posthoganalytics 6.7.0__py3-none-any.whl → 7.4.3__py3-none-any.whl

posthoganalytics/ai/langchain/callbacks.py

@@ -1,8 +1,8 @@
 try:
-    import langchain  # noqa: F401
+    import langchain_core  # noqa: F401
 except ImportError:
     raise ModuleNotFoundError(
-        "Please install LangChain to use this feature: 'pip install langchain'"
+        "Please install LangChain to use this feature: 'pip install langchain-core'"
     )
 
 import json
@@ -20,8 +20,14 @@ from typing import (
 )
 from uuid import UUID
 
-from langchain.callbacks.base import BaseCallbackHandler
-from langchain.schema.agent import AgentAction, AgentFinish
+try:
+    # LangChain 1.0+ and modern 0.x with langchain-core
+    from langchain_core.callbacks.base import BaseCallbackHandler
+    from langchain_core.agents import AgentAction, AgentFinish
+except (ImportError, ModuleNotFoundError):
+    # Fallback for older LangChain versions
+    from langchain.callbacks.base import BaseCallbackHandler
+    from langchain.schema.agent import AgentAction, AgentFinish
 from langchain_core.documents import Document
 from langchain_core.messages import (
     AIMessage,
@@ -37,6 +43,7 @@ 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")
@@ -72,6 +79,8 @@ class GenerationMetadata(SpanMetadata):
     """Base URL of the provider's API used in the run."""
     tools: Optional[List[Dict[str, Any]]] = None
     """Tools provided to the model."""
+    posthog_properties: Optional[Dict[str, Any]] = None
+    """PostHog properties of the run."""
 
 
 RunMetadata = Union[SpanMetadata, GenerationMetadata]
@@ -413,6 +422,8 @@ class CallbackHandler(BaseCallbackHandler):
                 generation.model = model
             if provider := metadata.get("ls_provider"):
                 generation.provider = provider
+
+            generation.posthog_properties = metadata.get("posthog_properties")
         try:
             base_url = serialized["kwargs"]["openai_api_base"]
             if base_url is not None:
@@ -480,11 +491,12 @@ 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,
             "$ai_span_id": run_id,
+            "$ai_framework": "langchain",
         }
         if parent_run_id is not None:
             event_properties["$ai_parent_id"] = parent_run_id
@@ -550,13 +562,17 @@ class CallbackHandler(BaseCallbackHandler):
             "$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,
             "$ai_base_url": run.base_url,
+            "$ai_framework": "langchain",
         }
 
+        if isinstance(run.posthog_properties, dict):
+            event_properties.update(run.posthog_properties)
+
         if run.tools:
             event_properties["$ai_tools"] = run.tools
 
@@ -566,7 +582,7 @@ class CallbackHandler(BaseCallbackHandler):
             event_properties["$ai_is_error"] = True
         else:
             # Add usage
-            usage = _parse_usage(output)
+            usage = _parse_usage(output, run.provider, run.model)
             event_properties["$ai_input_tokens"] = usage.input_tokens
             event_properties["$ai_output_tokens"] = usage.output_tokens
             event_properties["$ai_cache_creation_input_tokens"] = (
@@ -687,6 +703,8 @@ class ModelUsage:
 
