feat: cached tokens (#202)

* feat: cached tokens

* feat: add tool support

* chore: local test

* chore: isort black

* chore: bump v

* chore: remove import

* fix: types

* fix: black

* fix: mypy unpacking of None

* chore: mypy baseline

* feat: mypy fix

* fix: did things and stuff

* fix: mypy yourpy whos py?

* fix: things can be None

* fix: move test

* fix remove exampels from package

* fix: losing my py

CHANGELOG.md

@@ -1,3 +1,11 @@
+## 3.19.0 – 2025-03-04
+
+1. Add support for tool calls in OpenAI and Anthropic.
+2. Add support for cached tokens.
+
+## 3.18.1 – 2025-03-03
+
+1. Improve quota-limited feature flag logs
 
 ## 3.18.0 - 2025-02-28
 

llm_observability_examples.py

@@ -1,221 +0,0 @@
-import asyncio
-import os
-import uuid
-
-from pydantic import BaseModel
-
-import posthog
-from posthog.ai.openai import AsyncOpenAI, OpenAI
-
-# Example credentials - replace these with your own or use environment variables
-posthog.project_api_key = os.getenv("POSTHOG_PROJECT_API_KEY", "your-project-api-key")
-posthog.host = os.getenv("POSTHOG_HOST", "http://localhost:8000")  # Or https://app.posthog.com
-posthog.debug = True
-# change this to False to see usage events
-# posthog.privacy_mode = True
-
-openai_client = OpenAI(
-    api_key=os.getenv("OPENAI_API_KEY", "your-openai-api-key"),
-    posthog_client=posthog,
-)
-
-async_openai_client = AsyncOpenAI(
-    api_key=os.getenv("OPENAI_API_KEY", "your-openai-api-key"),
-    posthog_client=posthog,
-)
-
-
-def main_sync():
-    trace_id = str(uuid.uuid4())
-    print("Trace ID:", trace_id)
-    distinct_id = "test2_distinct_id"
-    properties = {"test_property": "test_value"}
-    groups = {"company": "test_company"}
-
-    try:
-        basic_openai_call(distinct_id, trace_id, properties, groups)
-        # streaming_openai_call(distinct_id, trace_id, properties, groups)
-        # embedding_openai_call(distinct_id, trace_id, properties, groups)
-        # image_openai_call()
-        # beta_openai_call(distinct_id, trace_id, properties, groups)
-    except Exception as e:
-        print("Error during OpenAI call:", str(e))
-
-
-async def main_async():
-    trace_id = str(uuid.uuid4())
-    print("Trace ID:", trace_id)
-    distinct_id = "test_distinct_id"
-    properties = {"test_property": "test_value"}
-    groups = {"company": "test_company"}
-
-    try:
-        await basic_async_openai_call(distinct_id, trace_id, properties, groups)
-        await streaming_async_openai_call(distinct_id, trace_id, properties, groups)
-        await embedding_async_openai_call(distinct_id, trace_id, properties, groups)
-        await image_async_openai_call()
-    except Exception as e:
-        print("Error during OpenAI call:", str(e))
-
-
-def basic_openai_call(distinct_id, trace_id, properties, groups):
-    response = openai_client.chat.completions.create(
-        model="gpt-4o-mini",
-        messages=[
-            {"role": "system", "content": "You are a complex problem solver."},
-            {"role": "user", "content": "Explain quantum computing in simple terms."},
-        ],
-        max_tokens=100,
-        temperature=0.7,
-        posthog_distinct_id=distinct_id,
-        posthog_trace_id=trace_id,
-        posthog_properties=properties,
-        posthog_groups=groups,
-    )
-    print(response)
-    if response and response.choices:
-        print("OpenAI response:", response.choices[0].message.content)
-    else:
-        print("No response or unexpected format returned.")
-    return response
-
-
-async def basic_async_openai_call(distinct_id, trace_id, properties, groups):
-    response = await async_openai_client.chat.completions.create(
-        model="gpt-4o-mini",
-        messages=[
-            {"role": "system", "content": "You are a complex problem solver."},
-            {"role": "user", "content": "Explain quantum computing in simple terms."},
-        ],
-        max_tokens=100,
-        temperature=0.7,
-        posthog_distinct_id=distinct_id,
-        posthog_trace_id=trace_id,
-        posthog_properties=properties,
-        posthog_groups=groups,
-    )
-    if response and hasattr(response, "choices"):
-        print("OpenAI response:", response.choices[0].message.content)
-    else:
-        print("No response or unexpected format returned.")
-    return response
-
-
-def streaming_openai_call(distinct_id, trace_id, properties, groups):
-
-    response = openai_client.chat.completions.create(
-        model="gpt-4o-mini",
-        messages=[
-            {"role": "system", "content": "You are a complex problem solver."},
-            {"role": "user", "content": "Explain quantum computing in simple terms."},
-        ],
-        max_tokens=100,
-        temperature=0.7,
-        stream=True,
-        posthog_distinct_id=distinct_id,
-        posthog_trace_id=trace_id,
-        posthog_properties=properties,
-        posthog_groups=groups,
-    )
-
-    for chunk in response:
-        if hasattr(chunk, "choices") and chunk.choices and len(chunk.choices) > 0:
-            print(chunk.choices[0].delta.content or "", end="")
-
-    return response
-
-
-async def streaming_async_openai_call(distinct_id, trace_id, properties, groups):
-    response = await async_openai_client.chat.completions.create(
-        model="gpt-4o-mini",
-        messages=[
-            {"role": "system", "content": "You are a complex problem solver."},
-            {"role": "user", "content": "Explain quantum computing in simple terms."},
-        ],
-        max_tokens=100,
-        temperature=0.7,
-        stream=True,
-        posthog_distinct_id=distinct_id,
-        posthog_trace_id=trace_id,
-        posthog_properties=properties,
-        posthog_groups=groups,
-    )
-
-    async for chunk in response:
-        if hasattr(chunk, "choices") and chunk.choices and len(chunk.choices) > 0:
-            print(chunk.choices[0].delta.content or "", end="")
-
-    return response
-
-
-# none instrumented
-def image_openai_call():
-    response = openai_client.images.generate(model="dall-e-3", prompt="A cute baby hedgehog", n=1, size="1024x1024")
-    print(response)
-    return response
-
-
-# none instrumented
-async def image_async_openai_call():
-    response = await async_openai_client.images.generate(
-        model="dall-e-3", prompt="A cute baby hedgehog", n=1, size="1024x1024"
-    )
-    print(response)
-    return response
-
-
-def embedding_openai_call(posthog_distinct_id, posthog_trace_id, posthog_properties, posthog_groups):
-    response = openai_client.embeddings.create(
-        input="The hedgehog is cute",
-        model="text-embedding-3-small",
-        posthog_distinct_id=posthog_distinct_id,
-        posthog_trace_id=posthog_trace_id,
-        posthog_properties=posthog_properties,
-        posthog_groups=posthog_groups,
-    )
-    print(response)
-    return response
-
-
-async def embedding_async_openai_call(posthog_distinct_id, posthog_trace_id, posthog_properties, posthog_groups):
-    response = await async_openai_client.embeddings.create(
-        input="The hedgehog is cute",
-        model="text-embedding-3-small",
-        posthog_distinct_id=posthog_distinct_id,
-        posthog_trace_id=posthog_trace_id,
-        posthog_properties=posthog_properties,
-        posthog_groups=posthog_groups,
-    )
-    print(response)
-    return response
-
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-
-def beta_openai_call(distinct_id, trace_id, properties, groups):
-    response = openai_client.beta.chat.completions.parse(
-        model="gpt-4o-mini",
-        messages=[
-            {"role": "system", "content": "Extract the event information."},
-            {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-        ],
-        response_format=CalendarEvent,
-        posthog_distinct_id=distinct_id,
-        posthog_trace_id=trace_id,
-        posthog_properties=properties,
-        posthog_groups=groups,
-    )
-    print(response)
-    return response
-
-
-# HOW TO RUN:
-# comment out one of these to run the other
-
-if __name__ == "__main__":
-    main_sync()
-# asyncio.run(main_async())

