|
| 1 | +import re |
| 2 | +from typing import Any |
| 3 | +from urllib.parse import urlparse |
| 4 | + |
| 5 | +REDACTED_IMAGE_PLACEHOLDER = "[base64 image redacted]" |
| 6 | + |
| 7 | + |
| 8 | +def is_base64_data_url(text: str) -> bool: |
| 9 | + return re.match(r"^data:([^;]+);base64,", text) is not None |
| 10 | + |
| 11 | + |
| 12 | +def is_valid_url(text: str) -> bool: |
| 13 | + try: |
| 14 | + result = urlparse(text) |
| 15 | + return bool(result.scheme and result.netloc) |
| 16 | + except Exception: |
| 17 | + pass |
| 18 | + |
| 19 | + return text.startswith(("/", "./", "../")) |
| 20 | + |
| 21 | + |
| 22 | +def is_raw_base64(text: str) -> bool: |
| 23 | + if is_valid_url(text): |
| 24 | + return False |
| 25 | + |
| 26 | + return len(text) > 20 and re.match(r"^[A-Za-z0-9+/]+=*$", text) is not None |
| 27 | + |
| 28 | + |
| 29 | +def redact_base64_data_url(value: Any) -> Any: |
| 30 | + if not isinstance(value, str): |
| 31 | + return value |
| 32 | + |
| 33 | + if is_base64_data_url(value): |
| 34 | + return REDACTED_IMAGE_PLACEHOLDER |
| 35 | + |
| 36 | + if is_raw_base64(value): |
| 37 | + return REDACTED_IMAGE_PLACEHOLDER |
| 38 | + |
| 39 | + return value |
| 40 | + |
| 41 | + |
| 42 | +def process_messages(messages: Any, transform_content_func) -> Any: |
| 43 | + if not messages: |
| 44 | + return messages |
| 45 | + |
| 46 | + def process_content(content: Any) -> Any: |
| 47 | + if isinstance(content, str): |
| 48 | + return content |
| 49 | + |
| 50 | + if not content: |
| 51 | + return content |
| 52 | + |
| 53 | + if isinstance(content, list): |
| 54 | + return [transform_content_func(item) for item in content] |
| 55 | + |
| 56 | + return transform_content_func(content) |
| 57 | + |
| 58 | + def process_message(msg: Any) -> Any: |
| 59 | + if not isinstance(msg, dict) or "content" not in msg: |
| 60 | + return msg |
| 61 | + return {**msg, "content": process_content(msg["content"])} |
| 62 | + |
| 63 | + if isinstance(messages, list): |
| 64 | + return [process_message(msg) for msg in messages] |
| 65 | + |
| 66 | + return process_message(messages) |
| 67 | + |
| 68 | + |
| 69 | +def sanitize_openai_image(item: Any) -> Any: |
| 70 | + if not isinstance(item, dict): |
| 71 | + return item |
| 72 | + |
| 73 | + if ( |
| 74 | + item.get("type") == "image_url" |
| 75 | + and isinstance(item.get("image_url"), dict) |
| 76 | + and "url" in item["image_url"] |
| 77 | + ): |
| 78 | + return { |
| 79 | + **item, |
| 80 | + "image_url": { |
| 81 | + **item["image_url"], |
| 82 | + "url": redact_base64_data_url(item["image_url"]["url"]), |
| 83 | + }, |
| 84 | + } |
| 85 | + |
| 86 | + return item |
| 87 | + |
| 88 | + |
| 89 | +def sanitize_openai_response_image(item: Any) -> Any: |
| 90 | + if not isinstance(item, dict): |
| 91 | + return item |
| 92 | + |
| 93 | + if item.get("type") == "input_image" and "image_url" in item: |
| 94 | + return { |
| 95 | + **item, |
| 96 | + "image_url": redact_base64_data_url(item["image_url"]), |
| 97 | + } |
| 98 | + |
| 99 | + return item |
| 100 | + |
| 101 | + |
| 102 | +def sanitize_anthropic_image(item: Any) -> Any: |
| 103 | + if not isinstance(item, dict): |
| 104 | + return item |
| 105 | + |
| 106 | + if ( |
| 107 | + item.get("type") == "image" |
| 108 | + and isinstance(item.get("source"), dict) |
| 109 | + and item["source"].get("type") == "base64" |
| 110 | + and "data" in item["source"] |
| 111 | + ): |
| 112 | + # For Anthropic, if the source type is "base64", we should always redact the data |
| 113 | + # The provider is explicitly telling us this is base64 data |
