Add safety attribution metadata to provider hooks

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

@@ -3,6 +3,7 @@
 from opentelemetry.instrumentation.fortifyroot import (
     SafetyDecision,
     SafetyLocation,
+    build_safety_metadata,
     clone_value,
     get_object_value,
     run_completion_safety,
@@ -17,6 +18,7 @@
 def _apply_prompt_safety(span, kwargs, span_name: str):
     try:
         request_type = _request_type(span_name)
+        request_model = kwargs.get("model")
         mutated_kwargs = kwargs
 
         prompt = kwargs.get("prompt")
@@ -28,6 +30,7 @@ def _apply_prompt_safety(span, kwargs, span_name: str):
                 request_type=request_type,
                 segment_index=0,
                 segment_role="user",
+                request_model=request_model,
             )
             if changed:
                 mutated_kwargs = dict(kwargs)
@@ -41,6 +44,7 @@ def _apply_prompt_safety(span, kwargs, span_name: str):
             request_type=request_type,
             segment_index=0,
             segment_role="system",
+            request_model=request_model,
         )
         if system_changed:
             if mutated_kwargs is kwargs:
@@ -62,6 +66,7 @@ def _apply_prompt_safety(span, kwargs, span_name: str):
                 request_type=request_type,
                 segment_index=index,
                 segment_role=role,
+                request_model=request_model,
             )
             if not changed:
                 continue
@@ -85,6 +90,7 @@ def _mask_prompt_content(
     request_type,
     segment_index,
     segment_role,
+    request_model=None,
 ):
     if isinstance(content, str):
         return _mask_prompt_text(
@@ -94,6 +100,7 @@ def _mask_prompt_content(
             request_type=request_type,
             segment_index=segment_index,
             segment_role=segment_role,
+            request_model=request_model,
         )
 
     if not isinstance(content, list):
@@ -114,6 +121,7 @@ def _mask_prompt_content(
             segment_index=segment_index,
             segment_role=segment_role,
             metadata={"block_index": block_index},
+            request_model=request_model,
         )
         if not changed:
             continue
@@ -127,6 +135,7 @@ def _mask_prompt_content(
 def _apply_completion_safety(span, response, span_name: str):
     try:
         request_type = _request_type(span_name)
+        response_model = get_object_value(response, "model")
 
         completion = get_object_value(response, "completion")
         if isinstance(completion, str):
@@ -137,6 +146,7 @@ def _apply_completion_safety(span, response, span_name: str):
                 request_type=request_type,
                 segment_index=0,
                 segment_role="assistant",
+                response_model=response_model,
             )
             if changed:
                 set_object_value(response, "completion", updated_completion)
@@ -166,6 +176,7 @@ def _apply_completion_safety(span, response, span_name: str):
                 request_type=request_type,
                 segment_index=index,
                 segment_role=role,
+                response_model=response_model,
             )
             if changed:
                 set_object_value(block, text_key, updated_text)
@@ -182,7 +193,13 @@ def _mask_prompt_text(
     segment_index,
     segment_role,
     metadata=None,
+    request_model=None,
 ):
+    metadata = build_safety_metadata(
+        metadata,
+        provider=PROVIDER,
+        request_model=request_model,
+    )
     result = run_prompt_safety(
         span=span,
         provider=PROVIDER,
@@ -205,7 +222,14 @@ def _mask_completion_text(
     request_type,
     segment_index,
     segment_role,
+    metadata=None,
+    response_model=None,
 ):
+    metadata = build_safety_metadata(
+        metadata,
+        provider=PROVIDER,
+        response_model=response_model,
+    )
     result = run_completion_safety(
         span=span,
         provider=PROVIDER,
@@ -215,6 +239,7 @@ def _mask_completion_text(
         request_type=request_type,
         segment_index=segment_index,
         segment_role=segment_role,
+        metadata=metadata,
     )
     return _resolve_masked_text(text, result)
 

packages/opentelemetry-instrumentation-anthropic/tests/test_safety_unit.py

@@ -52,6 +52,21 @@ def test_apply_prompt_safety_masks_prompt_system_and_messages(monkeypatch):
     assert updated["messages"][0]["content"][0]["text"] == "masked:msg-secret"
 
