posthoganalytics 6.7.0__tar.gz → 6.7.1__tar.gz

{posthoganalytics-6.7.0/posthoganalytics.egg-info → posthoganalytics-6.7.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: posthoganalytics
-Version: 6.7.0
+Version: 6.7.1
 Summary: Integrate PostHog into any python application.
 Home-page: https://github.com/posthog/posthog-python
 Author: Posthog

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics/ai/anthropic/anthropic.py

@@ -16,6 +16,7 @@ from posthoganalytics.ai.utils import (
     merge_system_prompt,
     with_privacy_mode,
 )
+from posthoganalytics.ai.sanitization import sanitize_anthropic
 from posthoganalytics.client import Client as PostHogClient
 from posthoganalytics import setup
 
@@ -184,7 +185,7 @@ class WrappedMessages(Messages):
             "$ai_input": with_privacy_mode(
                 self._client._ph_client,
                 posthog_privacy_mode,
-                merge_system_prompt(kwargs, "anthropic"),
+                sanitize_anthropic(merge_system_prompt(kwargs, "anthropic")),
             ),
             "$ai_output_choices": with_privacy_mode(
                 self._client._ph_client,

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics/ai/anthropic/anthropic_async.py

@@ -17,6 +17,7 @@ from posthoganalytics.ai.utils import (
     merge_system_prompt,
     with_privacy_mode,
 )
+from posthoganalytics.ai.sanitization import sanitize_anthropic
 from posthoganalytics.client import Client as PostHogClient
 
 
@@ -184,7 +185,7 @@ class AsyncWrappedMessages(AsyncMessages):
             "$ai_input": with_privacy_mode(
                 self._client._ph_client,
                 posthog_privacy_mode,
-                merge_system_prompt(kwargs, "anthropic"),
+                sanitize_anthropic(merge_system_prompt(kwargs, "anthropic")),
             ),
             "$ai_output_choices": with_privacy_mode(
                 self._client._ph_client,

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics/ai/gemini/gemini.py

@@ -16,6 +16,7 @@ from posthoganalytics.ai.utils import (
     get_model_params,
     with_privacy_mode,
 )
+from posthoganalytics.ai.sanitization import sanitize_gemini
 from posthoganalytics.client import Client as PostHogClient
 
 
@@ -347,7 +348,7 @@ class Models:
             "$ai_input": with_privacy_mode(
                 self._ph_client,
                 privacy_mode,
-                self._format_input(contents),
+                sanitize_gemini(self._format_input(contents)),
             ),
             "$ai_output_choices": with_privacy_mode(
                 self._ph_client,

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics/ai/langchain/callbacks.py

@@ -37,6 +37,7 @@ from pydantic import BaseModel
 
 from posthoganalytics import setup
 from posthoganalytics.ai.utils import get_model_params, with_privacy_mode
+from posthoganalytics.ai.sanitization import sanitize_langchain
 from posthoganalytics.client import Client
 
 log = logging.getLogger("posthog")
@@ -480,7 +481,7 @@ class CallbackHandler(BaseCallbackHandler):
         event_properties = {
             "$ai_trace_id": trace_id,
             "$ai_input_state": with_privacy_mode(
-                self._ph_client, self._privacy_mode, run.input
+                self._ph_client, self._privacy_mode, sanitize_langchain(run.input)
             ),
             "$ai_latency": run.latency,
             "$ai_span_name": run.name,
@@ -550,7 +551,7 @@ class CallbackHandler(BaseCallbackHandler):
             "$ai_model": run.model,
             "$ai_model_parameters": run.model_params,
             "$ai_input": with_privacy_mode(
-                self._ph_client, self._privacy_mode, run.input
+                self._ph_client, self._privacy_mode, sanitize_langchain(run.input)
             ),
             "$ai_http_status": 200,
             "$ai_latency": run.latency,

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics/ai/openai/openai.py

@@ -15,6 +15,7 @@ from posthoganalytics.ai.utils import (
     get_model_params,
     with_privacy_mode,
 )
+from posthoganalytics.ai.sanitization import sanitize_openai, sanitize_openai_response
 from posthoganalytics.client import Client as PostHogClient
 from posthoganalytics import setup
 
