feat(google-genai): Set system instruction attribute

sentry_sdk/_types.py

@@ -359,3 +359,7 @@ class SDKInfo(TypedDict):
     )
 
     HttpStatusCodeRange = Union[int, Container[int]]
+
+    class TextPart(TypedDict):
+        type: Literal["text"]
+        content: str

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/google_genai/utils.py

@@ -32,13 +32,16 @@
     event_from_exception,
     safe_serialize,
 )
-from google.genai.types import GenerateContentConfig
+from google.genai.types import GenerateContentConfig, Part, Content
+from itertools import chain
 
 if TYPE_CHECKING:
     from sentry_sdk.tracing import Span
+    from sentry_sdk._types import TextPart
     from google.genai.types import (
         GenerateContentResponse,
         ContentListUnion,
+        ContentUnionDict,
         Tool,
         Model,
         EmbedContentResponse,
@@ -720,6 +723,56 @@ def extract_finish_reasons(
     return finish_reasons if finish_reasons else None
 
 
+def _transform_system_instruction_one_level(
+    system_instructions: "ContentUnionDict",
+) -> "list[TextPart]":
+    text_parts: "list[TextPart]" = []
+
+    if isinstance(system_instructions, str):
+        return [{"type": "text", "content": system_instructions}]
+
+    if isinstance(system_instructions, Part) and system_instructions.text:
+        return [{"type": "text", "content": system_instructions.text}]
+
+    if isinstance(system_instructions, Content):
+        for part in system_instructions.parts or []:
+            if part.text:
+                text_parts.append({"type": "text", "content": part.text})
+        return text_parts
+
+    if isinstance(system_instructions, dict):
+        if system_instructions.get("text"):
+            return [{"type": "text", "content": system_instructions["text"]}]
+
+        parts = system_instructions.get("parts", [])
+        for part in parts:
+            if isinstance(part, Part) and part.text:
+                text_parts.append({"type": "text", "content": part.text})
+            elif isinstance(part, dict) and part.get("text"):
+                text_parts.append({"type": "text", "content": part["text"]})
+        return text_parts
+
+    return text_parts
+
+
+def _transform_system_instructions(
+    system_instructions: "ContentUnionDict",
+) -> "list[TextPart]":
+    text_parts: "list[TextPart]" = []
+
+    if isinstance(system_instructions, list):
+        text_parts = list(
+            chain.from_iterable(
+                _transform_system_instruction_one_level(instructions)
+                for instructions in system_instructions
+            )
+        )
+
+        return text_parts
+
+    return _transform_system_instruction_one_level(system_instructions)
+
+
 def set_span_data_for_request(
     span: "Span",
     integration: "Any",
@@ -743,25 +796,14 @@ def set_span_data_for_request(
         # Add system instruction if present
         if config and hasattr(config, "system_instruction"):
             system_instruction = config.system_instruction
-            if system_instruction:
-                system_messages = extract_contents_messages(system_instruction)
-                # System instruction should be a single system message
-                # Extract text from all messages and combine into one system message
-                system_texts = []
-                for msg in system_messages:
-                    content = msg.get("content")
-                    if isinstance(content, list):
-                        # Extract text from content parts
-                        for part in content:
-                            if isinstance(part, dict) and part.get("type") == "text":
-                                system_texts.append(part.get("text", ""))
-                    elif isinstance(content, str):
-                        system_texts.append(content)
-
-                if system_texts:
-                    messages.append(
-                        {"role": "system", "content": " ".join(system_texts)}
-                    )
+
+            if system_instruction is not None:
+                set_data_normalized(
+                    span,
+                    SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS,
+                    _transform_system_instructions(system_instruction),
+                    unpack=False,
+                )
 
         # Extract messages from contents
         contents_messages = extract_contents_messages(contents)

tests/integrations/google_genai/test_google_genai.py

@@ -4,6 +4,7 @@
 
 from google import genai
 from google.genai import types as genai_types
+from google.genai.types import Content, Part
 
 from sentry_sdk import start_transaction
 from sentry_sdk._types import BLOB_DATA_SUBSTITUTE
@@ -186,6 +187,7 @@ def test_nonstreaming_generate_content(
         response_texts = json.loads(response_text)
         assert response_texts == ["Hello! How can I help you today?"]
     else:
+        assert SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS not in invoke_span["data"]
         assert SPANDATA.GEN_AI_REQUEST_MESSAGES not in invoke_span["data"]
         assert SPANDATA.GEN_AI_RESPONSE_TEXT not in chat_span["data"]
 
@@ -202,8 +204,107 @@ def test_nonstreaming_generate_content(
     assert invoke_span["data"][SPANDATA.GEN_AI_REQUEST_MAX_TOKENS] == 100
 