 
+def test_apply_prompt_safety_passes_model_metadata(monkeypatch):
+    contexts = []
+
+    def _prompt(**kwargs):
+        contexts.append(kwargs)
+        return None
+
+    monkeypatch.setattr(safety, "run_prompt_safety", _prompt)
+    kwargs = {"model": "claude-3-5-sonnet", "messages": [{"role": "user", "content": "secret"}]}
+    safety._apply_prompt_safety(None, kwargs, "anthropic.chat")
+
+    assert contexts[0]["metadata"]["gen_ai.system"] == "Anthropic"
+    assert contexts[0]["metadata"]["gen_ai.request.model"] == "claude-3-5-sonnet"
+
+
 def test_apply_prompt_safety_returns_partial_update_when_messages_missing():
     monkeypatch = pytest.MonkeyPatch()
     monkeypatch.setattr(safety, "run_prompt_safety", lambda **kwargs: SafetyResult(text=f"masked:{kwargs['text']}", overall_action="MASK"))
@@ -80,6 +95,21 @@ def test_apply_completion_safety_masks_completion_and_content(monkeypatch):
     assert response.content[1]["thinking"] == "masked:thought-secret"
 
 
+def test_apply_completion_safety_passes_model_metadata(monkeypatch):
+    contexts = []
+
+    def _completion(**kwargs):
+        contexts.append(kwargs)
+        return None
+
+    monkeypatch.setattr(safety, "run_completion_safety", _completion)
+    response = SimpleNamespace(model="claude-3-5-sonnet", completion="secret")
+    safety._apply_completion_safety(None, response, "anthropic.chat")
+
+    assert contexts[0]["metadata"]["gen_ai.system"] == "Anthropic"
+    assert contexts[0]["metadata"]["gen_ai.response.model"] == "claude-3-5-sonnet"
+
+
 def test_anthropic_prompt_and_completion_helpers_cover_noop_branches(monkeypatch):
     monkeypatch.setattr(safety, "run_prompt_safety", lambda **kwargs: SafetyResult(text=kwargs["text"], overall_action="MASK"))
     updated, changed = safety._mask_prompt_content(

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

@@ -364,12 +364,17 @@ def stream_done(response_body):
 @dont_throw
 def _handle_call(span: Span, kwargs, response, metric_params, event_logger):
     request_body = json.loads(kwargs.get("body"))
-    response_body = _prepare_invoke_response(span, response, _BEDROCK_INVOKE_SPAN_NAME)
+    (provider, model_vendor, model) = _get_vendor_model(kwargs.get("modelId"))
+    response_body = _prepare_invoke_response(
+        span,
+        response,
+        _BEDROCK_INVOKE_SPAN_NAME,
+        response_model=model,
+    )
     headers = {}
     if "ResponseMetadata" in response:
         headers = response.get("ResponseMetadata").get("HTTPHeaders", {})
 
-    (provider, model_vendor, model) = _get_vendor_model(kwargs.get("modelId"))
     metric_params.vendor = provider
     metric_params.model = model
     metric_params.is_stream = False
@@ -400,8 +405,13 @@ def _handle_call(span: Span, kwargs, response, metric_params, event_logger):
 
 @dont_throw
 def _handle_converse(span, kwargs, response, metric_params, event_logger):
-    _apply_converse_completion_safety(span, response, _BEDROCK_CONVERSE_SPAN_NAME)
     (provider, model_vendor, model) = _get_vendor_model(kwargs.get("modelId"))
+    _apply_converse_completion_safety(
+        span,
+        response,
+        _BEDROCK_CONVERSE_SPAN_NAME,
+        response_model=model,
+    )
     guardrail_converse(span, response, provider, model, metric_params)
 
     set_converse_model_span_attributes(span, provider, model, kwargs)