Skip to content

fix(anthropic): emit prompt cache usage attributes #1450

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
jstar0 wants to merge 1 commit into
base: main
Choose a base branch
from
+108 −60
Open

fix(anthropic): emit prompt cache usage attributes #1450

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
69 changes: 32 additions & 37 deletions agentops/instrumentation/providers/anthropic/attributes/message.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,35 @@
)


def get_usage_attributes(usage: Any) -> AttributeMap:
"""Extract Anthropic token usage, including prompt-cache breakdowns.

Anthropic reports cache read and cache creation tokens separately from
native ``input_tokens``. Keep the existing native total unchanged and
expose those fields independently so consumers can choose their accounting
policy without double-counting cache tokens.
"""
attributes = {}

usage_fields = {
SpanAttributes.LLM_USAGE_PROMPT_TOKENS: "input_tokens",
SpanAttributes.LLM_USAGE_COMPLETION_TOKENS: "output_tokens",
SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS: "cache_read_input_tokens",
SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS: "cache_creation_input_tokens",
}
for attribute, field in usage_fields.items():
value = getattr(usage, field, None)
if value is not None:
attributes[attribute] = value

input_tokens = getattr(usage, "input_tokens", None)
output_tokens = getattr(usage, "output_tokens", None)
if input_tokens is not None and output_tokens is not None:
attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS] = input_tokens + output_tokens

return attributes


def get_message_attributes(
args: Optional[Tuple] = None, kwargs: Optional[Dict] = None, return_value: Any = None
) -> AttributeMap:
Expand Down Expand Up @@ -318,18 +347,7 @@ def get_message_response_attributes(response: "Message") -> AttributeMap:

# Extract usage information
if hasattr(response, "usage"):
usage = response.usage
if hasattr(usage, "input_tokens"):
input_tokens = usage.input_tokens
attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS] = input_tokens

if hasattr(usage, "output_tokens"):
output_tokens = usage.output_tokens
attributes[SpanAttributes.LLM_USAGE_COMPLETION_TOKENS] = output_tokens

if hasattr(usage, "input_tokens") and hasattr(usage, "output_tokens"):
total_tokens = usage.input_tokens + usage.output_tokens
attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS] = total_tokens
attributes.update(get_usage_attributes(response.usage))

# Extract stop reason if available
if hasattr(response, "stop_reason"):
Expand Down Expand Up @@ -435,19 +453,7 @@ def get_completion_response_attributes(response: "Completion") -> AttributeMap:

# Extract usage information (newer versions have this)
if hasattr(response, "usage"):
usage = response.usage
if hasattr(usage, "input_tokens"):
input_tokens = usage.input_tokens
attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS] = input_tokens

if hasattr(usage, "output_tokens"):
output_tokens = usage.output_tokens
attributes[SpanAttributes.LLM_USAGE_COMPLETION_TOKENS] = output_tokens

# Calculate total tokens if we have both input and output
if hasattr(usage, "input_tokens") and hasattr(usage, "output_tokens"):
total_tokens = usage.input_tokens + usage.output_tokens
attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS] = total_tokens
attributes.update(get_usage_attributes(response.usage))

return attributes

Expand Down Expand Up @@ -518,18 +524,7 @@ def get_stream_event_attributes(event: Any) -> AttributeMap:
elif event_type == "RawMessageStartEvent":
if hasattr(event, "message"):
if hasattr(event.message, "usage"):
usage = event.message.usage
if hasattr(usage, "input_tokens"):
input_tokens = usage.input_tokens
attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS] = input_tokens

if hasattr(usage, "output_tokens"):
output_tokens = usage.output_tokens
attributes[SpanAttributes.LLM_USAGE_COMPLETION_TOKENS] = output_tokens

if hasattr(usage, "input_tokens") and hasattr(usage, "output_tokens"):
total_tokens = usage.input_tokens + usage.output_tokens
attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS] = total_tokens
attributes.update(get_usage_attributes(event.message.usage))

elif event_type == "RawMessageDeltaEvent":
if hasattr(event, "delta"):
Expand Down
27 changes: 5 additions & 22 deletions agentops/instrumentation/providers/anthropic/stream_wrapper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@
from agentops.instrumentation.providers.anthropic.attributes.message import (
get_message_request_attributes,
get_stream_attributes,
get_usage_attributes,
)
from agentops.instrumentation.providers.anthropic.event_handler_wrapper import EventHandleWrapper

Expand Down Expand Up @@ -165,16 +166,8 @@ def __exit__(self, exc_type, exc_val, exc_tb):
span.set_attribute(MessageAttributes.COMPLETION_CONTENT.format(i=0), content_text)

if hasattr(final_message, "usage"):
usage = final_message.usage
if hasattr(usage, "input_tokens"):
span.set_attribute(SpanAttributes.LLM_USAGE_PROMPT_TOKENS, usage.input_tokens)

if hasattr(usage, "output_tokens"):
span.set_attribute(SpanAttributes.LLM_USAGE_COMPLETION_TOKENS, usage.output_tokens)

if hasattr(usage, "input_tokens") and hasattr(usage, "output_tokens"):
total_tokens = usage.input_tokens + usage.output_tokens
span.set_attribute(SpanAttributes.LLM_USAGE_TOTAL_TOKENS, total_tokens)
for key, value in get_usage_attributes(final_message.usage).items():
span.set_attribute(key, value)
except Exception as e:
logger.debug(f"Failed to extract final message data: {e}")
finally:
Expand Down Expand Up @@ -404,18 +397,8 @@ async def __aexit__(self, exc_type, exc_val, exc_tb):
)