 def _parse_usage_model(
     usage: Union[BaseModel, dict],
+    provider: Optional[str] = None,
+    model: Optional[str] = None,
 ) -> ModelUsage:
     if isinstance(usage, BaseModel):
         usage = usage.__dict__
@@ -749,15 +767,38 @@ def _parse_usage_model(
         "cache_read": "cache_read_tokens",
         "reasoning": "reasoning_tokens",
     }
-    return ModelUsage(
+    normalized_usage = ModelUsage(
         **{
             dataclass_key: parsed_usage.get(mapped_key) or 0
             for mapped_key, dataclass_key in field_mapping.items()
         },
     )
+    # For Anthropic providers, LangChain reports input_tokens as the sum of all input tokens.
+    # Our cost calculation expects them to be separate for Anthropic, so we subtract cache tokens.
+    # Both cache_read and cache_write tokens should be subtracted since Anthropic's raw API
+    # reports input_tokens as tokens NOT read from or used to create a cache.
+    # For other providers (OpenAI, etc.), input_tokens already excludes cache tokens as expected.
+    # Match logic consistent with plugin-server: exact match on provider OR substring match on model
+    is_anthropic = False
+    if provider and provider.lower() == "anthropic":
+        is_anthropic = True
+    elif model and "anthropic" in model.lower():
+        is_anthropic = True
+
+    if is_anthropic and normalized_usage.input_tokens:
+        cache_tokens = (normalized_usage.cache_read_tokens or 0) + (
+            normalized_usage.cache_write_tokens or 0
+        )
+        if cache_tokens > 0:
+            normalized_usage.input_tokens = max(
+                normalized_usage.input_tokens - cache_tokens, 0
+            )
+    return normalized_usage
 
 
-def _parse_usage(response: LLMResult) -> ModelUsage:
+def _parse_usage(
+    response: LLMResult, provider: Optional[str] = None, model: Optional[str] = None
+) -> ModelUsage:
     # langchain-anthropic uses the usage field
     llm_usage_keys = ["token_usage", "usage"]
     llm_usage: ModelUsage = ModelUsage(
@@ -771,13 +812,15 @@ def _parse_usage(response: LLMResult) -> ModelUsage:
     if response.llm_output is not None:
         for key in llm_usage_keys:
             if response.llm_output.get(key):
-                llm_usage = _parse_usage_model(response.llm_output[key])
+                llm_usage = _parse_usage_model(
+                    response.llm_output[key], provider, model
+                )
                 break
 
     if hasattr(response, "generations"):
         for generation in response.generations:
             if "usage" in generation:
-                llm_usage = _parse_usage_model(generation["usage"])
+                llm_usage = _parse_usage_model(generation["usage"], provider, model)
                 break
 
             for generation_chunk in generation:
@@ -785,7 +828,9 @@ def _parse_usage(response: LLMResult) -> ModelUsage:
                     "usage_metadata" in generation_chunk.generation_info
                 ):
                     llm_usage = _parse_usage_model(
-                        generation_chunk.generation_info["usage_metadata"]
+                        generation_chunk.generation_info["usage_metadata"],
+                        provider,
+                        model,
                     )
                     break
 
@@ -812,7 +857,7 @@ def _parse_usage(response: LLMResult) -> ModelUsage:
                     bedrock_anthropic_usage or bedrock_titan_usage or ollama_usage
                 )
                 if chunk_usage:
-                    llm_usage = _parse_usage_model(chunk_usage)
+                    llm_usage = _parse_usage_model(chunk_usage, provider, model)
                     break
 
     return llm_usage

posthoganalytics/ai/openai/__init__.py

@@ -1,5 +1,20 @@
 from .openai import OpenAI
 from .openai_async import AsyncOpenAI
 from .openai_providers import AsyncAzureOpenAI, AzureOpenAI
+from .openai_converter import (
+    format_openai_response,
+    format_openai_input,
+    extract_openai_tools,
+    format_openai_streaming_content,
+)
 
-__all__ = ["OpenAI", "AsyncOpenAI", "AzureOpenAI", "AsyncAzureOpenAI"]
+__all__ = [
+    "OpenAI",
+    "AsyncOpenAI",
+    "AzureOpenAI",
+    "AsyncAzureOpenAI",
+    "format_openai_response",
+    "format_openai_input",
+    "extract_openai_tools",
+    "format_openai_streaming_content",
+]

posthoganalytics/ai/openai/openai.py

@@ -2,6 +2,8 @@ import time
 import uuid
 from typing import Any, Dict, List, Optional
 
