Merge master into fix/llma-streaming-providers-with-tool-calls

Resolved conflicts by:
- Keeping both StreamingEventData approach and new sanitization imports
- Applying sanitization to formatted inputs before passing to StreamingEventData
- Ensuring privacy mode and special token fields are handled by capture_streaming_event

Author: Radu-Raicea

CHANGELOG.md

@@ -1,3 +1,7 @@
+# 6.7.1 - 2025-09-01
+
+- fix: Add base64 inline image sanitization
+
 # 6.7.0 - 2025-08-26
 
 - feat: Add support for feature flag dependencies

posthog/ai/anthropic/anthropic.py

@@ -20,6 +20,7 @@
     handle_anthropic_tool_delta,
     finalize_anthropic_tool_input,
 )
+from posthog.ai.sanitization import sanitize_anthropic
 from posthog.client import Client as PostHogClient
 from posthog import setup
 
@@ -219,12 +220,15 @@ def _capture_streaming_event(
         from posthog.ai.utils import capture_streaming_event
 
         # Prepare standardized event data
+        formatted_input = format_anthropic_streaming_input(kwargs)
+        sanitized_input = sanitize_anthropic(formatted_input)
+        
         event_data = StreamingEventData(
             provider="anthropic",
             model=kwargs.get("model"),
             base_url=str(self._client.base_url),
             kwargs=kwargs,
-            formatted_input=format_anthropic_streaming_input(kwargs),
+            formatted_input=sanitized_input,
             formatted_output=format_anthropic_streaming_output_complete(
                 content_blocks, accumulated_content
             ),

posthog/ai/anthropic/anthropic_async.py

@@ -26,6 +26,7 @@
     handle_anthropic_tool_delta,
     finalize_anthropic_tool_input,
 )
+from posthog.ai.sanitization import sanitize_anthropic
 from posthog.client import Client as PostHogClient
 
 
@@ -240,7 +241,7 @@ async def _capture_streaming_event(
             "$ai_input": with_privacy_mode(
                 self._client._ph_client,
                 posthog_privacy_mode,
-                merge_system_prompt(kwargs, "anthropic"),
+                sanitize_anthropic(merge_system_prompt(kwargs, "anthropic")),
             ),
             "$ai_output_choices": with_privacy_mode(
                 self._client._ph_client,

posthog/ai/gemini/gemini.py

@@ -21,6 +21,7 @@
     extract_gemini_content_from_chunk,
     format_gemini_streaming_output,
 )
+from posthog.ai.sanitization import sanitize_gemini
 from posthog.client import Client as PostHogClient
 
 
@@ -355,12 +356,15 @@ def _capture_streaming_event(
         from posthog.ai.gemini.gemini_converter import standardize_gemini_usage
 
         # Prepare standardized event data
+        formatted_input = self._format_input(contents)
+        sanitized_input = sanitize_gemini(formatted_input)
+        
         event_data = StreamingEventData(
             provider="gemini",
             model=model,
             base_url=self._base_url,
             kwargs=kwargs,
-            formatted_input=self._format_input(contents),
+            formatted_input=sanitized_input,
             formatted_output=format_gemini_streaming_output(output),
             usage_stats=standardize_gemini_usage(usage_stats),
             latency=latency,

posthog/ai/langchain/callbacks.py

@@ -37,6 +37,7 @@
 
 from posthog import setup
 from posthog.ai.utils import get_model_params, with_privacy_mode
+from posthog.ai.sanitization import sanitize_langchain
 from posthog.client import Client
 
 log = logging.getLogger("posthog")
@@ -480,7 +481,7 @@ def _capture_trace_or_span(
         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 @@ def _capture_generation(
             "$ai_model": run.model,
             "$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,

posthog/ai/openai/openai.py

@@ -20,6 +20,7 @@
     extract_openai_tool_calls_from_chunk,
     accumulate_openai_tool_calls,
 )
+from posthog.ai.sanitization import sanitize_openai, sanitize_openai_response
 from posthog.client import Client as PostHogClient
 from posthog import setup
 
@@ -183,12 +184,15 @@ def _capture_streaming_event(
         from posthog.ai.utils import capture_streaming_event
 
         # Prepare standardized event data
+        formatted_input = format_openai_streaming_input(kwargs, "responses")
+        sanitized_input = sanitize_openai_response(formatted_input)
+        
         event_data = StreamingEventData(
             provider="openai",
             model=kwargs.get("model"),
             base_url=str(self._client.base_url),
             kwargs=kwargs,
-            formatted_input=format_openai_streaming_input(kwargs, "responses"),
+            formatted_input=sanitized_input,
             formatted_output=format_openai_streaming_output(output, "responses"),
             usage_stats=standardize_openai_usage(usage_stats, "responses"),
             latency=latency,
@@ -397,12 +401,15 @@ def _capture_streaming_event(
         from posthog.ai.utils import capture_streaming_event
 
         # Prepare standardized event data
+        formatted_input = format_openai_streaming_input(kwargs, "chat")
+        sanitized_input = sanitize_openai(formatted_input)
+        
         event_data = StreamingEventData(
             provider="openai",
             model=kwargs.get("model"),
             base_url=str(self._client.base_url),
             kwargs=kwargs,
-            formatted_input=format_openai_streaming_input(kwargs, "chat"),
+            formatted_input=sanitized_input,
             formatted_output=format_openai_streaming_output(output, "chat", tool_calls),
             usage_stats=standardize_openai_usage(usage_stats, "chat"),
             latency=latency,
@@ -474,7 +481,9 @@ def create(
             "$ai_provider": "openai",
             "$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),

posthog/ai/openai/openai_async.py

@@ -23,6 +23,7 @@
     accumulate_openai_tool_calls,
     format_openai_streaming_output,
 )
+from posthog.ai.sanitization import sanitize_openai, sanitize_openai_response
 from posthog.client import Client as PostHogClient
 
 
@@ -187,7 +188,9 @@ async def _capture_streaming_event(
             "$ai_model": kwargs.get("model"),
             "$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,
@@ -416,7 +419,9 @@ async def _capture_streaming_event(
             "$ai_model": kwargs.get("model"),
             "$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,
@@ -510,7 +515,9 @@ async def create(
             "$ai_provider": "openai",
             "$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),