import inspect import json from collections import deque from copy import deepcopy from sys import getsizeof from typing import TYPE_CHECKING from sentry_sdk._types import BLOB_DATA_SUBSTITUTE if TYPE_CHECKING: from typing import Any, Callable, Dict, List, Optional, Tuple from sentry_sdk.tracing import Span import sentry_sdk from sentry_sdk.utils import logger MAX_GEN_AI_MESSAGE_BYTES = 20_000 # 20KB # Maximum characters when only a single message is left after bytes truncation MAX_SINGLE_MESSAGE_CONTENT_CHARS = 10_000 class GEN_AI_ALLOWED_MESSAGE_ROLES: SYSTEM = "system" USER = "user" ASSISTANT = "assistant" TOOL = "tool" GEN_AI_MESSAGE_ROLE_REVERSE_MAPPING = { GEN_AI_ALLOWED_MESSAGE_ROLES.SYSTEM: ["system"], GEN_AI_ALLOWED_MESSAGE_ROLES.USER: ["user", "human"], GEN_AI_ALLOWED_MESSAGE_ROLES.ASSISTANT: ["assistant", "ai"], GEN_AI_ALLOWED_MESSAGE_ROLES.TOOL: ["tool", "tool_call"], } GEN_AI_MESSAGE_ROLE_MAPPING = {} for target_role, source_roles in GEN_AI_MESSAGE_ROLE_REVERSE_MAPPING.items(): for source_role in source_roles: GEN_AI_MESSAGE_ROLE_MAPPING[source_role] = target_role def parse_data_uri(url: str) -> "Tuple[str, str]": """ Parse a data URI and return (mime_type, content). Data URI format (RFC 2397): data:[][;base64], Examples: data:image/jpeg;base64,/9j/4AAQ... → ("image/jpeg", "/9j/4AAQ...") data:text/plain,Hello → ("text/plain", "Hello") data:;base64,SGVsbG8= → ("", "SGVsbG8=") Raises: ValueError: If the URL is not a valid data URI (missing comma separator) """ if "," not in url: raise ValueError("Invalid data URI: missing comma separator") header, content = url.split(",", 1) # Extract mime type from header # Format: "data:[;param1][;param2]..." e.g. "data:image/jpeg;base64" # Remove "data:" prefix, then take everything before the first semicolon if header.startswith("data:"): mime_part = header[5:] # Remove "data:" prefix else: mime_part = header mime_type = mime_part.split(";")[0] return mime_type, content def get_modality_from_mime_type(mime_type: str) -> str: """ Infer the content modality from a MIME type string. Args: mime_type: A MIME type string (e.g., "image/jpeg", "audio/mp3") Returns: One of: "image", "audio", "video", or "document" Defaults to "image" for unknown or empty MIME types. Examples: "image/jpeg" -> "image" "audio/mp3" -> "audio" "video/mp4" -> "video" "application/pdf" -> "document" "text/plain" -> "document" """ if not mime_type: return "image" # Default fallback mime_lower = mime_type.lower() if mime_lower.startswith("image/"): return "image" elif mime_lower.startswith("audio/"): return "audio" elif mime_lower.startswith("video/"): return "video" elif mime_lower.startswith("application/") or mime_lower.startswith("text/"): return "document" else: return "image" # Default fallback for unknown types def transform_openai_content_part( content_part: "Dict[str, Any]", ) -> "Optional[Dict[str, Any]]": """ Transform an OpenAI/LiteLLM content part to Sentry's standardized format. This handles the OpenAI image_url format used by OpenAI and LiteLLM SDKs. Input format: - {"type": "image_url", "image_url": {"url": "..."}} - {"type": "image_url", "image_url": "..."} (string shorthand) Output format (one of): - {"type": "blob", "modality": "image", "mime_type": "...", "content": "..."} - {"type": "uri", "modality": "image", "mime_type": "", "uri": "..."} Args: content_part: A dictionary representing a content part from OpenAI/LiteLLM Returns: A transformed dictionary in standardized format, or None if the format is not OpenAI image_url format or transformation fails. """ if not isinstance(content_part, dict): return None block_type = content_part.get("type") if block_type != "image_url": return None image_url_data = content_part.get("image_url") if isinstance(image_url_data, str): url = image_url_data elif isinstance(image_url_data, dict): url = image_url_data.get("url", "") else: return None if not url: return None # Check if it's a data URI (base64 encoded) if url.startswith("data:"): try: mime_type, content = parse_data_uri(url) return { "type": "blob", "modality": get_modality_from_mime_type(mime_type), "mime_type": mime_type, "content": content, } except ValueError: # If parsing fails, return as URI return { "type": "uri", "modality": "image", "mime_type": "", "uri": url, } else: # Regular URL return { "type": "uri", "modality": "image", "mime_type": "", "uri": url, } def transform_anthropic_content_part( content_part: "Dict[str, Any]", ) -> "Optional[Dict[str, Any]]": """ Transform an Anthropic content part to Sentry's standardized format. This handles the Anthropic image and document formats with source dictionaries. Input format: - {"type": "image", "source": {"type": "base64", "media_type": "...", "data": "..."}} - {"type": "image", "source": {"type": "url", "media_type": "...", "url": "..."}} - {"type": "image", "source": {"type": "file", "media_type": "...", "file_id": "..."