mypy-baseline.txt

@@ -43,19 +43,13 @@ sentry_django_example/sentry_django_example/settings.py:0: error: Need type anno
 sentry_django_example/sentry_django_example/settings.py:0: error: Incompatible types in assignment (expression has type "str", variable has type "None")  [assignment]
 posthog/ai/openai/openai_async.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
 posthog/ai/openai/openai_async.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
-posthog/ai/openai/openai_async.py:0: error: Unpacked dict entry 11 has incompatible type "dict[str, Any] | None"; expected "SupportsKeysAndGetItem[str, Any]"  [dict-item]
 posthog/ai/openai/openai_async.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
-posthog/ai/openai/openai_async.py:0: error: Unpacked dict entry 8 has incompatible type "dict[str, Any] | None"; expected "SupportsKeysAndGetItem[str, Any]"  [dict-item]
 posthog/ai/openai/openai.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
 posthog/ai/openai/openai.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
-posthog/ai/openai/openai.py:0: error: Unpacked dict entry 11 has incompatible type "dict[str, Any] | None"; expected "SupportsKeysAndGetItem[str, Any]"  [dict-item]
 posthog/ai/openai/openai.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
-posthog/ai/openai/openai.py:0: error: Unpacked dict entry 8 has incompatible type "dict[str, Any] | None"; expected "SupportsKeysAndGetItem[str, Any]"  [dict-item]
 posthog/ai/anthropic/anthropic_async.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
 posthog/ai/anthropic/anthropic_async.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
 posthog/ai/anthropic/anthropic_async.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
 posthog/ai/anthropic/anthropic.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
 posthog/ai/anthropic/anthropic.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
 posthog/ai/anthropic/anthropic.py:0: error: Incompatible types in assignment (expression has type "UUID", variable has type "str | None")  [assignment]
-llm_observability_examples.py:0: error: Argument "posthog_client" to "OpenAI" has incompatible type Module; expected "Client"  [arg-type]
-llm_observability_examples.py:0: error: Argument "posthog_client" to "AsyncOpenAI" has incompatible type Module; expected "Client"  [arg-type]

