posthoganalytics 7.0.1__py3-none-any.whl → 7.4.1__py3-none-any.whl

posthoganalytics/ai/openai/openai.py

@@ -124,14 +124,23 @@ class WrappedResponses:
         start_time = time.time()
         usage_stats: TokenUsage = TokenUsage()
         final_content = []
+        model_from_response: Optional[str] = None
         response = self._original.create(**kwargs)
 
         def generator():
             nonlocal usage_stats
             nonlocal final_content  # noqa: F824
+            nonlocal model_from_response
 
             try:
                 for chunk in response:
+                    # Extract model from response object in chunk (for stored prompts)
+                    if hasattr(chunk, "response") and chunk.response:
+                        if model_from_response is None and hasattr(
+                            chunk.response, "model"
+                        ):
+                            model_from_response = chunk.response.model
+
                     # Extract usage stats from chunk
                     chunk_usage = extract_openai_usage_from_chunk(chunk, "responses")
 
@@ -161,6 +170,7 @@ class WrappedResponses:
                     latency,
                     output,
                     None,  # Responses API doesn't have tools
+                    model_from_response,
                 )
 
         return generator()
@@ -177,6 +187,7 @@ class WrappedResponses:
         latency: float,
         output: Any,
         available_tool_calls: Optional[List[Dict[str, Any]]] = None,
+        model_from_response: Optional[str] = None,
     ):
         from posthoganalytics.ai.types import StreamingEventData
         from posthoganalytics.ai.openai.openai_converter import (
@@ -189,9 +200,12 @@ class WrappedResponses:
         formatted_input = format_openai_streaming_input(kwargs, "responses")
         sanitized_input = sanitize_openai_response(formatted_input)
 
+        # Use model from kwargs, fallback to model from response
+        model = kwargs.get("model") or model_from_response or "unknown"
+
         event_data = StreamingEventData(
             provider="openai",
-            model=kwargs.get("model", "unknown"),
+            model=model,
             base_url=str(self._client.base_url),
             kwargs=kwargs,
             formatted_input=sanitized_input,
@@ -320,6 +334,7 @@ class WrappedCompletions:
         usage_stats: TokenUsage = TokenUsage()
         accumulated_content = []
         accumulated_tool_calls: Dict[int, Dict[str, Any]] = {}
+        model_from_response: Optional[str] = None
         if "stream_options" not in kwargs:
             kwargs["stream_options"] = {}
         kwargs["stream_options"]["include_usage"] = True
@@ -329,9 +344,14 @@ class WrappedCompletions:
             nonlocal usage_stats
             nonlocal accumulated_content  # noqa: F824
             nonlocal accumulated_tool_calls
+            nonlocal model_from_response
 
             try:
                 for chunk in response:
+                    # Extract model from chunk (Chat Completions chunks have model field)
+                    if model_from_response is None and hasattr(chunk, "model"):
+                        model_from_response = chunk.model
+
                     # Extract usage stats from chunk
                     chunk_usage = extract_openai_usage_from_chunk(chunk, "chat")
 
@@ -376,6 +396,7 @@ class WrappedCompletions:
                     accumulated_content,
                     tool_calls_list,
                     extract_available_tool_calls("openai", kwargs),
+                    model_from_response,
                 )
 
         return generator()
@@ -393,6 +414,7 @@ class WrappedCompletions:
         output: Any,
         tool_calls: Optional[List[Dict[str, Any]]] = None,
         available_tool_calls: Optional[List[Dict[str, Any]]] = None,
+        model_from_response: Optional[str] = None,
     ):
         from posthoganalytics.ai.types import StreamingEventData
         from posthoganalytics.ai.openai.openai_converter import (
@@ -405,9 +427,12 @@ class WrappedCompletions:
         formatted_input = format_openai_streaming_input(kwargs, "chat")
         sanitized_input = sanitize_openai(formatted_input)
 