 
+# Threatened bot to generate as many cases as possible
+@pytest.mark.parametrize(
+    "system_instructions,expected_texts",
+    [
+        ("You are a helpful assistant", ["You are a helpful assistant"]),
+        (
+            ["You are a translator", "Translate to French"],
+            ["You are a translator", "Translate to French"],
+        ),
+        (
+            Content(role="user", parts=[Part(text="You are a helpful assistant")]),
+            ["You are a helpful assistant"],
+        ),
+        (
+            Content(
+                role="user",
+                parts=[
+                    Part(text="You are a translator"),
+                    Part(text="Translate to French"),
+                ],
+            ),
+            ["You are a translator", "Translate to French"],
+        ),
+        (
+            {"parts": [{"text": "You are a helpful assistant"}], "role": "user"},
+            ["You are a helpful assistant"],
+        ),
+        (
+            {
+                "parts": [
+                    {"text": "You are a translator"},
+                    {"text": "Translate to French"},
+                ],
+                "role": "user",
+            },
+            ["You are a translator", "Translate to French"],
+        ),
+        (Part(text="You are a helpful assistant"), ["You are a helpful assistant"]),
+        ({"text": "You are a helpful assistant"}, ["You are a helpful assistant"]),
+        (
+            [Part(text="You are a translator"), Part(text="Translate to French")],
+            ["You are a translator", "Translate to French"],
+        ),
+        (
+            [{"text": "You are a translator"}, {"text": "Translate to French"}],
+            ["You are a translator", "Translate to French"],
+        ),
+        (
+            [Part(text="First instruction"), {"text": "Second instruction"}],
+            ["First instruction", "Second instruction"],
+        ),
+        (
+            {
+                "parts": [
+                    Part(text="First instruction"),
+                    {"text": "Second instruction"},
+                ],
+                "role": "user",
+            },
+            ["First instruction", "Second instruction"],
+        ),
+        (None, None),
+        ("", []),
+        ({}, []),
+        ({"parts": []}, []),
+        (Content(role="user", parts=[]), []),
+        (
+            {
+                "parts": [
+                    {"text": "Text part"},
+                    {"file_data": {"file_uri": "gs://bucket/file.pdf"}},
+                ],
+                "role": "user",
+            },
+            ["Text part"],
+        ),
+        (
+            {
+                "parts": [
+                    {"text": "First"},
+                    Part(text="Second"),
+                    {"text": "Third"},
+                ],
+                "role": "user",
+            },
+            ["First", "Second", "Third"],
+        ),
+        (
+            {
+                "parts": [
+                    Part(text="First"),
+                    Part(text="Second"),
+                    Part(text="Third"),
+                ],
+            },
+            ["First", "Second", "Third"],
+        ),
+    ],
+)
 def test_generate_content_with_system_instruction(
-    sentry_init, capture_events, mock_genai_client
+    sentry_init, capture_events, mock_genai_client, system_instructions, expected_texts
 ):
     sentry_init(
         integrations=[GoogleGenAIIntegration(include_prompts=True)],
@@ -219,7 +320,7 @@ def test_generate_content_with_system_instruction(
     ):
         with start_transaction(name="google_genai"):
             config = create_test_config(
-                system_instruction="You are a helpful assistant",
+                system_instruction=system_instructions,
                 temperature=0.5,
             )
             mock_genai_client.models.generate_content(
@@ -229,14 +330,21 @@ def test_generate_content_with_system_instruction(
     (event,) = events
     invoke_span = event["spans"][0]
 
-    # Check that system instruction is included in messages
+    if expected_texts is None:
+        assert SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS not in invoke_span["data"]
+        return
+
     # (PII is enabled and include_prompts is True in this test)
-    messages_str = invoke_span["data"][SPANDATA.GEN_AI_REQUEST_MESSAGES]
-    # Parse the JSON string to verify content
-    messages = json.loads(messages_str)
-    assert len(messages) == 2
-    assert messages[0] == {"role": "system", "content": "You are a helpful assistant"}
-    assert messages[1] == {"role": "user", "content": "What is 2+2?"}
+    system_instructions = json.loads(
+        invoke_span["data"][SPANDATA.GEN_AI_SYSTEM_INSTRUCTIONS]
+    )
+
+    if isinstance(expected_texts, str):
+        assert system_instructions == [{"type": "text", "content": expected_texts}]
+    else:
+        assert system_instructions == [
+            {"type": "text", "content": text} for text in expected_texts
+        ]
 
 
 def test_generate_content_with_tools(sentry_init, capture_events, mock_genai_client):
@@ -933,10 +1041,8 @@ def test_google_genai_message_truncation(
         with start_transaction(name="google_genai"):
             mock_genai_client.models.generate_content(
                 model="gemini-1.5-flash",
-                contents=small_content,
-                config=create_test_config(
-                    system_instruction=large_content,
-                ),
+                contents=[large_content, small_content],
+                config=create_test_config(),
             )
 
     (event,) = events