+from posthoganalytics.ai.types import TokenUsage
+
 try:
     import openai
 except ImportError:
@@ -12,9 +14,16 @@ except ImportError:
 from posthoganalytics.ai.utils import (
     call_llm_and_track_usage,
     extract_available_tool_calls,
-    get_model_params,
+    merge_usage_stats,
     with_privacy_mode,
 )
+from posthoganalytics.ai.openai.openai_converter import (
+    extract_openai_usage_from_chunk,
+    extract_openai_content_from_chunk,
+    extract_openai_tool_calls_from_chunk,
+    accumulate_openai_tool_calls,
+)
+from posthoganalytics.ai.sanitization import sanitize_openai, sanitize_openai_response
 from posthoganalytics.client import Client as PostHogClient
 from posthoganalytics import setup
 
@@ -33,6 +42,7 @@ class OpenAI(openai.OpenAI):
             posthog_client: If provided, events will be captured via this client instead of the global `posthog`.
             **openai_config: Any additional keyword args to set on openai (e.g. organization="xxx").
         """
+
         super().__init__(**kwargs)
         self._ph_client = posthog_client or setup()
 
@@ -112,45 +122,36 @@ class WrappedResponses:
         **kwargs: Any,
     ):
         start_time = time.time()
-        usage_stats: Dict[str, int] = {}
+        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:
-                    if hasattr(chunk, "type") and chunk.type == "response.completed":
-                        res = chunk.response
-                        if res.output and len(res.output) > 0:
-                            final_content.append(res.output[0])
-
-                    if hasattr(chunk, "usage") and chunk.usage:
-                        usage_stats = {
-                            k: getattr(chunk.usage, k, 0)
-                            for k in [
-                                "input_tokens",
-                                "output_tokens",
-                                "total_tokens",
-                            ]
-                        }
-
-                        # Add support for cached tokens
-                        if hasattr(chunk.usage, "output_tokens_details") and hasattr(
-                            chunk.usage.output_tokens_details, "reasoning_tokens"
+                    # 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"
                         ):
-                            usage_stats["reasoning_tokens"] = (
-                                chunk.usage.output_tokens_details.reasoning_tokens
-                            )
+                            model_from_response = chunk.response.model
 
-                        if hasattr(chunk.usage, "input_tokens_details") and hasattr(
-                            chunk.usage.input_tokens_details, "cached_tokens"
-                        ):
-                            usage_stats["cache_read_input_tokens"] = (
-                                chunk.usage.input_tokens_details.cached_tokens
-                            )
+                    # Extract usage stats from chunk
+                    chunk_usage = extract_openai_usage_from_chunk(chunk, "responses")
+
+                    if chunk_usage:
+                        merge_usage_stats(usage_stats, chunk_usage)
+
+                    # Extract content from chunk
+                    content = extract_openai_content_from_chunk(chunk, "responses")
+
+                    if content is not None:
+                        final_content.append(content)
 
                     yield chunk
 
@@ -168,7 +169,8 @@ class WrappedResponses:
                     usage_stats,
                     latency,
                     output,
-                    extract_available_tool_calls("openai", kwargs),
+                    None,  # Responses API doesn't have tools
+                    model_from_response,
                 )
 
