fix(langchain): handle prompt_tokens_details as dict in _parse_usage_model

langfuse/langchain/CallbackHandler.py

@@ -1256,8 +1256,25 @@ def _parse_usage_model(usage: Union[pydantic.BaseModel, dict]) -> Any:
                 if "output" in usage_model:
                     usage_model["output"] = max(0, usage_model["output"] - value)
 
-        # Vertex AI
+        # OpenAI / LiteLLM — prompt_tokens_details as dict
+        # e.g. {"cached_tokens": 12000}
         if "prompt_tokens_details" in usage_model and isinstance(
+            usage_model["prompt_tokens_details"], dict
+        ):
+            prompt_tokens_details = usage_model.pop("prompt_tokens_details")
+
+            for key, value in prompt_tokens_details.items():
+                if not isinstance(value, int):
+                    continue
+
+                usage_model[f"input_{key}"] = value
+
+                if "input" in usage_model:
+                    usage_model["input"] = max(0, usage_model["input"] - value)
+
+        # Vertex AI — prompt_tokens_details as list
+        # e.g. [{"modality": "text", "token_count": N}]
+        elif "prompt_tokens_details" in usage_model and isinstance(
             usage_model["prompt_tokens_details"], list
         ):
             prompt_tokens_details = usage_model.pop("prompt_tokens_details")

tests/test_parse_usage_model.py

@@ -16,6 +16,68 @@ def test_standard_tier_input_token_details():
     assert result["total"] == 14
 
 
+def test_prompt_tokens_details_dict_cached_tokens():
+    """OpenAI/LiteLLM: prompt_tokens_details as dict with cached_tokens."""
+    usage = {
+        "prompt_tokens": 15000,
+        "completion_tokens": 500,
+        "total_tokens": 15500,
+        "prompt_tokens_details": {"cached_tokens": 12000},
+    }
+    result = _parse_usage_model(usage)
+    assert result["input"] == 3000  # 15000 - 12000
+    assert result["output"] == 500
+    assert result["total"] == 15500
+    assert result["input_cached_tokens"] == 12000
+
+
+def test_prompt_tokens_details_dict_with_cache_creation():
+    """OpenAI/LiteLLM: prompt_tokens_details dict + top-level cache_creation."""
+    usage = {
+        "prompt_tokens": 15000,
+        "completion_tokens": 500,
+        "total_tokens": 15500,
+        "prompt_tokens_details": {"cached_tokens": 12000},
+        "cache_creation_input_tokens": 3000,
+    }
+    result = _parse_usage_model(usage)
+    assert result["input"] == 3000  # 15000 - 12000 (cached_tokens only subtracted here)
+    assert result["input_cached_tokens"] == 12000
+    assert result["cache_creation_input_tokens"] == 3000
+
+
+def test_prompt_tokens_details_list_vertex_ai():
+    """Vertex AI: prompt_tokens_details as list — existing behavior preserved."""
+    usage = {
+        "prompt_token_count": 1000,
+        "candidates_token_count": 200,
+        "total_token_count": 1200,
+        "prompt_tokens_details": [
+            {"modality": "text", "token_count": 800},
+            {"modality": "image", "token_count": 200},
+        ],
+    }
+    result = _parse_usage_model(usage)
+    assert result["input"] == 0  # 1000 - 800 - 200
+    assert result["output"] == 200
+    assert result["total"] == 1200
+    assert result["input_modality_text"] == 800
+    assert result["input_modality_image"] == 200
+
+
+def test_prompt_tokens_details_dict_empty():
+    """Empty dict prompt_tokens_details — no crash, input unchanged."""
+    usage = {
+        "prompt_tokens": 5000,
+        "completion_tokens": 100,
+        "total_tokens": 5100,
+        "prompt_tokens_details": {},
+    }
+    result = _parse_usage_model(usage)
+    assert result["input"] == 5000
+    assert result["output"] == 100
+
+
 def test_priority_tier_not_subtracted():
     """Priority tier: 'priority' and 'priority_*' keys must NOT be subtracted."""
     usage = {