feat(ai): Add original input length meta attribute

sentry_sdk/ai/_openai_completions_api.py

@@ -0,0 +1,48 @@
+from collections.abc import Iterable
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from sentry_sdk._types import TextPart
+
+    from openai.types.chat import (
+        ChatCompletionMessageParam,
+        ChatCompletionSystemMessageParam,
+    )
+
+
+def _is_system_instruction(message: "ChatCompletionMessageParam") -> bool:
+    return isinstance(message, dict) and message.get("role") == "system"
+
+
+def _get_system_instructions(
+    messages: "Iterable[ChatCompletionMessageParam]",
+) -> "list[ChatCompletionMessageParam]":
+    if not isinstance(messages, Iterable):
+        return []
+
+    return [message for message in messages if _is_system_instruction(message)]
+
+
+def _transform_system_instructions(
+    system_instructions: "list[ChatCompletionSystemMessageParam]",
+) -> "list[TextPart]":
+    instruction_text_parts: "list[TextPart]" = []
+
+    for instruction in system_instructions:
+        if not isinstance(instruction, dict):
+            continue
+
+        content = instruction.get("content")
+
+        if isinstance(content, str):
+            instruction_text_parts.append({"type": "text", "content": content})
+
+        elif isinstance(content, list):
+            for part in content:
+                if isinstance(part, dict) and part.get("type") == "text":
+                    text = part.get("text", "")
+                    if text:
+                        instruction_text_parts.append({"type": "text", "content": text})
+
+    return instruction_text_parts

sentry_sdk/ai/_openai_responses_api.py

@@ -0,0 +1,22 @@
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from typing import Union
+
+    from openai.types.responses import ResponseInputParam, ResponseInputItemParam
+
+
+def _is_system_instruction(message: "ResponseInputItemParam") -> bool:
+    if not isinstance(message, dict) or not message.get("role") == "system":
+        return False
+
+    return "type" not in message or message["type"] == "message"
+
+
+def _get_system_instructions(
+    messages: "Union[str, ResponseInputParam]",
+) -> "list[ResponseInputItemParam]":
+    if not isinstance(messages, list):
+        return []
+
+    return [message for message in messages if _is_system_instruction(message)]

sentry_sdk/ai/utils.py

@@ -14,6 +14,7 @@
 
 import sentry_sdk
 from sentry_sdk.utils import logger
+from sentry_sdk.consts import SPANDATA
 
 MAX_GEN_AI_MESSAGE_BYTES = 20_000  # 20KB
 # Maximum characters when only a single message is left after bytes truncation
@@ -682,6 +683,28 @@ def truncate_messages_by_size(
 
 
 def truncate_and_annotate_messages(
+    messages: "Optional[List[Dict[str, Any]]]",
+    span: "Any",
+    scope: "Any",
+    max_single_message_chars: int = MAX_SINGLE_MESSAGE_CONTENT_CHARS,
+) -> "Optional[List[Dict[str, Any]]]":
+    if not messages:
+        return None
+
+    messages = redact_blob_message_parts(messages)
+
+    truncated_message = _truncate_single_message_content_if_present(
+        deepcopy(messages[-1]), max_chars=max_single_message_chars
+    )
+    if len(messages) > 1:
+        scope._gen_ai_original_message_count[span.span_id] = len(messages)
+
+    span.set_data(SPANDATA.META_GEN_AI_ORIGINAL_INPUT_MESSAGES_LENGTH, len(messages))
+
+    return [truncated_message]
+
+
+def truncate_and_annotate_embedding_inputs(
     messages: "Optional[List[Dict[str, Any]]]",
     span: "Any",
     scope: "Any",

sentry_sdk/consts.py

@@ -863,6 +863,14 @@ class SPANDATA:
     Example: "a1b2c3d4e5f6"
     """
 
+    META_GEN_AI_ORIGINAL_INPUT_MESSAGES_LENGTH = (
+        "sentry.sdk_meta.gen_ai.input.messages.original_length"
+    )
+    """
+    The original number of input non-system instruction messages, before SDK trimming.
+    Example: 4
+    """
+
 
 class SPANSTATUS:
     """

sentry_sdk/integrations/litellm.py

@@ -9,6 +9,7 @@
     set_data_normalized,
     truncate_and_annotate_messages,
     transform_openai_content_part,
+    truncate_and_annotate_embedding_inputs,
 )
 from sentry_sdk.consts import SPANDATA
 from sentry_sdk.integrations import DidNotEnable, Integration
@@ -118,7 +119,9 @@ def _input_callback(kwargs: "Dict[str, Any]") -> None:
                     if isinstance(embedding_input, list)
                     else [embedding_input]
                 )
-                messages_data = truncate_and_annotate_messages(input_list, span, scope)
+                messages_data = truncate_and_annotate_embedding_inputs(
+                    input_list, span, scope
+                )
                 if messages_data is not None:
                     set_data_normalized(
                         span,