@@ -194,7 +195,9 @@ class WrappedResponses:
             "$ai_model": kwargs.get("model"),
             "$ai_model_parameters": get_model_params(kwargs),
             "$ai_input": with_privacy_mode(
-                self._client._ph_client, posthog_privacy_mode, kwargs.get("input")
+                self._client._ph_client,
+                posthog_privacy_mode,
+                sanitize_openai_response(kwargs.get("input")),
             ),
             "$ai_output_choices": with_privacy_mode(
                 self._client._ph_client,
@@ -427,7 +430,9 @@ class WrappedCompletions:
             "$ai_model": kwargs.get("model"),
             "$ai_model_parameters": get_model_params(kwargs),
             "$ai_input": with_privacy_mode(
-                self._client._ph_client, posthog_privacy_mode, kwargs.get("messages")
+                self._client._ph_client,
+                posthog_privacy_mode,
+                sanitize_openai(kwargs.get("messages")),
             ),
             "$ai_output_choices": with_privacy_mode(
                 self._client._ph_client,
@@ -518,7 +523,9 @@ class WrappedEmbeddings:
             "$ai_provider": "openai",
             "$ai_model": kwargs.get("model"),
             "$ai_input": with_privacy_mode(
-                self._client._ph_client, posthog_privacy_mode, kwargs.get("input")
+                self._client._ph_client,
+                posthog_privacy_mode,
+                sanitize_openai_response(kwargs.get("input")),
             ),
             "$ai_http_status": 200,
             "$ai_input_tokens": usage_stats.get("prompt_tokens", 0),

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics/ai/openai/openai_async.py

@@ -16,6 +16,7 @@ from posthoganalytics.ai.utils import (
     get_model_params,
     with_privacy_mode,
 )
+from posthoganalytics.ai.sanitization import sanitize_openai, sanitize_openai_response
 from posthoganalytics.client import Client as PostHogClient
 
 
@@ -195,7 +196,9 @@ class WrappedResponses:
             "$ai_model": kwargs.get("model"),
             "$ai_model_parameters": get_model_params(kwargs),
             "$ai_input": with_privacy_mode(
-                self._client._ph_client, posthog_privacy_mode, kwargs.get("input")
+                self._client._ph_client,
+                posthog_privacy_mode,
+                sanitize_openai_response(kwargs.get("input")),
             ),
             "$ai_output_choices": with_privacy_mode(
                 self._client._ph_client,
@@ -431,7 +434,9 @@ class WrappedCompletions:
             "$ai_model": kwargs.get("model"),
             "$ai_model_parameters": get_model_params(kwargs),
             "$ai_input": with_privacy_mode(
-                self._client._ph_client, posthog_privacy_mode, kwargs.get("messages")
+                self._client._ph_client,
+                posthog_privacy_mode,
+                sanitize_openai(kwargs.get("messages")),
             ),
             "$ai_output_choices": with_privacy_mode(
                 self._client._ph_client,
@@ -522,7 +527,9 @@ class WrappedEmbeddings:
             "$ai_provider": "openai",
             "$ai_model": kwargs.get("model"),
             "$ai_input": with_privacy_mode(
-                self._client._ph_client, posthog_privacy_mode, kwargs.get("input")
+                self._client._ph_client,
+                posthog_privacy_mode,
+                sanitize_openai_response(kwargs.get("input")),
             ),
             "$ai_http_status": 200,
             "$ai_input_tokens": usage_stats.get("prompt_tokens", 0),

