fix(langchain): cached token usage

langfuse/callback/langchain.py

@@ -1096,12 +1096,18 @@ def _parse_usage_model(usage: typing.Union[pydantic.BaseModel, dict]):
             for key, value in input_token_details.items():
                 usage_model[f"input_{key}"] = value
 
+                if "input" in usage_model:
+                    usage_model["input"] -= value
+
         if "output_token_details" in usage_model:
             output_token_details = usage_model.pop("output_token_details", {})
 
             for key, value in output_token_details.items():
                 usage_model[f"output_{key}"] = value
 
+                if "output" in usage_model:
+                    usage_model["output"] -= value
+
     return usage_model if usage_model else None
 
 

tests/test_langchain.py

@@ -2318,3 +2318,49 @@ def call_model(state: MessagesState):
             assert observation.level == "DEFAULT"
 
     assert hidden_count > 0
+
+
+def test_cached_token_usage():
+    prompt = ChatPromptTemplate.from_messages(
+        [
+            (
+                "system",
+                (
+                    "This is a test prompt to reproduce the issue. "
+                    "The prompt needs 1024 tokens to enable cache." * 100
+                ),
+            ),
+            ("user", "Reply to this message {test_param}."),
+        ]
+    )
+    chat = ChatOpenAI(model="gpt-4o-mini")
+    chain = prompt | chat
+    handler = CallbackHandler()
+    config = {"callbacks": [handler]} if handler else {}
+
+    chain.invoke({"test_param": "in a funny way"}, config)
+
+    # invoke again to force cached token usage
+    chain.invoke({"test_param": "in a funny way"}, config)
+
+    handler.flush()
+
+    trace = get_api().trace.get(handler.get_trace_id())
+
+    generation = next((o for o in trace.observations if o.type == "GENERATION"))
+
+    assert generation.usage_details["input_cache_read"] > 0
+    assert (
+        generation.usage_details["input"]
+        + generation.usage_details["input_cache_read"]
+        + generation.usage_details["output"]
+        == generation.usage_details["total"]
+    )
+
+    assert generation.cost_details["input_cache_read"] > 0
+    assert (
+        generation.cost_details["input"]
+        + generation.cost_details["input_cache_read"]
+        + generation.cost_details["output"]
+        == generation.cost_details["total"]
+    )