posthog/ai/anthropic/anthropic.py

@@ -125,6 +125,8 @@ class WrappedMessages(Messages):
                             for k in [
                                 "input_tokens",
                                 "output_tokens",
+                                "cache_read_input_tokens",
+                                "cache_creation_input_tokens",
                             ]
                         }
 
@@ -184,6 +186,8 @@ class WrappedMessages(Messages):
             "$ai_http_status": 200,
             "$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),
+            "$ai_cache_creation_input_tokens": usage_stats.get("cache_creation_input_tokens", 0),
             "$ai_latency": latency,
             "$ai_trace_id": posthog_trace_id,
             "$ai_base_url": str(self._client.base_url),

posthog/ai/anthropic/anthropic_async.py

@@ -125,6 +125,8 @@ class AsyncWrappedMessages(AsyncMessages):
                             for k in [
                                 "input_tokens",
                                 "output_tokens",
+                                "cache_read_input_tokens",
+                                "cache_creation_input_tokens",
                             ]
                         }
 
@@ -184,6 +186,8 @@ class AsyncWrappedMessages(AsyncMessages):
             "$ai_http_status": 200,
             "$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),
+            "$ai_cache_creation_input_tokens": usage_stats.get("cache_creation_input_tokens", 0),
             "$ai_latency": latency,
             "$ai_trace_id": posthog_trace_id,
             "$ai_base_url": str(self._client.base_url),

posthog/ai/langchain/callbacks.py

@@ -516,7 +516,11 @@ class CallbackHandler(BaseCallbackHandler):
             "$ai_base_url": run.base_url,
         }
         if run.tools:
-            event_properties["$ai_tools"] = run.tools
+            event_properties["$ai_tools"] = with_privacy_mode(
+                self._client,
+                self._privacy_mode,
+                run.tools,
+            )
 
         if isinstance(output, BaseException):
             event_properties["$ai_http_status"] = _get_http_status(output)

posthog/ai/openai/openai.py

@@ -1,6 +1,6 @@
 import time
 import uuid
-from typing import Any, Dict, Optional
+from typing import Any, Dict, List, Optional
 
 try:
     import openai
@@ -92,6 +92,7 @@ class WrappedCompletions(openai.resources.chat.completions.Completions):
         start_time = time.time()
         usage_stats: Dict[str, int] = {}
         accumulated_content = []
+        accumulated_tools = {}
         if "stream_options" not in kwargs:
             kwargs["stream_options"] = {}
         kwargs["stream_options"]["include_usage"] = True
@@ -100,6 +101,8 @@ class WrappedCompletions(openai.resources.chat.completions.Completions):
         def generator():
             nonlocal usage_stats
             nonlocal accumulated_content
+            nonlocal accumulated_tools
+
             try:
                 for chunk in response:
                     if hasattr(chunk, "usage") and chunk.usage:
@@ -112,17 +115,36 @@ class WrappedCompletions(openai.resources.chat.completions.Completions):
                             ]
                         }
 
+                        # 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
+
                     if hasattr(chunk, "choices") and chunk.choices and len(chunk.choices) > 0:
                         content = chunk.choices[0].delta.content
                         if content:
                             accumulated_content.append(content)
 
+                        # Process tool calls
+                        tool_calls = getattr(chunk.choices[0].delta, "tool_calls", None)
+                        if tool_calls:
+                            for tool_call in tool_calls:
+                                index = tool_call.index
+                                if index not in accumulated_tools:
+                                    accumulated_tools[index] = tool_call
+                                else:
+                                    # Append arguments for existing tool calls
+                                    if hasattr(tool_call, "function") and hasattr(tool_call.function, "arguments"):
+                                        accumulated_tools[index].function.arguments += tool_call.function.arguments
+
                     yield chunk
 
             finally:
                 end_time = time.time()
                 latency = end_time - start_time
                 output = "".join(accumulated_content)
+                tools = list(accumulated_tools.values()) if accumulated_tools else None
                 self._capture_streaming_event(
                     posthog_distinct_id,
                     posthog_trace_id,
@@ -133,6 +155,7 @@ class WrappedCompletions(openai.resources.chat.completions.Completions):
                     usage_stats,
                     latency,
                     output,
+                    tools,
                 )
 