         return generator()
@@ -181,52 +183,44 @@ class WrappedResponses:
         posthog_privacy_mode: bool,
         posthog_groups: Optional[Dict[str, Any]],
         kwargs: Dict[str, Any],
-        usage_stats: Dict[str, int],
+        usage_stats: TokenUsage,
         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())
-
-        event_properties = {
-            "$ai_provider": "openai",
-            "$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")
-            ),
-            "$ai_output_choices": with_privacy_mode(
-                self._client._ph_client,
-                posthog_privacy_mode,
-                output,
-            ),
-            "$ai_http_status": 200,
-            "$ai_input_tokens": usage_stats.get("input_tokens", 0),
-            "$ai_output_tokens": usage_stats.get("output_tokens", 0),
-            "$ai_cache_read_input_tokens": usage_stats.get(
-                "cache_read_input_tokens", 0
-            ),
-            "$ai_reasoning_tokens": usage_stats.get("reasoning_tokens", 0),
-            "$ai_latency": latency,
-            "$ai_trace_id": posthog_trace_id,
-            "$ai_base_url": str(self._client.base_url),
-            **(posthog_properties or {}),
-        }
-
-        if available_tool_calls:
-            event_properties["$ai_tools"] = available_tool_calls
-
-        if posthog_distinct_id is None:
-            event_properties["$process_person_profile"] = False
+        from posthoganalytics.ai.types import StreamingEventData
+        from posthoganalytics.ai.openai.openai_converter import (
+            format_openai_streaming_input,
+            format_openai_streaming_output,
+        )
+        from posthoganalytics.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)
+
+        # 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=model,
+            base_url=str(self._client.base_url),
+            kwargs=kwargs,
+            formatted_input=sanitized_input,
+            formatted_output=format_openai_streaming_output(output, "responses"),
+            usage_stats=usage_stats,
+            latency=latency,
+            distinct_id=posthog_distinct_id,
+            trace_id=posthog_trace_id,
+            properties=posthog_properties,
+            privacy_mode=posthog_privacy_mode,
+            groups=posthog_groups,
+        )
 
-        if hasattr(self._client._ph_client, "capture"):
-            self._client._ph_client.capture(
-                distinct_id=posthog_distinct_id or posthog_trace_id,
-                event="$ai_generation",
-                properties=event_properties,
-                groups=posthog_groups,
-            )
+        # Use the common capture function
+        capture_streaming_event(self._client._ph_client, event_data)
 
     def parse(
         self,
@@ -337,8 +331,10 @@ class WrappedCompletions:
         **kwargs: Any,
     ):
         start_time = time.time()
-        usage_stats: Dict[str, int] = {}
+        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
@@ -347,50 +343,47 @@ class WrappedCompletions:
         def generator():
             nonlocal usage_stats
             nonlocal accumulated_content  # noqa: F824
+            nonlocal accumulated_tool_calls
+            nonlocal model_from_response
 
             try:
                 for chunk in response:
-                    if hasattr(chunk, "usage") and chunk.usage:
-                        usage_stats = {
-                            k: getattr(chunk.usage, k, 0)
-                            for k in [
-                                "prompt_tokens",
-                                "completion_tokens",
-                                "total_tokens",
-                            ]
-                        }
-
-                        # Add support for cached tokens
-                        if hasattr(chunk.usage, "prompt_tokens_details") and hasattr(
-                            chunk.usage.prompt_tokens_details, "cached_tokens"
-                        ):
-                            usage_stats["cache_read_input_tokens"] = (
-                                chunk.usage.prompt_tokens_details.cached_tokens
-                            )
+                    # 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
 
-                        if hasattr(chunk.usage, "output_tokens_details") and hasattr(
-                            chunk.usage.output_tokens_details, "reasoning_tokens"
-                        ):
-                            usage_stats["reasoning_tokens"] = (
-                                chunk.usage.output_tokens_details.reasoning_tokens
-                            )
-
-                    if (
-                        hasattr(chunk, "choices")
-                        and chunk.choices
-                        and len(chunk.choices) > 0
-                    ):
-                        if chunk.choices[0].delta and chunk.choices[0].delta.content:
-                            content = chunk.choices[0].delta.content
-                            if content:
-                                accumulated_content.append(content)
+                    # Extract usage stats from chunk
+                    chunk_usage = extract_openai_usage_from_chunk(chunk, "chat")
+
+                    if chunk_usage:
+                        merge_usage_stats(usage_stats, chunk_usage)
+
+                    # Extract content from chunk
+                    content = extract_openai_content_from_chunk(chunk, "chat")
+
+                    if content is not None:
+                        accumulated_content.append(content)
+
+                    # Extract and accumulate tool calls from chunk
+                    chunk_tool_calls = extract_openai_tool_calls_from_chunk(chunk)
+                    if chunk_tool_calls:
+                        accumulate_openai_tool_calls(
+                            accumulated_tool_calls, chunk_tool_calls
+                        )
 
