fix(llm-observability): parallel traces

posthog/ai/langchain/callbacks.py

@@ -32,13 +32,22 @@
 
 
 class RunMetadata(TypedDict, total=False):
-    messages: Union[List[Dict[str, Any]], List[str]]
+    input: Any
+    """Input of the run: messages, prompt variables, etc."""
+    name: str
+    """Name of the run: chain name, model name, etc."""
     provider: str
+    """Provider of the run: OpenAI, Anthropic"""
     model: str
+    """Model used in the run"""
     model_params: Dict[str, Any]
+    """Model parameters of the run: temperature, max_tokens, etc."""
     base_url: str
+    """Base URL of the provider's API used in the run."""
     start_time: float
+    """Start time of the run."""
     end_time: float
+    """End time of the run."""
 
 
 RunStorage = Dict[UUID, RunMetadata]
@@ -119,8 +128,7 @@ def on_chain_start(
         self._log_debug_event("on_chain_start", run_id, parent_run_id, inputs=inputs)
         self._set_parent_of_run(run_id, parent_run_id)
         if parent_run_id is None and self._trace_name is None:
-            self._trace_name = self._get_langchain_run_name(serialized, **kwargs)
-            self._trace_input = inputs
+            self._set_span_metadata(run_id, self._get_langchain_run_name(serialized, **kwargs), inputs)
 
     def on_chat_model_start(
         self,
@@ -134,7 +142,7 @@ def on_chat_model_start(
         self._log_debug_event("on_chat_model_start", run_id, parent_run_id, messages=messages)
         self._set_parent_of_run(run_id, parent_run_id)
         input = [_convert_message_to_dict(message) for row in messages for message in row]
-        self._set_run_metadata(serialized, run_id, input, **kwargs)
+        self._set_llm_metadata(serialized, run_id, input, **kwargs)
 
     def on_llm_start(
         self,
@@ -147,7 +155,7 @@ def on_llm_start(
     ):
         self._log_debug_event("on_llm_start", run_id, parent_run_id, prompts=prompts)
         self._set_parent_of_run(run_id, parent_run_id)
-        self._set_run_metadata(serialized, run_id, prompts, **kwargs)
+        self._set_llm_metadata(serialized, run_id, prompts, **kwargs)
 
     def on_llm_new_token(
         self,
@@ -204,7 +212,7 @@ def on_chain_end(
         self._pop_parent_of_run(run_id)
 
         if parent_run_id is None:
-            self._capture_trace(run_id, outputs=outputs)
+            self._pop_trace_and_capture(run_id, outputs=outputs)
 
     def on_chain_error(
         self,
@@ -218,7 +226,7 @@ def on_chain_error(
         self._pop_parent_of_run(run_id)
 
         if parent_run_id is None:
-            self._capture_trace(run_id, outputs=None)
+            self._pop_trace_and_capture(run_id, outputs=None)
 
     def on_llm_end(
         self,
@@ -253,7 +261,7 @@ def on_llm_end(
             "$ai_provider": run.get("provider"),
             "$ai_model": run.get("model"),
             "$ai_model_parameters": run.get("model_params"),
-            "$ai_input": with_privacy_mode(self._client, self._privacy_mode, run.get("messages")),
+            "$ai_input": with_privacy_mode(self._client, self._privacy_mode, run.get("input")),
             "$ai_output_choices": with_privacy_mode(self._client, self._privacy_mode, output),
             "$ai_http_status": 200,
             "$ai_input_tokens": input_tokens,
@@ -292,7 +300,7 @@ def on_llm_error(
             "$ai_provider": run.get("provider"),
             "$ai_model": run.get("model"),
             "$ai_model_parameters": run.get("model_params"),
-            "$ai_input": with_privacy_mode(self._client, self._privacy_mode, run.get("messages")),
+            "$ai_input": with_privacy_mode(self._client, self._privacy_mode, run.get("input")),
             "$ai_http_status": _get_http_status(error),
             "$ai_latency": latency,
             "$ai_trace_id": trace_id,
@@ -377,7 +385,14 @@ def _find_root_run(self, run_id: UUID) -> UUID:
             id = self._parent_tree[id]
         return id
 
-    def _set_run_metadata(
+    def _set_span_metadata(self, run_id: UUID, name: str, input: Any):
+        self._runs[run_id] = {
+            "name": name,
+            "input": input,
+            "start_time": time.time(),
+        }
+
+    def _set_llm_metadata(
         self,
         serialized: Dict[str, Any],
         run_id: UUID,
@@ -387,7 +402,7 @@ def _set_run_metadata(
         **kwargs,
     ):
         run: RunMetadata = {
-            "messages": messages,
+            "input": messages,
             "start_time": time.time(),
         }
         if isinstance(invocation_params, dict):
@@ -450,12 +465,16 @@ def _get_langchain_run_name(self, serialized: Optional[Dict[str, Any]], **kwargs
         except (KeyError, TypeError):
             pass
 
