Skip to content

The agent engineering platform.

anthropic: update streaming usage metadata (#31158)

ccurmee34f9fd6f71b
diff --git a/libs/partners/anthropic/langchain_anthropic/chat_models.py b/libs/partners/anthropic/langchain_anthropic/chat_models.py
index 26c6a7f39b..e3e1bebc40 100644
--- a/libs/partners/anthropic/langchain_anthropic/chat_models.py
+++ b/libs/partners/anthropic/langchain_anthropic/chat_models.py
@@ -1744,6 +1744,12 @@ def _make_message_chunk_from_anthropic_event(
# See https://github.com/anthropics/anthropic-sdk-python/blob/main/src/anthropic/lib/streaming/_messages.py # noqa: E501
if event.type == "message_start" and stream_usage:
usage_metadata = _create_usage_metadata(event.message.usage)
+ # We pick up a cumulative count of output_tokens at the end of the stream,
+ # so here we zero out to avoid double counting.
+ usage_metadata["total_tokens"] = (
+ usage_metadata["total_tokens"] - usage_metadata["output_tokens"]
+ )
+ usage_metadata["output_tokens"] = 0
if hasattr(event.message, "model"):
response_metadata = {"model_name": event.message.model}
else:
@@ -1817,7 +1823,11 @@ def _make_message_chunk_from_anthropic_event(
tool_call_chunks=[tool_call_chunk], # type: ignore
)
elif event.type == "message_delta" and stream_usage:
- usage_metadata = _create_usage_metadata(event.usage)
+ usage_metadata = UsageMetadata(
+ input_tokens=0,
+ output_tokens=event.usage.output_tokens,
+ total_tokens=event.usage.output_tokens,
+ )
message_chunk = AIMessageChunk(
content="",
usage_metadata=usage_metadata,
diff --git a/libs/partners/anthropic/tests/integration_tests/test_chat_models.py b/libs/partners/anthropic/tests/integration_tests/test_chat_models.py
index 5c41c3f878..befdca0ec1 100644
--- a/libs/partners/anthropic/tests/integration_tests/test_chat_models.py
+++ b/libs/partners/anthropic/tests/integration_tests/test_chat_models.py
@@ -46,7 +46,7 @@ def test_stream() -> None:
if token.usage_metadata is not None:
if token.usage_metadata.get("input_tokens"):
chunks_with_input_token_counts += 1
- elif token.usage_metadata.get("output_tokens"):
+ if token.usage_metadata.get("output_tokens"):
chunks_with_output_token_counts += 1
chunks_with_model_name += int("model_name" in token.response_metadata)
if chunks_with_input_token_counts != 1 or chunks_with_output_token_counts != 1:
@@ -85,7 +85,7 @@ async def test_astream() -> None:
if token.usage_metadata is not None:
if token.usage_metadata.get("input_tokens"):
chunks_with_input_token_counts += 1
- elif token.usage_metadata.get("output_tokens"):
+ if token.usage_metadata.get("output_tokens"):
chunks_with_output_token_counts += 1
if chunks_with_input_token_counts != 1 or chunks_with_output_token_counts != 1:
raise AssertionError(