posthoganalytics-6.7.1/posthoganalytics/ai/sanitization.py

@@ -0,0 +1,226 @@
+import re
+from typing import Any
+from urllib.parse import urlparse
+
+REDACTED_IMAGE_PLACEHOLDER = "[base64 image redacted]"
+
+
+def is_base64_data_url(text: str) -> bool:
+    return re.match(r"^data:([^;]+);base64,", text) is not None
+
+
+def is_valid_url(text: str) -> bool:
+    try:
+        result = urlparse(text)
+        return bool(result.scheme and result.netloc)
+    except Exception:
+        pass
+
+    return text.startswith(("/", "./", "../"))
+
+
+def is_raw_base64(text: str) -> bool:
+    if is_valid_url(text):
+        return False
+
+    return len(text) > 20 and re.match(r"^[A-Za-z0-9+/]+=*$", text) is not None
+
+
+def redact_base64_data_url(value: Any) -> Any:
+    if not isinstance(value, str):
+        return value
+
+    if is_base64_data_url(value):
+        return REDACTED_IMAGE_PLACEHOLDER
+
+    if is_raw_base64(value):
+        return REDACTED_IMAGE_PLACEHOLDER
+
+    return value
+
+
+def process_messages(messages: Any, transform_content_func) -> Any:
+    if not messages:
+        return messages
+
+    def process_content(content: Any) -> Any:
+        if isinstance(content, str):
+            return content
+
+        if not content:
+            return content
+
+        if isinstance(content, list):
+            return [transform_content_func(item) for item in content]
+
+        return transform_content_func(content)
+
+    def process_message(msg: Any) -> Any:
+        if not isinstance(msg, dict) or "content" not in msg:
+            return msg
+        return {**msg, "content": process_content(msg["content"])}
+
+    if isinstance(messages, list):
+        return [process_message(msg) for msg in messages]
+
+    return process_message(messages)
+
+
+def sanitize_openai_image(item: Any) -> Any:
+    if not isinstance(item, dict):
+        return item
+
+    if (
+        item.get("type") == "image_url"
+        and isinstance(item.get("image_url"), dict)
+        and "url" in item["image_url"]
+    ):
+        return {
+            **item,
+            "image_url": {
+                **item["image_url"],
+                "url": redact_base64_data_url(item["image_url"]["url"]),
+            },
+        }
+
+    return item
+
+
+def sanitize_openai_response_image(item: Any) -> Any:
+    if not isinstance(item, dict):
+        return item
+
+    if item.get("type") == "input_image" and "image_url" in item:
+        return {
+            **item,
+            "image_url": redact_base64_data_url(item["image_url"]),
+        }
+
+    return item
+
+
+def sanitize_anthropic_image(item: Any) -> Any:
+    if not isinstance(item, dict):
+        return item
+
+    if (
+        item.get("type") == "image"
+        and isinstance(item.get("source"), dict)
+        and item["source"].get("type") == "base64"
+        and "data" in item["source"]
+    ):
+        # For Anthropic, if the source type is "base64", we should always redact the data
+        # The provider is explicitly telling us this is base64 data
+        return {
+            **item,
+            "source": {
+                **item["source"],
+                "data": REDACTED_IMAGE_PLACEHOLDER,
+            },
+        }
+
+    return item
+
+
+def sanitize_gemini_part(part: Any) -> Any:
+    if not isinstance(part, dict):
+        return part
+
+    if (
+        "inline_data" in part
+        and isinstance(part["inline_data"], dict)
+        and "data" in part["inline_data"]
+    ):
+        # For Gemini, the inline_data structure indicates base64 data
+        # We should redact any string data in this context
+        return {
+            **part,
+            "inline_data": {
+                **part["inline_data"],
+                "data": REDACTED_IMAGE_PLACEHOLDER,
+            },
+        }
+
+    return part
+
+
+def process_gemini_item(item: Any) -> Any:
+    if not isinstance(item, dict):
+        return item
+
+    if "parts" in item and item["parts"]:
+        parts = item["parts"]
+        if isinstance(parts, list):
+            parts = [sanitize_gemini_part(part) for part in parts]
+        else:
+            parts = sanitize_gemini_part(parts)
+
+        return {**item, "parts": parts}
+
+    return item
+
+
+def sanitize_langchain_image(item: Any) -> Any:
+    if not isinstance(item, dict):
+        return item
+
+    if (
+        item.get("type") == "image_url"
+        and isinstance(item.get("image_url"), dict)
+        and "url" in item["image_url"]
+    ):
+        return {
+            **item,
+            "image_url": {
+                **item["image_url"],
+                "url": redact_base64_data_url(item["image_url"]["url"]),
+            },
+        }
+
+    if item.get("type") == "image" and "data" in item:
+        return {**item, "data": redact_base64_data_url(item["data"])}
+
+    if (
+        item.get("type") == "image"
+        and isinstance(item.get("source"), dict)
+        and "data" in item["source"]
+    ):
+        # Anthropic style - raw base64 in structured format, always redact
+        return {
+            **item,
+            "source": {
+                **item["source"],
+                "data": REDACTED_IMAGE_PLACEHOLDER,
+            },
+        }
+
+    if item.get("type") == "media" and "data" in item:
+        return {**item, "data": redact_base64_data_url(item["data"])}
+
+    return item
+
+
+def sanitize_openai(data: Any) -> Any:
+    return process_messages(data, sanitize_openai_image)
+
+
+def sanitize_openai_response(data: Any) -> Any:
+    return process_messages(data, sanitize_openai_response_image)
+
+
+def sanitize_anthropic(data: Any) -> Any:
+    return process_messages(data, sanitize_anthropic_image)
+
+
+def sanitize_gemini(data: Any) -> Any:
+    if not data:
+        return data
+
+    if isinstance(data, list):
+        return [process_gemini_item(item) for item in data]
+
+    return process_gemini_item(data)
+
+
+def sanitize_langchain(data: Any) -> Any:
+    return process_messages(data, sanitize_langchain_image)