if hasattr(final_message, "usage"):
usage = final_message.usage
if hasattr(usage, "input_tokens"):
span.set_attribute(SpanAttributes.LLM_USAGE_PROMPT_TOKENS, usage.input_tokens)

if hasattr(usage, "output_tokens"):
span.set_attribute(
SpanAttributes.LLM_USAGE_COMPLETION_TOKENS, usage.output_tokens
)

if hasattr(usage, "input_tokens") and hasattr(usage, "output_tokens"):
total_tokens = usage.input_tokens + usage.output_tokens
span.set_attribute(SpanAttributes.LLM_USAGE_TOTAL_TOKENS, total_tokens)
for key, value in get_usage_attributes(final_message.usage).items():
span.set_attribute(key, value)
except Exception as e:
logger.debug(f"Failed to extract final async message data: {e}")
finally:
Expand Down
30 changes: 30 additions & 0 deletions tests/unit/instrumentation/anthropic/test_attributes.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
"""Tests for Anthropic attribute extraction functionality."""

from types import SimpleNamespace

from agentops.semconv import (
InstrumentationAttributes,
SpanAttributes,
Expand All @@ -12,6 +14,7 @@
extract_request_attributes,
)
from agentops.instrumentation.providers.anthropic.attributes.message import (
get_message_response_attributes,
get_message_request_attributes,
get_stream_attributes,
get_stream_event_attributes,
Expand Down Expand Up @@ -72,6 +75,29 @@ def test_get_message_request_attributes():
assert MessageAttributes.PROMPT_CONTENT.format(i=1) in attributes


def test_get_message_response_attributes_exposes_cache_usage_separately():
"""Anthropic cache usage is reported without changing native token totals."""
response = SimpleNamespace(
id="msg_cache",
model="claude-sonnet",
content=[],
usage=SimpleNamespace(
input_tokens=100,
output_tokens=20,
cache_read_input_tokens=500,
cache_creation_input_tokens=50,
),
)

attributes = get_message_response_attributes(response)

assert attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS] == 100
assert attributes[SpanAttributes.LLM_USAGE_COMPLETION_TOKENS] == 20
assert attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS] == 120
assert attributes[SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS] == 500
assert attributes[SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS] == 50


# Stream Attributes Tests
def test_get_stream_attributes():
"""Test extraction of stream attributes."""
Expand Down Expand Up @@ -108,6 +134,8 @@ class MockUsage:
def __init__(self):
self.input_tokens = 10
self.output_tokens = 5
self.cache_read_input_tokens = 40
self.cache_creation_input_tokens = 30

class MockMessage:
def __init__(self):
Expand All @@ -124,6 +152,8 @@ def __init__(self):
assert attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS] == 10
assert attributes[SpanAttributes.LLM_USAGE_COMPLETION_TOKENS] == 5
assert attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS] == 15
assert attributes[SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS] == 40
assert attributes[SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS] == 30


def test_get_stream_event_attributes_raw_message_delta():
Expand Down
42 changes: 41 additions & 1 deletion tests/unit/instrumentation/anthropic/test_stream_wrapper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
import pytest
from types import SimpleNamespace
from unittest.mock import MagicMock
from opentelemetry.trace import SpanKind

Expand Down Expand Up @@ -108,7 +109,12 @@ def test_stream_final_message_attributes(mock_tracer, mock_stream_manager):

final_message = MagicMock()
final_message.content = [MagicMock(text="Final response")]
final_message.usage = MagicMock(input_tokens=10, output_tokens=20)
final_message.usage = SimpleNamespace(
input_tokens=10,
output_tokens=20,
cache_read_input_tokens=40,
cache_creation_input_tokens=30,
)
mock_stream_manager._MessageStreamManager__stream._MessageStream__final_message_snapshot = final_message

result = wrapper(wrapped, None, [], {})
Expand All @@ -123,3 +129,37 @@ def test_stream_final_message_attributes(mock_tracer, mock_stream_manager):
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_PROMPT_TOKENS, 10)
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_COMPLETION_TOKENS, 20)
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_TOTAL_TOKENS, 30)
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS, 40)
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS, 30)


@pytest.mark.asyncio
async def test_async_stream_final_message_attributes_include_cache_usage(mock_tracer, mock_async_stream_manager):
"""Async stream final usage includes both Anthropic cache token fields."""
wrapper = messages_stream_async_wrapper(mock_tracer)
wrapped = MagicMock(return_value=mock_async_stream_manager)
final_message = SimpleNamespace(
content=[SimpleNamespace(text="Final response")],
usage=SimpleNamespace(
input_tokens=10,
output_tokens=20,
cache_read_input_tokens=40,
cache_creation_input_tokens=30,
),
)
mock_async_stream_manager._AsyncMessageStreamManager__stream._AsyncMessageStream__final_message_snapshot = (
final_message
)

result = wrapper(wrapped, None, [], {})

async with result as stream:
async for _ in stream.text_stream:
pass

span = mock_tracer.start_span.return_value
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_PROMPT_TOKENS, 10)
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_COMPLETION_TOKENS, 20)
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_TOTAL_TOKENS, 30)
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS, 40)
span.set_attribute.assert_any_call(SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS, 30)