feat(langchain): Set system instruction attribute

sentry_sdk/consts.py

@@ -542,6 +542,12 @@ class SPANDATA:
     Example: 2048
     """
 
+    GEN_AI_SYSTEM_INSTRUCTIONS = "gen_ai.system_instructions"
+    """
+    The system instructions passed to the model.
+    Example: [{"type": "text", "text": "You are a helpful assistant."},{"type": "text", "text": "Be concise and clear."}]
+    """
+
     GEN_AI_REQUEST_MESSAGES = "gen_ai.request.messages"
     """
     The messages passed to the model. The "content" can be a string or an array of objects.

sentry_sdk/integrations/langchain.py

@@ -189,6 +189,22 @@ def _get_current_agent() -> "Optional[str]":
     return None
 
 
+def _set_system_prompt(
+    span: "sentry_sdk.tracing.Span", messages: "List[List[BaseMessage]]"
+) -> None:
+    for list_ in messages:
+        for message in list_:
+            if message.type == "system":
+                system_prompt = message.content
+                set_data_normalized(
+                    span,
+                    SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS,
+                    system_prompt,
+                    unpack=False,
+                )
+                return
+
+
 class LangchainIntegration(Integration):
     identifier = "langchain"
     origin = f"auto.ai.{identifier}"
@@ -430,9 +446,14 @@ def on_chat_model_start(
             _set_tools_on_span(span, all_params.get("tools"))
 
             if should_send_default_pii() and self.include_prompts:
+                _set_system_prompt(span, messages)
+
                 normalized_messages = []
                 for list_ in messages:
                     for message in list_:
+                        if message.type == "system":
+                            continue
+
                         normalized_messages.append(
                             self._normalize_langchain_message(message)
                         )

tests/integrations/langchain/test_langchain.py

@@ -219,14 +219,14 @@ def test_langchain_agent(
     if send_default_pii and include_prompts:
         assert (
             "You are very powerful"
-            in chat_spans[0]["data"][SPANDATA.GEN_AI_REQUEST_MESSAGES]
+            in chat_spans[0]["data"][SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS]
         )
         assert "5" in chat_spans[0]["data"][SPANDATA.GEN_AI_RESPONSE_TEXT]
         assert "word" in tool_exec_span["data"][SPANDATA.GEN_AI_TOOL_INPUT]
         assert 5 == int(tool_exec_span["data"][SPANDATA.GEN_AI_TOOL_OUTPUT])
         assert (
             "You are very powerful"
-            in chat_spans[1]["data"][SPANDATA.GEN_AI_REQUEST_MESSAGES]
+            in chat_spans[1]["data"][SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS]
         )
         assert "5" in chat_spans[1]["data"][SPANDATA.GEN_AI_RESPONSE_TEXT]
 
@@ -243,8 +243,10 @@ def test_langchain_agent(
             tool_call_str = str(tool_calls_data)
             assert "get_word_length" in tool_call_str
     else:
+        assert SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS not in chat_spans[0].get("data", {})
         assert SPANDATA.GEN_AI_REQUEST_MESSAGES not in chat_spans[0].get("data", {})
         assert SPANDATA.GEN_AI_RESPONSE_TEXT not in chat_spans[0].get("data", {})
+        assert SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS not in chat_spans[1].get("data", {})
         assert SPANDATA.GEN_AI_REQUEST_MESSAGES not in chat_spans[1].get("data", {})
         assert SPANDATA.GEN_AI_RESPONSE_TEXT not in chat_spans[1].get("data", {})
         assert SPANDATA.GEN_AI_TOOL_INPUT not in tool_exec_span.get("data", {})