-    def _capture_trace(self, run_id: UUID, *, outputs: Optional[Dict[str, Any]]):
+    def _pop_trace_and_capture(self, run_id: UUID, *, outputs: Optional[Dict[str, Any]]):
         trace_id = self._get_trace_id(run_id)
+        run = self._pop_run_metadata(run_id)
+        if not run:
+            return
         event_properties = {
-            "$ai_trace_name": self._trace_name,
+            "$ai_trace_name": run.get("name"),
             "$ai_trace_id": trace_id,
-            "$ai_input_state": with_privacy_mode(self._client, self._privacy_mode, self._trace_input),
+            "$ai_input_state": with_privacy_mode(self._client, self._privacy_mode, run.get("input")),
+            "$ai_latency": run.get("end_time", 0) - run.get("start_time", 0),
             **self._properties,
         }
         if outputs is not None:

posthog/test/ai/langchain/test_callbacks.py

@@ -1,3 +1,4 @@
+import asyncio
 import logging
 import math
 import os
@@ -67,7 +68,7 @@ def test_metadata_capture(mock_client):
     callbacks = CallbackHandler(mock_client)
     run_id = uuid.uuid4()
     with patch("time.time", return_value=1234567890):
-        callbacks._set_run_metadata(
+        callbacks._set_llm_metadata(
             {"kwargs": {"openai_api_base": "https://us.posthog.com"}},
             run_id,
             messages=[{"role": "user", "content": "Who won the world series in 2020?"}],
@@ -76,7 +77,7 @@ def test_metadata_capture(mock_client):
         )
     expected = {
         "model": "hog-mini",
-        "messages": [{"role": "user", "content": "Who won the world series in 2020?"}],
+        "input": [{"role": "user", "content": "Who won the world series in 2020?"}],
         "start_time": 1234567890,
         "model_params": {"temperature": 0.5},
         "provider": "posthog",
@@ -90,6 +91,19 @@ def test_metadata_capture(mock_client):
     callbacks._pop_run_metadata(uuid.uuid4())  # should not raise
 
 
+def test_run_metadata_capture(mock_client):
+    callbacks = CallbackHandler(mock_client)
+    run_id = uuid.uuid4()
+    with patch("time.time", return_value=1234567890):
+        callbacks._set_span_metadata(run_id, "test", 1)
+    expected = {
+        "name": "test",
+        "input": 1,
+        "start_time": 1234567890,
+    }
+    assert callbacks._runs[run_id] == expected
+
+
 @pytest.mark.parametrize("stream", [True, False])
 def test_basic_chat_chain(mock_client, stream):
     prompt = ChatPromptTemplate.from_messages(
@@ -514,7 +528,11 @@ def test_callbacks_logic(mock_client):
     assert callbacks._parent_tree == {}
 
     def assert_intermediary_run(m):
-        assert callbacks._runs == {}
+        assert len(callbacks._runs) != 0
+        run = next(iter(callbacks._runs.values()))
+        assert run["name"] == "RunnableSequence"
+        assert run["input"] == {}
+        assert run["start_time"] is not None
         assert len(callbacks._parent_tree.items()) == 1
         return [m]
 
@@ -981,3 +999,33 @@ def test_tool_calls(mock_client):
         }
     ]
     assert "additional_kwargs" not in generation_call["properties"]["$ai_output_choices"][0]
+
+
+async def test_async_traces(mock_client):
+    async def sleep(x):  # -> Any:
+        await asyncio.sleep(0.1)
+        return x
+
+    prompt = ChatPromptTemplate.from_messages([("user", "Foo")])
+    chain1 = RunnableLambda(sleep)
+    chain2 = prompt | FakeMessagesListChatModel(responses=[AIMessage(content="Bar")])
+
+    cb = CallbackHandler(mock_client)
+
+    start_time = time.time()
+    await asyncio.gather(
+        chain1.ainvoke({}, config={"callbacks": [cb]}),
+        chain2.ainvoke({}, config={"callbacks": [cb]}),
+    )
+    approximate_latency = math.floor(time.time() - start_time)
+    assert mock_client.capture.call_count == 3
+
+    first_call, second_call, third_call = mock_client.capture.call_args_list
+    assert first_call[1]["event"] == "$ai_generation"
+    assert second_call[1]["event"] == "$ai_trace"
+    assert second_call[1]["properties"]["$ai_trace_name"] == "RunnableSequence"
+    assert third_call[1]["event"] == "$ai_trace"
+    assert third_call[1]["properties"]["$ai_trace_name"] == "sleep"
+    assert (
+        min(approximate_latency - 1, 0) <= math.floor(third_call[1]["properties"]["$ai_latency"]) <= approximate_latency
+    )

posthog/version.py

@@ -1,4 +1,4 @@
-VERSION = "3.9.2"
+VERSION = "3.9.3"
 
 if __name__ == "__main__":
     print(VERSION, end="")  # noqa: T201