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics/ai/utils.py

@@ -5,6 +5,12 @@ from typing import Any, Callable, Dict, List, Optional
 from httpx import URL
 
 from posthoganalytics.client import Client as PostHogClient
+from posthoganalytics.ai.sanitization import (
+    sanitize_openai,
+    sanitize_anthropic,
+    sanitize_gemini,
+    sanitize_langchain,
+)
 
 
 def get_model_params(kwargs: Dict[str, Any]) -> Dict[str, Any]:
@@ -422,12 +428,15 @@ def call_llm_and_track_usage(
             usage = get_usage(response, provider)
 
         messages = merge_system_prompt(kwargs, provider)
+        sanitized_messages = sanitize_messages(messages, provider)
 
         event_properties = {
             "$ai_provider": provider,
             "$ai_model": kwargs.get("model"),
             "$ai_model_parameters": get_model_params(kwargs),
-            "$ai_input": with_privacy_mode(ph_client, posthog_privacy_mode, messages),
+            "$ai_input": with_privacy_mode(
+                ph_client, posthog_privacy_mode, sanitized_messages
+            ),
             "$ai_output_choices": with_privacy_mode(
                 ph_client, posthog_privacy_mode, format_response(response, provider)
             ),
@@ -536,12 +545,15 @@ async def call_llm_and_track_usage_async(
             usage = get_usage(response, provider)
 
         messages = merge_system_prompt(kwargs, provider)
+        sanitized_messages = sanitize_messages(messages, provider)
 
         event_properties = {
             "$ai_provider": provider,
             "$ai_model": kwargs.get("model"),
             "$ai_model_parameters": get_model_params(kwargs),
-            "$ai_input": with_privacy_mode(ph_client, posthog_privacy_mode, messages),
+            "$ai_input": with_privacy_mode(
+                ph_client, posthog_privacy_mode, sanitized_messages
+            ),
             "$ai_output_choices": with_privacy_mode(
                 ph_client, posthog_privacy_mode, format_response(response, provider)
             ),
@@ -600,6 +612,19 @@ async def call_llm_and_track_usage_async(
     return response
 
 
+def sanitize_messages(data: Any, provider: str) -> Any:
+    """Sanitize messages using provider-specific sanitization functions."""
+    if provider == "anthropic":
+        return sanitize_anthropic(data)
+    elif provider == "openai":
+        return sanitize_openai(data)
+    elif provider == "gemini":
+        return sanitize_gemini(data)
+    elif provider == "langchain":
+        return sanitize_langchain(data)
+    return data
+
+
 def with_privacy_mode(ph_client: PostHogClient, privacy_mode: bool, value: Any):
     if ph_client.privacy_mode or privacy_mode:
         return None

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics/version.py

@@ -1,4 +1,4 @@
-VERSION = "6.7.0"
+VERSION = "6.7.1"
 
 if __name__ == "__main__":
     print(VERSION, end="")  # noqa: T201

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1/posthoganalytics.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: posthoganalytics
-Version: 6.7.0
+Version: 6.7.1
 Summary: Integrate PostHog into any python application.
 Home-page: https://github.com/posthog/posthog-python
 Author: Posthog

{posthoganalytics-6.7.0 → posthoganalytics-6.7.1}/posthoganalytics.egg-info/SOURCES.txt

@@ -24,6 +24,7 @@ posthoganalytics.egg-info/dependency_links.txt
 posthoganalytics.egg-info/requires.txt
 posthoganalytics.egg-info/top_level.txt
 posthoganalytics/ai/__init__.py
+posthoganalytics/ai/sanitization.py
 posthoganalytics/ai/utils.py
 posthoganalytics/ai/anthropic/__init__.py
 posthoganalytics/ai/anthropic/anthropic.py