| 114 | + return { |
| 115 | + **item, |
| 116 | + "source": { |
| 117 | + **item["source"], |
| 118 | + "data": REDACTED_IMAGE_PLACEHOLDER, |
| 119 | + }, |
| 120 | + } |
| 121 | + |
| 122 | + return item |
| 123 | + |
| 124 | + |
| 125 | +def sanitize_gemini_part(part: Any) -> Any: |
| 126 | + if not isinstance(part, dict): |
| 127 | + return part |
| 128 | + |
| 129 | + if ( |
| 130 | + "inline_data" in part |
| 131 | + and isinstance(part["inline_data"], dict) |
| 132 | + and "data" in part["inline_data"] |
| 133 | + ): |
| 134 | + # For Gemini, the inline_data structure indicates base64 data |
| 135 | + # We should redact any string data in this context |
| 136 | + return { |
| 137 | + **part, |
| 138 | + "inline_data": { |
| 139 | + **part["inline_data"], |
| 140 | + "data": REDACTED_IMAGE_PLACEHOLDER, |
| 141 | + }, |
| 142 | + } |
| 143 | + |
| 144 | + return part |
| 145 | + |
| 146 | + |
| 147 | +def process_gemini_item(item: Any) -> Any: |
| 148 | + if not isinstance(item, dict): |
| 149 | + return item |
| 150 | + |
| 151 | + if "parts" in item and item["parts"]: |
| 152 | + parts = item["parts"] |
| 153 | + if isinstance(parts, list): |
| 154 | + parts = [sanitize_gemini_part(part) for part in parts] |
| 155 | + else: |
| 156 | + parts = sanitize_gemini_part(parts) |
| 157 | + |
| 158 | + return {**item, "parts": parts} |
| 159 | + |
| 160 | + return item |
| 161 | + |
| 162 | + |
| 163 | +def sanitize_langchain_image(item: Any) -> Any: |
| 164 | + if not isinstance(item, dict): |
| 165 | + return item |
| 166 | + |
| 167 | + if ( |
| 168 | + item.get("type") == "image_url" |
| 169 | + and isinstance(item.get("image_url"), dict) |
| 170 | + and "url" in item["image_url"] |
| 171 | + ): |
| 172 | + return { |
| 173 | + **item, |
| 174 | + "image_url": { |
| 175 | + **item["image_url"], |
| 176 | + "url": redact_base64_data_url(item["image_url"]["url"]), |
| 177 | + }, |
| 178 | + } |
| 179 | + |
| 180 | + if item.get("type") == "image" and "data" in item: |
| 181 | + return {**item, "data": redact_base64_data_url(item["data"])} |
| 182 | + |
| 183 | + if ( |
| 184 | + item.get("type") == "image" |
| 185 | + and isinstance(item.get("source"), dict) |
| 186 | + and "data" in item["source"] |
| 187 | + ): |
| 188 | + # Anthropic style - raw base64 in structured format, always redact |
| 189 | + return { |
| 190 | + **item, |
| 191 | + "source": { |
| 192 | + **item["source"], |
| 193 | + "data": REDACTED_IMAGE_PLACEHOLDER, |
| 194 | + }, |
| 195 | + } |
| 196 | + |
| 197 | + if item.get("type") == "media" and "data" in item: |
| 198 | + return {**item, "data": redact_base64_data_url(item["data"])} |
| 199 | + |
| 200 | + return item |
| 201 | + |
| 202 | + |
| 203 | +def sanitize_openai(data: Any) -> Any: |
| 204 | + return process_messages(data, sanitize_openai_image) |
| 205 | + |
| 206 | + |
| 207 | +def sanitize_openai_response(data: Any) -> Any: |
| 208 | + return process_messages(data, sanitize_openai_response_image) |
| 209 | + |
| 210 | + |
| 211 | +def sanitize_anthropic(data: Any) -> Any: |
| 212 | + return process_messages(data, sanitize_anthropic_image) |
| 213 | + |
| 214 | + |
| 215 | +def sanitize_gemini(data: Any) -> Any: |
| 216 | + if not data: |
| 217 | + return data |
| 218 | + |
| 219 | + if isinstance(data, list): |
| 220 | + return [process_gemini_item(item) for item in data] |
| 221 | + |
| 222 | + return process_gemini_item(data) |
| 223 | + |
| 224 | + |
| 225 | +def sanitize_langchain(data: Any) -> Any: |
| 226 | + return process_messages(data, sanitize_langchain_image) |
0 commit comments