         return generator()
@@ -148,6 +171,7 @@ class WrappedCompletions(openai.resources.chat.completions.Completions):
         usage_stats: Dict[str, int],
         latency: float,
         output: str,
+        tool_calls: Optional[List[Dict[str, Any]]] = None,
     ):
         if posthog_trace_id is None:
             posthog_trace_id = uuid.uuid4()
@@ -165,12 +189,20 @@ class WrappedCompletions(openai.resources.chat.completions.Completions):
             "$ai_http_status": 200,
             "$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),
             "$ai_latency": latency,
             "$ai_trace_id": posthog_trace_id,
             "$ai_base_url": str(self._client.base_url),
-            **posthog_properties,
+            **(posthog_properties or {}),
         }
 
+        if tool_calls:
+            event_properties["$ai_tools"] = with_privacy_mode(
+                self._client._ph_client,
+                posthog_privacy_mode,
+                tool_calls,
+            )
+
         if posthog_distinct_id is None:
             event_properties["$process_person_profile"] = False
 
@@ -234,7 +266,7 @@ class WrappedEmbeddings(openai.resources.embeddings.Embeddings):
             "$ai_latency": latency,
             "$ai_trace_id": posthog_trace_id,
             "$ai_base_url": str(self._client.base_url),
-            **posthog_properties,
+            **(posthog_properties or {}),
         }
 
         if posthog_distinct_id is None:

posthog/ai/openai/openai_async.py

@@ -1,6 +1,6 @@
 import time
 import uuid
-from typing import Any, Dict, Optional
+from typing import Any, Dict, List, Optional
 
 try:
     import openai
@@ -93,13 +93,14 @@ class WrappedCompletions(openai.resources.chat.completions.AsyncCompletions):
         start_time = time.time()
         usage_stats: Dict[str, int] = {}
         accumulated_content = []
+        accumulated_tools = {}
         if "stream_options" not in kwargs:
             kwargs["stream_options"] = {}
         kwargs["stream_options"]["include_usage"] = True
         response = await super().create(**kwargs)
 
         async def async_generator():
-            nonlocal usage_stats, accumulated_content
+            nonlocal usage_stats, accumulated_content, accumulated_tools
             try:
                 async for chunk in response:
                     if hasattr(chunk, "usage") and chunk.usage:
@@ -111,17 +112,37 @@ class WrappedCompletions(openai.resources.chat.completions.AsyncCompletions):
                                 "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
+
                     if hasattr(chunk, "choices") and chunk.choices and len(chunk.choices) > 0:
                         content = chunk.choices[0].delta.content
                         if content:
                             accumulated_content.append(content)
 
+                        # Process tool calls
+                        tool_calls = getattr(chunk.choices[0].delta, "tool_calls", None)
+                        if tool_calls:
+                            for tool_call in tool_calls:
+                                index = tool_call.index
+                                if index not in accumulated_tools:
+                                    accumulated_tools[index] = tool_call
+                                else:
+                                    # Append arguments for existing tool calls
+                                    if hasattr(tool_call, "function") and hasattr(tool_call.function, "arguments"):
+                                        accumulated_tools[index].function.arguments += tool_call.function.arguments
+
                     yield chunk
 
             finally:
                 end_time = time.time()
                 latency = end_time - start_time
                 output = "".join(accumulated_content)
+                tools = list(accumulated_tools.values()) if accumulated_tools else None
                 await self._capture_streaming_event(
                     posthog_distinct_id,
                     posthog_trace_id,
@@ -132,6 +153,7 @@ class WrappedCompletions(openai.resources.chat.completions.AsyncCompletions):
                     usage_stats,
                     latency,
                     output,
+                    tools,
                 )
 
         return async_generator()
@@ -147,6 +169,7 @@ class WrappedCompletions(openai.resources.chat.completions.AsyncCompletions):
         usage_stats: Dict[str, int],
         latency: float,
         output: str,
+        tool_calls: Optional[List[Dict[str, Any]]] = None,
     ):
         if posthog_trace_id is None:
             posthog_trace_id = uuid.uuid4()
@@ -164,12 +187,20 @@ class WrappedCompletions(openai.resources.chat.completions.AsyncCompletions):
             "$ai_http_status": 200,
             "$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),
             "$ai_latency": latency,
             "$ai_trace_id": posthog_trace_id,
             "$ai_base_url": str(self._client.base_url),
-            **posthog_properties,
+            **(posthog_properties or {}),
         }
 
+        if tool_calls:
+            event_properties["$ai_tools"] = with_privacy_mode(
+                self._client._ph_client,
+                posthog_privacy_mode,
+                tool_calls,
+            )
+
         if posthog_distinct_id is None:
             event_properties["$process_person_profile"] = False
 