+        # Use model from kwargs, fallback to model from response
+        model = kwargs.get("model") or model_from_response or "unknown"
+
         event_data = StreamingEventData(
             provider="openai",
-            model=kwargs.get("model", "unknown"),
+            model=model,
             base_url=str(self._client.base_url),
             kwargs=kwargs,
             formatted_input=sanitized_input,

posthoganalytics/ai/openai/openai_async.py

@@ -128,14 +128,23 @@ class WrappedResponses:
         start_time = time.time()
         usage_stats: TokenUsage = TokenUsage()
         final_content = []
+        model_from_response: Optional[str] = None
         response = await self._original.create(**kwargs)
 
         async def async_generator():
             nonlocal usage_stats
             nonlocal final_content  # noqa: F824
+            nonlocal model_from_response
 
             try:
                 async for chunk in response:
+                    # Extract model from response object in chunk (for stored prompts)
+                    if hasattr(chunk, "response") and chunk.response:
+                        if model_from_response is None and hasattr(
+                            chunk.response, "model"
+                        ):
+                            model_from_response = chunk.response.model
+
                     # Extract usage stats from chunk
                     chunk_usage = extract_openai_usage_from_chunk(chunk, "responses")
 
@@ -166,6 +175,7 @@ class WrappedResponses:
                     latency,
                     output,
                     extract_available_tool_calls("openai", kwargs),
+                    model_from_response,
                 )
 
         return async_generator()
@@ -182,13 +192,17 @@ class WrappedResponses:
         latency: float,
         output: Any,
         available_tool_calls: Optional[List[Dict[str, Any]]] = None,
+        model_from_response: Optional[str] = None,
     ):
         if posthog_trace_id is None:
             posthog_trace_id = str(uuid.uuid4())
 
+        # Use model from kwargs, fallback to model from response
+        model = kwargs.get("model") or model_from_response or "unknown"
+
         event_properties = {
             "$ai_provider": "openai",
-            "$ai_model": kwargs.get("model"),
+            "$ai_model": model,
             "$ai_model_parameters": get_model_params(kwargs),
             "$ai_input": with_privacy_mode(
                 self._client._ph_client,
@@ -350,6 +364,7 @@ class WrappedCompletions:
         usage_stats: TokenUsage = TokenUsage()
         accumulated_content = []
         accumulated_tool_calls: Dict[int, Dict[str, Any]] = {}
+        model_from_response: Optional[str] = None
 
         if "stream_options" not in kwargs:
             kwargs["stream_options"] = {}
@@ -360,9 +375,14 @@ class WrappedCompletions:
             nonlocal usage_stats
             nonlocal accumulated_content  # noqa: F824
             nonlocal accumulated_tool_calls
+            nonlocal model_from_response
 
             try:
                 async for chunk in response:
+                    # Extract model from chunk (Chat Completions chunks have model field)
+                    if model_from_response is None and hasattr(chunk, "model"):
+                        model_from_response = chunk.model
+
                     # Extract usage stats from chunk
                     chunk_usage = extract_openai_usage_from_chunk(chunk, "chat")
                     if chunk_usage:
@@ -405,6 +425,7 @@ class WrappedCompletions:
                     accumulated_content,
                     tool_calls_list,
                     extract_available_tool_calls("openai", kwargs),
+                    model_from_response,
                 )
 
         return async_generator()
@@ -422,13 +443,17 @@ class WrappedCompletions:
         output: Any,
         tool_calls: Optional[List[Dict[str, Any]]] = None,
         available_tool_calls: Optional[List[Dict[str, Any]]] = None,
+        model_from_response: Optional[str] = None,
     ):
         if posthog_trace_id is None:
             posthog_trace_id = str(uuid.uuid4())
 
+        # Use model from kwargs, fallback to model from response
+        model = kwargs.get("model") or model_from_response or "unknown"
+
         event_properties = {
             "$ai_provider": "openai",
-            "$ai_model": kwargs.get("model"),
+            "$ai_model": model,
             "$ai_model_parameters": get_model_params(kwargs),
             "$ai_input": with_privacy_mode(
                 self._client._ph_client,

posthoganalytics/ai/openai/openai_converter.py

@@ -67,6 +67,12 @@ def format_openai_response(response: Any) -> List[FormattedMessage]:
                             }
                         )
 
