata(
generation
)
@@ -1073,7 +1074,7 @@ class BaseChatModel(BaseLanguageModel[BaseMessage], ABC):
chunk.message.response_metadata = _gen_info_and_msg_metadata(chunk)
if run_manager:
if chunk.message.id is None:
- chunk.message.id = f"run-{run_manager.run_id}"
+ chunk.message.id = f"{_LC_ID_PREFIX}-{run_manager.run_id}"
await run_manager.on_llm_new_token(
cast("str", chunk.message.content), chunk=chunk
)
@@ -1089,7 +1090,7 @@ class BaseChatModel(BaseLanguageModel[BaseMessage], ABC):
# Add response metadata to each generation
for idx, generation in enumerate(result.generations):
if run_manager and generation.message.id is None:
- generation.message.id = f"run-{run_manager.run_id}-{idx}"
+ generation.message.id = f"{_LC_ID_PREFIX}-{run_manager.run_id}-{idx}"
generation.message.response_metadata = _gen_info_and_msg_metadata(
generation
)
diff --git a/libs/core/langchain_core/messages/ai.py b/libs/core/langchain_core/messages/ai.py
index 918f717aa5..6b06786418 100644
--- a/libs/core/langchain_core/messages/ai.py
+++ b/libs/core/langchain_core/messages/ai.py
@@ -36,6 +36,9 @@ from langchain_core.utils.usage import _dict_int_op
logger = logging.getLogger(__name__)
+_LC_ID_PREFIX = "run-"
+
+
class InputTokenDetails(TypedDict, total=False):
"""Breakdown of input token counts.
@@ -418,10 +421,19 @@ def add_ai_message_chunks(
usage_metadata = None
id = None
- for id_ in [left.id] + [o.id for o in others]:
- if id_:
+ candidates = [left.id] + [o.id for o in others]
+ # first pass: pick the first non‐run-* id
+ for id_ in candidates:
+ if id_ and not id_.startswith(_LC_ID_PREFIX):
id = id_
break
+ else:
+ # second pass: no provider-assigned id found, just take the first non‐null
+ for id_ in candidates:
+ if id_:
+ id = id_
+ break
+
return left.__class__(
example=left.example,
content=content,
diff --git a/libs/core/tests/unit_tests/test_messages.py b/libs/core/tests/unit_tests/test_messages.py
index 5899468537..e6cc725cfc 100644
--- a/libs/core/tests/unit_tests/test_messages.py
+++ b/libs/core/tests/unit_tests/test_messages.py
@@ -178,6 +178,22 @@ def test_message_chunks() -> None:
assert AIMessageChunk(content="") + left == left
assert right + AIMessageChunk(content="") == right
+ # Test ID order of precedence
+ null_id = AIMessageChunk(content="", id=None)
+ default_id = AIMessageChunk(
+ content="", id="run-abc123"
+ ) # LangChain-assigned run ID
+ meaningful_id = AIMessageChunk(content="", id="msg_def456") # provider-assigned ID
+
+ assert (null_id + default_id).id == "run-abc123"
+ assert (default_id + null_id).id == "run-abc123"
+
+ assert (null_id + meaningful_id).id == "msg_def456"
+ assert (meaningful_id + null_id).id == "msg_def456"
+
+ assert (default_id + meaningful_id).id == "msg_def456"
+ assert (meaningful_id + default_id).id == "msg_def456"
+
def test_chat_message_chunks() -> None:
assert ChatMessageChunk(role="User", content="I am", id="ai4") + ChatMessageChunk(
diff --git a/libs/partners/openai/langchain_openai/chat_models/base.py b/libs/partners/openai/langchain_openai/chat_models/base.py
index 7b13029f50..64bb56a5a2 100644
--- a/libs/partners/openai/langchain_openai/chat_models/base.py
+++ b/libs/partners/openai/langchain_openai/chat_models/base.py
@@ -3127,6 +3127,7 @@ def _construct_responses_api_input(messages: Sequence[BaseMessage]) -> list:
reasoning_items = []
if reasoning := lc_msg.additional_kwargs.get("reasoning"):
reasoning_items.append(_pop_summary_index_from_reasoning(reasoning))
+ input_.extend(reasoning_items)
# Function calls
function_calls = []
if tool_calls := msg.pop("tool_calls", None):
@@ -3185,13 +3186,11 @@ def _construct_responses_api_input(messages: Sequence[BaseMessage]) -> list:
pass
msg["content"] = new_blocks
if msg["content"]:
+ if lc_msg.id and lc_msg.id.startswith("msg_"):
+ msg["id"] = lc_msg.id
input_.append(msg)
input_.extend(function_calls)
- if computer_calls:
- # Hack: we only add reasoning items if computer calls are present. See:
- # https://community.openai.com/t/how-to-solve-badrequesterror-400-item-rs-of-type-reasoning-was-provided-without-its-required-following-item-error-in-responses-api/1151686/5
- input_.extend(reasoning_items)
- input_.extend(computer_calls)
+ input_.extend(computer_calls)
elif msg["role"] == "user":
if isinstance(msg["content"], list):
new_blocks = []
diff --git a/libs/partners/openai/tests/integration_tests/chat_models/test_responses_api.py b/libs/partners/openai/tests/integration_tests/chat_models/test_responses_api.py
index 2b22e6d4c3..d05cfa82c2 100644
--- a/libs/partners/openai/tests/integration_tests/chat_models/test_responses_api.py
+++ b/libs/partners/openai/tests/integration_tests/chat_models/test_responses_api.py
@@ -271,7 +271,7 @@ def test_function_calling_and_structured_output() -> None:
"""return x * y"""
return x * y
- llm = ChatOpenAI(model=MODEL_NAME)
+ llm = ChatOpenAI(model=MODEL_NAME, use_responses_api=True)
bound_llm = llm.bind_tools([multiply], response_format=Foo, strict=True)
# Test structured output
response = llm.invoke("how are ya", response_format=Foo)