Skip to content

Python SDK for AI agent monitoring, LLM cost tracking, benchmarking, and more. Integrates with most LLMs and agent frameworks including CrewAI, Agno, OpenAI Agents SDK, Langchain, Autogen, AG2, and CamelAI

fix: anthropic provider on raw message (#507)

teocns7e1c32f61233
diff --git a/agentops/llms/anthropic.py b/agentops/llms/anthropic.py
index bc63d85c..0ec1c184 100644
--- a/agentops/llms/anthropic.py
+++ b/agentops/llms/anthropic.py
@@ -1,3 +1,4 @@
+import json
import pprint
from typing import Optional
@@ -5,9 +6,9 @@ from agentops.llms.instrumented_provider import InstrumentedProvider
from agentops.time_travel import fetch_completion_override_from_time_travel_cache
from ..event import ErrorEvent, LLMEvent, ToolEvent
-from ..session import Session
-from ..log_config import logger
from ..helpers import check_call_stack_for_agent_id, get_ISO_time
+from ..log_config import logger
+from ..session import Session
from ..singleton import singleton
@@ -24,9 +25,9 @@ class AnthropicProvider(InstrumentedProvider):
def handle_response(self, response, kwargs, init_timestamp, session: Optional[Session] = None):
"""Handle responses for Anthropic"""
- from anthropic import Stream, AsyncStream
- from anthropic.resources import AsyncMessages
import anthropic.resources.beta.messages.messages as beta_messages
+ from anthropic import AsyncStream, Stream
+ from anthropic.resources import AsyncMessages
from anthropic.types import Message
llm_event = LLMEvent(init_timestamp=init_timestamp, params=kwargs)
@@ -118,17 +119,63 @@ class AnthropicProvider(InstrumentedProvider):
# Handle object responses
try:
- llm_event.returns = response.model_dump()
- llm_event.agent_id = check_call_stack_for_agent_id()
- llm_event.prompt = kwargs["messages"]
- llm_event.prompt_tokens = response.usage.input_tokens
- llm_event.completion = {
- "role": "assistant",
- "content": response.content[0].text,
- }
- llm_event.completion_tokens = response.usage.output_tokens
- llm_event.model = response.model
+ # Naively handle AttributeError("'LegacyAPIResponse' object has no attribute 'model_dump'")
+ if hasattr(response, "model_dump"):
+ # This bets on the fact that the response object has a model_dump method
+ llm_event.returns = response.model_dump()
+ llm_event.prompt_tokens = response.usage.input_tokens
+ llm_event.completion_tokens = response.usage.output_tokens
+
+ llm_event.completion = {
+ "role": "assistant",
+ "content": response.content[0].text,
+ }
+ llm_event.model = response.model
+
+ else:
+ """Handle raw response data from the Anthropic API.
+
+ The raw response has the following structure:
+ {
+ 'id': str, # Message ID (e.g. 'msg_018Gk9N2pcWaYLS7mxXbPD5i')
+ 'type': str, # Type of response (e.g. 'message')
+ 'role': str, # Role of responder (e.g. 'assistant')
+ 'model': str, # Model used (e.g. 'claude-3-5-sonnet-20241022')
+ 'content': List[Dict], # List of content blocks with 'type' and 'text'
+ 'stop_reason': str, # Reason for stopping (e.g. 'end_turn')
+ 'stop_sequence': Any, # Stop sequence used, if any
+ 'usage': { # Token usage statistics
+ 'input_tokens': int,
+ 'output_tokens': int
+ }
+ }
+
+ Note: We import Anthropic types here since the package must be installed
+ for raw responses to be available; doing so in the global scope would
+ result in dependencies error since this provider is not lazily imported (tests fail)
+ """
+ from anthropic import APIResponse
+ from anthropic._legacy_response import LegacyAPIResponse
+
+ assert isinstance(response, (APIResponse, LegacyAPIResponse)), (
+ f"Expected APIResponse or LegacyAPIResponse, got {type(response)}. "
+ "This is likely caused by changes in the Anthropic SDK and the integrations with AgentOps needs update."
+ "Please open an issue at https://github.com/AgentOps-AI/agentops/issues"
+ )
+ response_data = json.loads(response.text)
+ llm_event.returns = response_data
+ llm_event.model = response_data["model"]
+ llm_event.completion = {
+ "role": response_data.get("role"),
+ "content": response_data.get("content")[0].get("text") if response_data.get("content") else "",
+ }
+ if usage := response_data.get("usage"):
+ llm_event.prompt_tokens = usage.get("input_tokens")
+ llm_event.completion_tokens = usage.get("output_tokens")
+
llm_event.end_timestamp = get_ISO_time()
+ llm_event.prompt = kwargs["messages"]
+ llm_event.agent_id = check_call_stack_for_agent_id()
self._safe_record(session, llm_event)
except Exception as e:
@@ -148,8 +195,8 @@ class AnthropicProvider(InstrumentedProvider):
self._override_async_completion()
def _override_completion(self):
- from anthropic.resources import messages
import anthropic.resources.beta.messages.messages as beta_messages
+ from anthropic.resources import messages
from anthropic.types import (
Message,
RawContentBlockDeltaEvent,
@@ -168,6 +215,7 @@ class AnthropicProvider(InstrumentedProvider):
def patched_function(*args, **kwargs):
init_timestamp = get_ISO_time()
session = kwargs.get("session", None)
+
if "session" in kwargs.keys():
del kwargs["session"]
@@ -212,6 +260,7 @@ class AnthropicProvider(InstrumentedProvider):
beta_messages.Messages.create = create_patched_function(is_beta=True)
def _override_async_completion(self):
+ import anthropic.resources.beta.messages.messages as beta_messages
from anthropic.resources import messages
from anthropic.types import (
Message,
@@ -222,7 +271,6 @@ class AnthropicProvider(InstrumentedProvider):
RawMessageStartEvent,
RawMessageStopEvent,
)
- import anthropic.resources.beta.messages.messages as beta_messages
# Store the original method
self.original_create_async = messages.AsyncMessages.create