@@ -235,7 +266,7 @@ class WrappedEmbeddings(openai.resources.embeddings.AsyncEmbeddings):
             "$ai_latency": latency,
             "$ai_trace_id": posthog_trace_id,
             "$ai_base_url": str(self._client.base_url),
-            **posthog_properties,
+            **(posthog_properties or {}),
         }
 
         if posthog_distinct_id is None:

posthog/ai/utils.py

@@ -34,15 +34,25 @@ def get_usage(response, provider: str) -> Dict[str, Any]:
         return {
             "input_tokens": response.usage.input_tokens,
             "output_tokens": response.usage.output_tokens,
+            "cache_read_input_tokens": response.usage.cache_read_input_tokens,
+            "cache_creation_input_tokens": response.usage.cache_creation_input_tokens,
         }
     elif provider == "openai":
+        cached_tokens = 0
+        if hasattr(response.usage, "prompt_tokens_details") and hasattr(
+            response.usage.prompt_tokens_details, "cached_tokens"
+        ):
+            cached_tokens = response.usage.prompt_tokens_details.cached_tokens
         return {
             "input_tokens": response.usage.prompt_tokens,
             "output_tokens": response.usage.completion_tokens,
+            "cache_read_input_tokens": cached_tokens,
         }
     return {
         "input_tokens": 0,
         "output_tokens": 0,
+        "cache_read_input_tokens": 0,
+        "cache_creation_input_tokens": 0,
     }
 
 
@@ -86,6 +96,21 @@ def format_response_openai(response):
     return output
 
 
+def format_tool_calls(response, provider: str):
+    if provider == "anthropic":
+        if hasattr(response, "tools") and response.tools and len(response.tools) > 0:
+            return response.tools
+    elif provider == "openai":
+        if (
+            hasattr(response, "choices")
+            and response.choices
+            and hasattr(response.choices[0].message, "tool_calls")
+            and response.choices[0].message.tool_calls
+        ):
+            return response.choices[0].message.tool_calls
+    return None
+
+
 def merge_system_prompt(kwargs: Dict[str, Any], provider: str):
     if provider != "anthropic":
         return kwargs.get("messages")
@@ -157,6 +182,16 @@ def call_llm_and_track_usage(
             **(error_params or {}),
         }
 
+        tool_calls = format_tool_calls(response, provider)
+        if tool_calls:
+            event_properties["$ai_tools"] = with_privacy_mode(ph_client, posthog_privacy_mode, tool_calls)
+
+        if usage.get("cache_read_input_tokens") is not None and usage.get("cache_read_input_tokens", 0) > 0:
+            event_properties["$ai_cache_read_input_tokens"] = usage.get("cache_read_input_tokens", 0)
+
+        if usage.get("cache_creation_input_tokens") is not None and usage.get("cache_creation_input_tokens", 0) > 0:
+            event_properties["$ai_cache_creation_input_tokens"] = usage.get("cache_creation_input_tokens", 0)
+
         if posthog_distinct_id is None:
             event_properties["$process_person_profile"] = False
 
@@ -233,6 +268,16 @@ async def call_llm_and_track_usage_async(
             **(error_params or {}),
         }
 
+        tool_calls = format_tool_calls(response, provider)
+        if tool_calls:
+            event_properties["$ai_tools"] = with_privacy_mode(ph_client, posthog_privacy_mode, tool_calls)
+
+        if usage.get("cache_read_input_tokens") is not None and usage.get("cache_read_input_tokens", 0) > 0:
+            event_properties["$ai_cache_read_input_tokens"] = usage.get("cache_read_input_tokens", 0)
+
+        if usage.get("cache_creation_input_tokens") is not None and usage.get("cache_creation_input_tokens", 0) > 0:
+            event_properties["$ai_cache_creation_input_tokens"] = usage.get("cache_creation_input_tokens", 0)
+
         if posthog_distinct_id is None:
             event_properties["$process_person_profile"] = False
 

posthog/client.py

@@ -659,7 +659,9 @@ class Client(object):
                         "More information: https://posthog.com/docs/api/overview",
                     )
             elif e.status == 402:
-                self.log.warning("[FEATURE FLAGS] PostHog feature flags quota limited")
+                self.log.warning(
+                    "[FEATURE FLAGS] PostHog feature flags quota limited, resetting feature flag data.  Learn more about billing limits at https://posthog.com/docs/billing/limits-alerts"
+                )
                 # Reset all feature flag data when quota limited
                 self.feature_flags = []
                 self.feature_flags_by_key = {}

posthog/request.py

@@ -78,7 +78,9 @@ def _process_response(
             and isinstance(response["quotaLimited"], list)
             and "feature_flags" in response["quotaLimited"]
         ):
-            log.warning("PostHog feature flags quota limited")
+            log.warning(
+                "[FEATURE FLAGS] PostHog feature flags quota limited, resetting feature flag data.  Learn more about billing limits at https://posthog.com/docs/billing/limits-alerts"
+            )
             raise QuotaLimitError(res.status_code, "Feature flags quota limited")
         return response
     try:

posthog/test/ai/anthropic/test_anthropic.py

@@ -55,6 +55,28 @@ def mock_anthropic_stream():
     return stream_generator()
 