}} - {"type": "document", "source": {...}} (same source formats) Output format (one of): - {"type": "blob", "modality": "...", "mime_type": "...", "content": "..."} - {"type": "uri", "modality": "...", "mime_type": "...", "uri": "..."} - {"type": "file", "modality": "...", "mime_type": "...", "file_id": "..."} Args: content_part: A dictionary representing a content part from Anthropic Returns: A transformed dictionary in standardized format, or None if the format is not Anthropic format or transformation fails. """ if not isinstance(content_part, dict): return None block_type = content_part.get("type") if block_type not in ("image", "document") or "source" not in content_part: return None source = content_part.get("source") if not isinstance(source, dict): return None source_type = source.get("type") media_type = source.get("media_type", "") modality = ( "document" if block_type == "document" else get_modality_from_mime_type(media_type) ) if source_type == "base64": return { "type": "blob", "modality": modality, "mime_type": media_type, "content": source.get("data", ""), } elif source_type == "url": return { "type": "uri", "modality": modality, "mime_type": media_type, "uri": source.get("url", ""), } elif source_type == "file": return { "type": "file", "modality": modality, "mime_type": media_type, "file_id": source.get("file_id", ""), } return None def transform_google_content_part( content_part: "Dict[str, Any]", ) -> "Optional[Dict[str, Any]]": """ Transform a Google GenAI content part to Sentry's standardized format. This handles the Google GenAI inline_data and file_data formats. Input format: - {"inline_data": {"mime_type": "...", "data": "..."}} - {"file_data": {"mime_type": "...", "file_uri": "..."}} Output format (one of): - {"type": "blob", "modality": "...", "mime_type": "...", "content": "..."} - {"type": "uri", "modality": "...", "mime_type": "...", "uri": "..."} Args: content_part: A dictionary representing a content part from Google GenAI Returns: A transformed dictionary in standardized format, or None if the format is not Google format or transformation fails. """ if not isinstance(content_part, dict): return None # Handle Google inline_data format if "inline_data" in content_part: inline_data = content_part.get("inline_data") if isinstance(inline_data, dict): mime_type = inline_data.get("mime_type", "") return { "type": "blob", "modality": get_modality_from_mime_type(mime_type), "mime_type": mime_type, "content": inline_data.get("data", ""), } return None # Handle Google file_data format if "file_data" in content_part: file_data = content_part.get("file_data") if isinstance(file_data, dict): mime_type = file_data.get("mime_type", "") return { "type": "uri", "modality": get_modality_from_mime_type(mime_type), "mime_type": mime_type, "uri": file_data.get("file_uri", ""), } return None return None def transform_generic_content_part( content_part: "Dict[str, Any]", ) -> "Optional[Dict[str, Any]]": """ Transform a generic/LangChain-style content part to Sentry's standardized format. This handles generic formats where the type indicates the modality and the data is provided via direct base64, url, or file_id fields. Input format: - {"type": "image", "base64": "...", "mime_type": "..."} - {"type": "audio", "url": "...", "mime_type": "..."} - {"type": "video", "base64": "...", "mime_type": "..."} - {"type": "file", "file_id": "...", "mime_type": "..."} Output format (one of): - {"type": "blob", "modality": "...", "mime_type": "...", "content": "..."} - {"type": "uri", "modality": "...", "mime_type": "...", "uri": "..."} - {"type": "file", "modality": "...", "mime_type": "...", "file_id": "..."} Args: content_part: A dictionary representing a content part in generic format Returns: A transformed dictionary in standardized format, or None if the format is not generic format or transformation fails. """ if not isinstance(content_part, dict): return None block_type = content_part.get("type") if block_type not in ("image", "audio", "video", "file"): return None # Ensure it's not Anthropic format (which also uses type: "image") if "source" in content_part: return None mime_type = content_part.get("mime_type", "") modality = block_type if block_type != "file" else "document" # Check for base64 encoded content if "base64" in content_part: return { "type": "blob", "modality": modality, "mime_type": mime_type, "content": content_part.get("base64", ""), } # Check for URL reference elif "url" in content_part: return { "type": "uri", "modality": modality, "mime_type": mime_type, "uri": content_part.get("url", ""), } # Check for file_id reference elif "file_id" in content_part: return { "type": "file", "modality": modality, "mime_type": mime_type, "file_id": content_part.get("file_id", ""), } return None def transform_content_part( content_part: "Dict[str, Any]", ) -> "Optional[Dict[str, Any]]": """ Transform a content part from various AI SDK formats to Sentry's standardized format. This is a heuristic dispatcher that detects the format and delegates to the appropriate SDK-specific transformer. For direct SDK integration, prefer using the specific transformers directly: - transform_openai_content_part() for OpenAI/LiteLLM - transform_anthropic_content_part() for Anthropic - transform_google_content_part() for Google GenAI - transform_generic_content_part() for LangChain and other generic formats Detection order: 1. OpenAI: type == "image_url" 2. Google: "inline_data" or "file_data" keys present 3. Anthropic: type in ("image", "document") with "source" key 4. Generic: type in ("image", "audio", "video", "file") with base64/url/file_id Output format (one of): - {"type": "blob", "modality": "...", "mime_type": "...", "content": "..."} - {"type": "uri", "modality": "...", "mime_type": "...", "uri": "..."} - {"type": "file", "modality": "...", "mime_type": "...", "file_id": "..."} Args: content_part: A dictionary representing a content part from an AI SDK Returns: A transformed dictionary in standardized format, or None if the format is unrecognized or transformation fails. """ if not isinstance(content_part, dict): return None # Try OpenAI format first (most common, clear indicator) result = transform_openai_content_part(content_part) if result is not None: return result # Try Google format (unique keys make it easy to detect) result = transform_google_content_part(content_part) if result is not None: return result # Try Anthropic format (has "source" key) result = transform_anthropic_content_part(content_part) if result is not None: return result # Try generic format as fallback result = transform_generic_content_part(content_part) if result is not None: return result # Unrecognized format return None def transform_message_content(content: "Any") -> "Any": """ Transform message content, handling both string content and list of content blocks. For list content, each item is transformed using transform_content_part(). Items that cannot be transformed (return None) are kept as-is. Args: content: Message content - can be a string, list of content blocks, or other Returns: - String content: returned as-is - List content: list with each transformable item converted to standardized format - Other: returned as-is """ if isinstance(content, str): return content if isinstance(content, (list, tuple)): transformed = [] for item in content: if isinstance(item, dict): result = transform_content_part(item) # If transformation succeeded, use the result; otherwise keep original transformed.append(result if result is not None else item) else: transformed.append(item) return transformed return content def _normalize_data(data: "Any", unpack: bool = True) -> "Any": # convert pydantic data (e.g. OpenAI v1+) to json compatible format if hasattr(data, "model_dump"): # Check if it's a class (type) rather than an instance # Model classes can be passed as arguments (e.g., for schema definitions) if inspect.isclass(data): return f"" try: return _normalize_data(data.model_dump(), unpack=unpack) except Exception as e: logger.warning("Could not convert pydantic data to JSON: %s", e) return data if isinstance(data, (int, float, bool, str)) else str(data) if isinstance(data, list): if unpack and len(data) == 1: return _normalize_data(data[0], unpack=unpack) # remove empty dimensions return list(_normalize_data(x, unpack=unpack) for x in data) if isinstance(data, dict): return {k: _normalize_data(v, unpack=unpack) for (k, v) in data.items()} return data if isinstance(data, (int, float, bool, str)) else str(data) def set_data_normalized( span: "Span", key: str, value: "Any", unpack: bool = True ) -> None: normalized = _normalize_data(value, unpack=unpack) if isinstance(normalized, (int, float, bool, str)): span.set_data(key, normalized) else: span.set_data(key, json.dumps(normalized)) def normalize_message_role(role: str) -> str: """ Normalize a message role to one of the 4 allowed gen_ai role values. Maps "ai" -> "assistant" and keeps other standard roles unchanged. """ return GEN_AI_MESSAGE_ROLE_MAPPING.get(role, role) def normalize_message_roles(messages: "list[dict[str, Any]]") -> "list[dict[str, Any]]": """ Normalize roles in a list of messages to use standard gen_ai role values. Creates a deep copy to avoid modifying the original messages. """ normalized_messages = [] for message in messages: if not isinstance(message, dict): normalized_messages.append(message) continue normalized_message = message.copy() if "role" in message: normalized_message["role"] = normalize_message_role(message["role"]) normalized_messages.append(normalized_message) return normalized_messages def get_start_span_function() -> "Callable[..., Any]": current_span = sentry_sdk.get_current_span() transaction_exists = ( current_span