fix(anthropic): support with_raw_response wrapper for span generation

packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/__init__.py

@@ -130,8 +130,8 @@ async def _aset_token_usage(
     token_histogram: Histogram = None,
     choice_counter: Counter = None,
 ):
-    if not isinstance(response, dict):
-        response = response.__dict__
+    from opentelemetry.instrumentation.anthropic.utils import _extract_response_data
+    response = _extract_response_data(response)
 
     if usage := response.get("usage"):
         prompt_tokens = usage.input_tokens
@@ -223,8 +223,8 @@ def _set_token_usage(
     token_histogram: Histogram = None,
     choice_counter: Counter = None,
 ):
-    if not isinstance(response, dict):
-        response = response.__dict__
+    from opentelemetry.instrumentation.anthropic.utils import _extract_response_data
+    response = _extract_response_data(response)
 
     if usage := response.get("usage"):
         prompt_tokens = usage.input_tokens

packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/span_utils.py

@@ -8,6 +8,7 @@
     dont_throw,
     model_as_dict,
     should_send_prompts,
+    _extract_response_data,
 )
 from opentelemetry.semconv._incubating.attributes.gen_ai_attributes import (
     GEN_AI_RESPONSE_ID,
@@ -170,6 +171,7 @@ def _set_span_completions(span, response):
         return
     from opentelemetry.instrumentation.anthropic import set_span_attribute
 
+    response = _extract_response_data(response)
     index = 0
     prefix = f"{SpanAttributes.LLM_COMPLETIONS}.{index}"
     set_span_attribute(span, f"{prefix}.finish_reason", response.get("stop_reason"))
@@ -236,8 +238,7 @@ def _set_span_completions(span, response):
 def set_response_attributes(span, response):
     from opentelemetry.instrumentation.anthropic import set_span_attribute
 
-    if not isinstance(response, dict):
-        response = response.__dict__
+    response = _extract_response_data(response)
     set_span_attribute(span, SpanAttributes.LLM_RESPONSE_MODEL, response.get("model"))
     set_span_attribute(span, GEN_AI_RESPONSE_ID, response.get("id"))
 

packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/utils.py

@@ -61,10 +61,32 @@ def _handle_exception(e, func, logger):
     return async_wrapper if asyncio.iscoroutinefunction(func) else sync_wrapper
 
 
+def _extract_response_data(response):
+    """Extract the actual response data from both regular and with_raw_response wrapped responses."""
+    if isinstance(response, dict):
+        return response
+
+    # Handle with_raw_response wrapped responses
+    if hasattr(response, 'parse') and callable(response.parse):
+        try:
+            # For with_raw_response, parse() gives us the actual response object
+            parsed_response = response.parse()
+            if not isinstance(parsed_response, dict):
+                parsed_response = parsed_response.__dict__
+            return parsed_response
+        except Exception:
+            pass
+
+    # Fallback to __dict__ for regular response objects
+    if hasattr(response, '__dict__'):
+        return response.__dict__
+
+    return {}
+
+
 @dont_throw
 def shared_metrics_attributes(response):
-    if not isinstance(response, dict):
-        response = response.__dict__
+    response = _extract_response_data(response)
 
     common_attributes = Config.get_common_metrics_attributes()
 

packages/opentelemetry-instrumentation-anthropic/pyproject.toml

@@ -36,7 +36,7 @@ pytest = "^8.2.2"
 pytest-sugar = "1.0.0"
 
 [tool.poetry.group.test.dependencies]
-anthropic = ">=0.36.0"
+anthropic = {extras = ["bedrock"], version = ">=0.36.0"}
 pytest = "^8.2.2"
 pytest-sugar = "1.0.0"
 vcrpy = "^6.0.1"