 
+@pytest.fixture
+def mock_anthropic_response_with_cached_tokens():
+    # Create a mock Usage object with cached_tokens in input_tokens_details
+    usage = Usage(
+        input_tokens=20,
+        output_tokens=10,
+        cache_read_input_tokens=15,
+        cache_creation_input_tokens=2,
+    )
+
+    return Message(
+        id="msg_123",
+        type="message",
+        role="assistant",
+        content=[{"type": "text", "text": "Test response"}],
+        model="claude-3-opus-20240229",
+        usage=usage,
+        stop_reason="end_turn",
+        stop_sequence=None,
+    )
+
+
 def test_basic_completion(mock_client, mock_anthropic_response):
     with patch("anthropic.resources.Messages.create", return_value=mock_anthropic_response):
         client = Anthropic(api_key="test-key", posthog_client=mock_client)
@@ -339,3 +361,34 @@ def test_error(mock_client, mock_anthropic_response):
         props = call_args["properties"]
         assert props["$ai_is_error"] is True
         assert props["$ai_error"] == "Test error"
+
+
+def test_cached_tokens(mock_client, mock_anthropic_response_with_cached_tokens):
+    with patch("anthropic.resources.Messages.create", return_value=mock_anthropic_response_with_cached_tokens):
+        client = Anthropic(api_key="test-key", posthog_client=mock_client)
+        response = client.messages.create(
+            model="claude-3-opus-20240229",
+            messages=[{"role": "user", "content": "Hello"}],
+            posthog_distinct_id="test-id",
+            posthog_properties={"foo": "bar"},
+        )
+
+        assert response == mock_anthropic_response_with_cached_tokens
+        assert mock_client.capture.call_count == 1
+
+        call_args = mock_client.capture.call_args[1]
+        props = call_args["properties"]
+
+        assert call_args["distinct_id"] == "test-id"
+        assert call_args["event"] == "$ai_generation"
+        assert props["$ai_provider"] == "anthropic"
+        assert props["$ai_model"] == "claude-3-opus-20240229"
+        assert props["$ai_input"] == [{"role": "user", "content": "Hello"}]
+        assert props["$ai_output_choices"] == [{"role": "assistant", "content": "Test response"}]
+        assert props["$ai_input_tokens"] == 20
+        assert props["$ai_output_tokens"] == 10
+        assert props["$ai_cache_read_input_tokens"] == 15
+        assert props["$ai_cache_creation_input_tokens"] == 2
+        assert props["$ai_http_status"] == 200
+        assert props["foo"] == "bar"
+        assert isinstance(props["$ai_latency"], float)

posthog/test/ai/openai/test_openai.py

@@ -1,9 +1,14 @@
+import json
 import time
 from unittest.mock import patch
 
 import pytest
 from openai.types.chat import ChatCompletion, ChatCompletionMessage
 from openai.types.chat.chat_completion import Choice
+from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
+from openai.types.chat.chat_completion_chunk import Choice as ChoiceChunk
+from openai.types.chat.chat_completion_chunk import ChoiceDelta, ChoiceDeltaToolCall, ChoiceDeltaToolCallFunction
+from openai.types.chat.chat_completion_message_tool_call import ChatCompletionMessageToolCall, Function
 from openai.types.completion_usage import CompletionUsage
 from openai.types.create_embedding_response import CreateEmbeddingResponse, Usage
 from openai.types.embedding import Embedding
@@ -62,6 +67,67 @@ def mock_embedding_response():
     )
 