+                # Handle audio output (gpt-4o-audio-preview)
+                if hasattr(choice.message, "audio") and choice.message.audio:
+                    # Convert Pydantic model to dict to capture all fields from OpenAI
+                    audio_dict = choice.message.audio.model_dump()
+                    content.append({"type": "audio", **audio_dict})
+
         if content:
             output.append(
                 {

posthoganalytics/ai/sanitization.py

@@ -1,3 +1,4 @@
+import os
 import re
 from typing import Any
 from urllib.parse import urlparse
@@ -5,6 +6,15 @@ from urllib.parse import urlparse
 REDACTED_IMAGE_PLACEHOLDER = "[base64 image redacted]"
 
 
+def _is_multimodal_enabled() -> bool:
+    """Check if multimodal capture is enabled via environment variable."""
+    return os.environ.get("_INTERNAL_LLMA_MULTIMODAL", "").lower() in (
+        "true",
+        "1",
+        "yes",
+    )
+
+
 def is_base64_data_url(text: str) -> bool:
     return re.match(r"^data:([^;]+);base64,", text) is not None
 
@@ -27,6 +37,9 @@ def is_raw_base64(text: str) -> bool:
 
 
 def redact_base64_data_url(value: Any) -> Any:
+    if _is_multimodal_enabled():
+        return value
+
     if not isinstance(value, str):
         return value
 
@@ -83,6 +96,11 @@ def sanitize_openai_image(item: Any) -> Any:
             },
         }
 
+    if item.get("type") == "audio" and "data" in item:
+        if _is_multimodal_enabled():
+            return item
+        return {**item, "data": REDACTED_IMAGE_PLACEHOLDER}
+
     return item
 
 
@@ -100,6 +118,9 @@ def sanitize_openai_response_image(item: Any) -> Any:
 
 
 def sanitize_anthropic_image(item: Any) -> Any:
+    if _is_multimodal_enabled():
+        return item
+
     if not isinstance(item, dict):
         return item
 
@@ -109,8 +130,6 @@ def sanitize_anthropic_image(item: Any) -> Any:
         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": {
@@ -123,6 +142,9 @@ def sanitize_anthropic_image(item: Any) -> Any:
 
 
 def sanitize_gemini_part(part: Any) -> Any:
+    if _is_multimodal_enabled():
+        return part
+
     if not isinstance(part, dict):
         return part
 
@@ -131,8 +153,6 @@ def sanitize_gemini_part(part: Any) -> Any:
         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": {
@@ -185,7 +205,9 @@ def sanitize_langchain_image(item: Any) -> Any:
         and isinstance(item.get("source"), dict)
         and "data" in item["source"]
     ):
-        # Anthropic style - raw base64 in structured format, always redact
+        if _is_multimodal_enabled():
+            return item
+
         return {
             **item,
             "source": {

posthoganalytics/ai/utils.py

@@ -285,7 +285,7 @@ def call_llm_and_track_usage(
 
         event_properties = {
             "$ai_provider": provider,
-            "$ai_model": kwargs.get("model"),
+            "$ai_model": kwargs.get("model") or getattr(response, "model", None),
             "$ai_model_parameters": get_model_params(kwargs),
             "$ai_input": with_privacy_mode(
                 ph_client, posthog_privacy_mode, sanitized_messages
@@ -396,7 +396,7 @@ async def call_llm_and_track_usage_async(
 
         event_properties = {
             "$ai_provider": provider,
-            "$ai_model": kwargs.get("model"),
+            "$ai_model": kwargs.get("model") or getattr(response, "model", None),
             "$ai_model_parameters": get_model_params(kwargs),
             "$ai_input": with_privacy_mode(
                 ph_client, posthog_privacy_mode, sanitized_messages