is not None and current_span.containing_transaction is not None ) return sentry_sdk.start_span if transaction_exists else sentry_sdk.start_transaction def _truncate_single_message_content_if_present( message: "Dict[str, Any]", max_chars: int ) -> "Dict[str, Any]": """ Truncate a message's content to at most `max_chars` characters and append an ellipsis if truncation occurs. """ if not isinstance(message, dict) or "content" not in message: return message content = message["content"] if not isinstance(content, str) or len(content) <= max_chars: return message message["content"] = content[:max_chars] + "..." return message def _find_truncation_index(messages: "List[Dict[str, Any]]", max_bytes: int) -> int: """ Find the index of the first message that would exceed the max bytes limit. Compute the individual message sizes, and return the index of the first message from the back of the list that would exceed the max bytes limit. """ running_sum = 0 for idx in range(len(messages) - 1, -1, -1): size = len(json.dumps(messages[idx], separators=(",", ":")).encode("utf-8")) running_sum += size if running_sum > max_bytes: return idx + 1 return 0 def redact_blob_message_parts( messages: "List[Dict[str, Any]]", ) -> "List[Dict[str, Any]]": """ Redact blob message parts from the messages by replacing blob content with "[Filtered]". This function creates a deep copy of messages that contain blob content to avoid mutating the original message dictionaries. Messages without blob content are returned as-is to minimize copying overhead. e.g: { "role": "user", "content": [ { "text": "How many ponies do you see in the image?", "type": "text" }, { "type": "blob", "modality": "image", "mime_type": "image/jpeg", "content": "data:image/jpeg;base64,..." } ] } becomes: { "role": "user", "content": [ { "text": "How many ponies do you see in the image?", "type": "text" }, { "type": "blob", "modality": "image", "mime_type": "image/jpeg", "content": "[Filtered]" } ] } """ # First pass: check if any message contains blob content has_blobs = False for message in messages: if not isinstance(message, dict): continue content = message.get("content") if isinstance(content, list): for item in content: if isinstance(item, dict) and item.get("type") == "blob": has_blobs = True break if has_blobs: break # If no blobs found, return original messages to avoid unnecessary copying if not has_blobs: return messages # Deep copy messages to avoid mutating the original messages_copy = deepcopy(messages) # Second pass: redact blob content in the copy for message in messages_copy: if not isinstance(message, dict): continue content = message.get("content") if isinstance(content, list): for item in content: if isinstance(item, dict) and item.get("type") == "blob": item["content"] = BLOB_DATA_SUBSTITUTE return messages_copy def truncate_messages_by_size( messages: "List[Dict[str, Any]]", max_bytes: int = MAX_GEN_AI_MESSAGE_BYTES, max_single_message_chars: int = MAX_SINGLE_MESSAGE_CONTENT_CHARS, ) -> "Tuple[List[Dict[str, Any]], int]": """ Returns a truncated messages list, consisting of - the last message, with its content truncated to `max_single_message_chars` characters, if the last message's size exceeds `max_bytes` bytes; otherwise, - the maximum number of messages, starting from the end of the `messages` list, whose total serialized size does not exceed `max_bytes` bytes. In the single message case, the serialized message size may exceed `max_bytes`, because truncation is based only on character count in that case. """ serialized_json = json.dumps(messages, separators=(",", ":")) current_size = len(serialized_json.encode("utf-8")) if current_size <= max_bytes: return messages, 0 truncation_index = _find_truncation_index(messages, max_bytes) if truncation_index < len(messages): truncated_messages = messages[truncation_index:] else: truncation_index = len(messages) - 1 truncated_messages = messages[-1:] if len(truncated_messages) == 1: truncated_messages[0] = _truncate_single_message_content_if_present( deepcopy(truncated_messages[0]), max_chars=max_single_message_chars ) return truncated_messages, truncation_index 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) return [truncated_message] def truncate_and_annotate_embedding_inputs( messages: "Optional[List[Dict[str, Any]]]", span: "Any", scope: "Any", max_bytes: int = MAX_GEN_AI_MESSAGE_BYTES, ) -> "Optional[List[Dict[str, Any]]]": if not messages: return None messages = redact_blob_message_parts(messages) truncated_messages, removed_count = truncate_messages_by_size(messages, max_bytes) if removed_count > 0: scope._gen_ai_original_message_count[span.span_id] = len(messages) return truncated_messages def set_conversation_id(conversation_id: str) -> None: """ Set the conversation_id in the scope. """ scope = sentry_sdk.get_current_scope() scope.set_conversation_id(conversation_id)