 
+@pytest.fixture
+def mock_openai_response_with_cached_tokens():
+    return ChatCompletion(
+        id="test",
+        model="gpt-4",
+        object="chat.completion",
+        created=int(time.time()),
+        choices=[
+            Choice(
+                finish_reason="stop",
+                index=0,
+                message=ChatCompletionMessage(
+                    content="Test response",
+                    role="assistant",
+                ),
+            )
+        ],
+        usage=CompletionUsage(
+            completion_tokens=10,
+            prompt_tokens=20,
+            total_tokens=30,
+            prompt_tokens_details={"cached_tokens": 15},
+        ),
+    )
+
+
+@pytest.fixture
+def mock_openai_response_with_tool_calls():
+    return ChatCompletion(
+        id="test",
+        model="gpt-4",
+        object="chat.completion",
+        created=int(time.time()),
+        choices=[
+            Choice(
+                finish_reason="tool_calls",
+                index=0,
+                message=ChatCompletionMessage(
+                    content="I'll check the weather for you.",
+                    role="assistant",
+                    tool_calls=[
+                        ChatCompletionMessageToolCall(
+                            id="call_abc123",
+                            type="function",
+                            function=Function(
+                                name="get_weather",
+                                arguments='{"location": "San Francisco", "unit": "celsius"}',
+                            ),
+                        )
+                    ],
+                ),
+            )
+        ],
+        usage=CompletionUsage(
+            completion_tokens=15,
+            prompt_tokens=20,
+            total_tokens=35,
+        ),
+    )
+
+
 def test_basic_completion(mock_client, mock_openai_response):
     with patch("openai.resources.chat.completions.Completions.create", return_value=mock_openai_response):
         client = OpenAI(api_key="test-key", posthog_client=mock_client)
@@ -187,3 +253,249 @@ def test_error(mock_client, mock_openai_response):
         props = call_args["properties"]
         assert props["$ai_is_error"] is True
         assert props["$ai_error"] == "Test error"
+
+
+def test_cached_tokens(mock_client, mock_openai_response_with_cached_tokens):
+    with patch(
+        "openai.resources.chat.completions.Completions.create", return_value=mock_openai_response_with_cached_tokens
+    ):
+        client = OpenAI(api_key="test-key", posthog_client=mock_client)
+        response = client.chat.completions.create(
+            model="gpt-4",
+            messages=[{"role": "user", "content": "Hello"}],
+            posthog_distinct_id="test-id",
+            posthog_properties={"foo": "bar"},
+        )
+
+        assert response == mock_openai_response_with_cached_tokens
+        assert mock_client.capture.call_count == 1
+
+        call_args = mock_client.capture.call_args[1]
+        props = call_args["properties"]
+
+        assert call_args["distinct_id"] == "test-id"
+        assert call_args["event"] == "$ai_generation"
+        assert props["$ai_provider"] == "openai"
+        assert props["$ai_model"] == "gpt-4"
+        assert props["$ai_input"] == [{"role": "user", "content": "Hello"}]
+        assert props["$ai_output_choices"] == [{"role": "assistant", "content": "Test response"}]
+        assert props["$ai_input_tokens"] == 20
+        assert props["$ai_output_tokens"] == 10
+        assert props["$ai_cache_read_input_tokens"] == 15
+        assert props["$ai_http_status"] == 200
+        assert props["foo"] == "bar"
+        assert isinstance(props["$ai_latency"], float)
+
+
+def test_tool_calls(mock_client, mock_openai_response_with_tool_calls):
+    with patch(
+        "openai.resources.chat.completions.Completions.create", return_value=mock_openai_response_with_tool_calls
+    ):
+        client = OpenAI(api_key="test-key", posthog_client=mock_client)
+        response = client.chat.completions.create(
+            model="gpt-4",
+            messages=[{"role": "user", "content": "What's the weather in San Francisco?"}],
+            tools=[
+                {
+                    "type": "function",
+                    "function": {"name": "get_weather", "description": "Get weather", "parameters": {}},
+                }
+            ],
+            posthog_distinct_id="test-id",
+        )
+
+        assert response == mock_openai_response_with_tool_calls
+        assert mock_client.capture.call_count == 1
+
+        call_args = mock_client.capture.call_args[1]
+        props = call_args["properties"]
+
+        assert call_args["distinct_id"] == "test-id"
+        assert call_args["event"] == "$ai_generation"
+        assert props["$ai_provider"] == "openai"
+        assert props["$ai_model"] == "gpt-4"
+        assert props["$ai_input"] == [{"role": "user", "content": "What's the weather in San Francisco?"}]
+        assert props["$ai_output_choices"] == [{"role": "assistant", "content": "I'll check the weather for you."}]
+
+        # Check that tool calls are properly captured
+        assert "$ai_tools" in props
+        tool_calls = props["$ai_tools"]
+        assert len(tool_calls) == 1
+
+        # Verify the tool call details
+        tool_call = tool_calls[0]
+        assert tool_call.id == "call_abc123"
+        assert tool_call.type == "function"
+        assert tool_call.function.name == "get_weather"
+
+        # Verify the arguments
+        arguments = tool_call.function.arguments
+        parsed_args = json.loads(arguments)
+        assert parsed_args == {"location": "San Francisco", "unit": "celsius"}
+
+        # Check token usage
+        assert props["$ai_input_tokens"] == 20
+        assert props["$ai_output_tokens"] == 15
+        assert props["$ai_http_status"] == 200
+
+
+def test_streaming_with_tool_calls(mock_client):
+    # Create mock tool call chunks that will be returned in sequence