                     yield chunk
 
             finally:
                 end_time = time.time()
                 latency = end_time - start_time
-                output = "".join(accumulated_content)
+
+                # Convert accumulated tool calls dict to list
+                tool_calls_list = (
+                    list(accumulated_tool_calls.values())
+                    if accumulated_tool_calls
+                    else None
+                )
+
                 self._capture_streaming_event(
                     posthog_distinct_id,
                     posthog_trace_id,
@@ -400,8 +393,10 @@ class WrappedCompletions:
                     kwargs,
                     usage_stats,
                     latency,
-                    output,
+                    accumulated_content,
+                    tool_calls_list,
                     extract_available_tool_calls("openai", kwargs),
+                    model_from_response,
                 )
 
         return generator()
@@ -414,52 +409,45 @@ class WrappedCompletions:
         posthog_privacy_mode: bool,
         posthog_groups: Optional[Dict[str, Any]],
         kwargs: Dict[str, Any],
-        usage_stats: Dict[str, int],
+        usage_stats: TokenUsage,
         latency: float,
         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())
-
-        event_properties = {
-            "$ai_provider": "openai",
-            "$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")
-            ),
-            "$ai_output_choices": with_privacy_mode(
-                self._client._ph_client,
-                posthog_privacy_mode,
-                [{"content": output, "role": "assistant"}],
-            ),
-            "$ai_http_status": 200,
-            "$ai_input_tokens": usage_stats.get("prompt_tokens", 0),
-            "$ai_output_tokens": usage_stats.get("completion_tokens", 0),
-            "$ai_cache_read_input_tokens": usage_stats.get(
-                "cache_read_input_tokens", 0
-            ),
-            "$ai_reasoning_tokens": usage_stats.get("reasoning_tokens", 0),
-            "$ai_latency": latency,
-            "$ai_trace_id": posthog_trace_id,
-            "$ai_base_url": str(self._client.base_url),
-            **(posthog_properties or {}),
-        }
-
-        if available_tool_calls:
-            event_properties["$ai_tools"] = available_tool_calls
-
-        if posthog_distinct_id is None:
-            event_properties["$process_person_profile"] = False
+        from posthoganalytics.ai.types import StreamingEventData
+        from posthoganalytics.ai.openai.openai_converter import (
+            format_openai_streaming_input,
+            format_openai_streaming_output,
+        )
+        from posthoganalytics.ai.utils import capture_streaming_event
+
+        # Prepare standardized event data
+        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=model,
+            base_url=str(self._client.base_url),
+            kwargs=kwargs,
+            formatted_input=sanitized_input,
+            formatted_output=format_openai_streaming_output(output, "chat", tool_calls),
+            usage_stats=usage_stats,
+            latency=latency,
+            distinct_id=posthog_distinct_id,
+            trace_id=posthog_trace_id,
+            properties=posthog_properties,
+            privacy_mode=posthog_privacy_mode,
+            groups=posthog_groups,
+        )
 
-        if hasattr(self._client._ph_client, "capture"):
-            self._client._ph_client.capture(
-                distinct_id=posthog_distinct_id or posthog_trace_id,
-                event="$ai_generation",
-                properties=event_properties,
-                groups=posthog_groups,
-            )
+        # Use the common capture function
+        capture_streaming_event(self._client._ph_client, event_data)
 
 
 class WrappedEmbeddings:
@@ -496,6 +484,7 @@ class WrappedEmbeddings:
         Returns:
             The response from OpenAI's embeddings.create call.
         """
+
         if posthog_trace_id is None:
             posthog_trace_id = str(uuid.uuid4())
 
@@ -518,7 +507,9 @@ 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),