wip

Author: shellmayr

sentry_sdk/ai/utils.py

@@ -112,22 +112,51 @@ def truncate_messages_by_size(messages, max_bytes=MAX_GEN_AI_MESSAGE_BYTES):
     if not messages:
         return messages
 
-    # make a list out of the messages in case it's just a string? why is this needed?
     truncated_messages = list(messages)
 
-    # while there is more than one message, serialize and measure the size, and if it's too big, remove the oldest message
     while len(truncated_messages) > 1:
-        serialized = serialize(
-            truncated_messages, is_vars=False, max_value_length=round(max_bytes * 0.8)
-        )
-        serialized_json = json.dumps(serialized, separators=(",", ":"))
+        serialized_json = json.dumps(truncated_messages, separators=(",", ":"))
         current_size = len(serialized_json.encode("utf-8"))
 
         if current_size <= max_bytes:
             break
 
         truncated_messages.pop(0)
 
+    serialized_json = json.dumps(truncated_messages, separators=(",", ":"))
+    current_size = len(serialized_json.encode("utf-8"))
+
+    if current_size > max_bytes and len(truncated_messages) == 1:
+        message = truncated_messages[0].copy()
+        content = message.get("content", "")
+
+        if isinstance(content, str):
+            max_content_length = max_bytes // 2
+            while True:
+                message["content"] = content[:max_content_length]
+                test_json = json.dumps([message], separators=(",", ":"))
+                if len(test_json.encode("utf-8")) <= max_bytes:
+                    break
+                max_content_length = int(max_content_length * 0.9)
+                if max_content_length < 100:
+                    message["content"] = ""
+                    break
+
+            truncated_messages = [message]
+        elif isinstance(content, list):
+            content_copy = list(content)
+            while len(content_copy) > 0:
+                message["content"] = content_copy
+                test_json = json.dumps([message], separators=(",", ":"))
+                if len(test_json.encode("utf-8")) <= max_bytes:
+                    break
+                content_copy = content_copy[:-1]
+
+            if len(content_copy) == 0:
+                message["content"] = []
+
+            truncated_messages = [message]
+
     return truncated_messages
 
 

sentry_sdk/integrations/openai_agents/utils.py

@@ -142,7 +142,9 @@ def _set_input_data(span, get_response_kwargs):
         role_normalized_messages, span, scope
     )
     if messages_data is not None:
-        span.set_data(SPANDATA.GEN_AI_REQUEST_MESSAGES, messages_data)
+        set_data_normalized(
+            span, SPANDATA.GEN_AI_REQUEST_MESSAGES, messages_data, unpack=False
+        )
 
 
 def _set_output_data(span, result):

tests/integrations/openai_agents/test_openai_agents.py

@@ -143,7 +143,10 @@ async def test_agent_invocation_span(
     assert transaction["contexts"]["trace"]["origin"] == "auto.ai.openai_agents"
 
     assert invoke_agent_span["description"] == "invoke_agent test_agent"
-    assert invoke_agent_span["data"]["gen_ai.request.messages"] == [
+    messages_data = invoke_agent_span["data"]["gen_ai.request.messages"]
+    assert isinstance(messages_data, str)
+    parsed_messages = json.loads(messages_data)
+    assert parsed_messages == [
         {
             "content": [{"text": "You are a helpful test assistant.", "type": "text"}],
             "role": "system",
@@ -484,7 +487,10 @@ def simple_test_tool(message: str) -> str:
     assert ai_client_span1["data"]["gen_ai.agent.name"] == "test_agent"
     assert ai_client_span1["data"]["gen_ai.request.available_tools"] == available_tools
     assert ai_client_span1["data"]["gen_ai.request.max_tokens"] == 100
-    assert ai_client_span1["data"]["gen_ai.request.messages"] == [
+    messages_data = ai_client_span1["data"]["gen_ai.request.messages"]
+    assert isinstance(messages_data, str)
+    parsed_messages = json.loads(messages_data)
+    assert parsed_messages == [
         {
             "role": "system",
             "content": [{"type": "text", "text": "You are a helpful test assistant."}],