+    tool_call_chunks = [
+        ChatCompletionChunk(
+            id="chunk1",
+            model="gpt-4",
+            object="chat.completion.chunk",
+            created=1234567890,
+            choices=[
+                ChoiceChunk(
+                    index=0,
+                    delta=ChoiceDelta(
+                        role="assistant",
+                        tool_calls=[
+                            ChoiceDeltaToolCall(
+                                index=0,
+                                id="call_abc123",
+                                type="function",
+                                function=ChoiceDeltaToolCallFunction(
+                                    name="get_weather",
+                                    arguments='{"location": "',
+                                ),
+                            )
+                        ],
+                    ),
+                    finish_reason=None,
+                )
+            ],
+        ),
+        ChatCompletionChunk(
+            id="chunk2",
+            model="gpt-4",
+            object="chat.completion.chunk",
+            created=1234567891,
+            choices=[
+                ChoiceChunk(
+                    index=0,
+                    delta=ChoiceDelta(
+                        tool_calls=[
+                            ChoiceDeltaToolCall(
+                                index=0,
+                                id="call_abc123",
+                                type="function",
+                                function=ChoiceDeltaToolCallFunction(
+                                    arguments='San Francisco"',
+                                ),
+                            )
+                        ],
+                    ),
+                    finish_reason=None,
+                )
+            ],
+        ),
+        ChatCompletionChunk(
+            id="chunk3",
+            model="gpt-4",
+            object="chat.completion.chunk",
+            created=1234567892,
+            choices=[
+                ChoiceChunk(
+                    index=0,
+                    delta=ChoiceDelta(
+                        tool_calls=[
+                            ChoiceDeltaToolCall(
+                                index=0,
+                                id="call_abc123",
+                                type="function",
+                                function=ChoiceDeltaToolCallFunction(
+                                    arguments=', "unit": "celsius"}',
+                                ),
+                            )
+                        ],
+                    ),
+                    finish_reason=None,
+                )
+            ],
+        ),
+        ChatCompletionChunk(
+            id="chunk4",
+            model="gpt-4",
+            object="chat.completion.chunk",
+            created=1234567893,
+            choices=[
+                ChoiceChunk(
+                    index=0,
+                    delta=ChoiceDelta(
+                        content="The weather in San Francisco is 15°C.",
+                    ),
+                    finish_reason=None,
+                )
+            ],
+            usage=CompletionUsage(
+                prompt_tokens=20,
+                completion_tokens=15,
+                total_tokens=35,
+            ),
+        ),
+    ]
+
+    # Mock the create method to return our chunks
+    with patch("openai.resources.chat.completions.Completions.create") as mock_create:
+        # Set up the mock to return our chunks when iterated
+        mock_create.return_value = tool_call_chunks
+
+        client = OpenAI(api_key="test-key", posthog_client=mock_client)
+
+        # Call the streaming method
+        response_generator = client.chat.completions.create(
+            model="gpt-4",
+            messages=[{"role": "user", "content": "What's the weather in San Francisco?"}],
+            tools=[
+                {
+                    "type": "function",
+                    "function": {"name": "get_weather", "description": "Get weather", "parameters": {}},
+                }
+            ],
+            stream=True,
+            posthog_distinct_id="test-id",
+        )
+
+        # Consume the generator to trigger the event capture
+        chunks = list(response_generator)
+
+        # Verify the chunks were returned correctly
+        assert len(chunks) == 4
+        assert chunks == tool_call_chunks
+
+        # Verify the capture was called with the right arguments
+        assert mock_client.capture.call_count == 1
+
+        call_args = mock_client.capture.call_args[1]
+        props = call_args["properties"]
+
+        assert call_args["distinct_id"] == "test-id"
+        assert call_args["event"] == "$ai_generation"
+        assert props["$ai_provider"] == "openai"
+        assert props["$ai_model"] == "gpt-4"
+
+        # Check that the tool calls were properly accumulated
+        assert "$ai_tools" in props
+        tool_calls = props["$ai_tools"]
+        assert len(tool_calls) == 1
+
+        # Verify the complete tool call was properly assembled
+        tool_call = tool_calls[0]
+        assert tool_call.id == "call_abc123"
+        assert tool_call.type == "function"
+        assert tool_call.function.name == "get_weather"
+
+        # Verify the arguments were concatenated correctly
+        arguments = tool_call.function.arguments
+        parsed_args = json.loads(arguments)
+        assert parsed_args == {"location": "San Francisco", "unit": "celsius"}
+
+        # Check that the content was also accumulated
+        assert props["$ai_output_choices"][0]["content"] == "The weather in San Francisco is 15°C."
+
+        # Check token usage
+        assert props["$ai_input_tokens"] == 20
+        assert props["$ai_output_tokens"] == 15

posthog/version.py

@@ -1,4 +1,4 @@
-VERSION = "3.18.0"
+VERSION = "3.19.0"
 
 if __name__ == "__main__":
     print(VERSION, end="")  # noqa: T201