Skip to content

LLM-Driven Extraction of Unstructured Data — Built for API Deployments & ETL Pipeline Workflows

UN-3249 [FIX] Use actual model name for Azure OpenAI cost tracking (#1805)

Praveen Kumare0c2af0168c9
diff --git a/unstract/sdk1/src/unstract/sdk1/adapters/base1.py b/unstract/sdk1/src/unstract/sdk1/adapters/base1.py
index 7a578d486..f3aa073a1 100644
--- a/unstract/sdk1/src/unstract/sdk1/adapters/base1.py
+++ b/unstract/sdk1/src/unstract/sdk1/adapters/base1.py
@@ -248,6 +248,9 @@ class AzureOpenAILLMParameters(BaseChatCompletionParameters):
@staticmethod
def validate(adapter_metadata: dict[str, "Any"]) -> dict[str, "Any"]:
+ # Capture user-provided model name before deployment_name overwrites it
+ original_model = adapter_metadata.get("model", "")
+
adapter_metadata["model"] = AzureOpenAILLMParameters.validate_model(
adapter_metadata
)
@@ -296,6 +299,14 @@ class AzureOpenAILLMParameters(BaseChatCompletionParameters):
"reasoning_effort", "medium"
)
+ # Preserve actual model name for cost tracking (deployment_name is used
+ # for LiteLLM routing but doesn't match pricing table entries)
+ if original_model:
+ cost_model = original_model
+ if not cost_model.startswith("azure/"):
+ cost_model = f"azure/{cost_model}"
+ validated["cost_model"] = cost_model
+
return validated
@staticmethod
@@ -720,6 +731,9 @@ class AzureOpenAIEmbeddingParameters(BaseEmbeddingParameters):
@staticmethod
def validate(adapter_metadata: dict[str, "Any"]) -> dict[str, "Any"]:
+ # Capture user-provided model name before deployment_name overwrites it
+ original_model = adapter_metadata.get("model", "")
+
adapter_metadata["model"] = AzureOpenAIEmbeddingParameters.validate_model(
adapter_metadata
)
@@ -733,7 +747,17 @@ class AzureOpenAIEmbeddingParameters(BaseEmbeddingParameters):
if "num_retries" in adapter_metadata and not adapter_metadata.get("max_retries"):
adapter_metadata["max_retries"] = adapter_metadata["num_retries"]
- return AzureOpenAIEmbeddingParameters(**adapter_metadata).model_dump()
+ result = AzureOpenAIEmbeddingParameters(**adapter_metadata).model_dump()
+
+ # Preserve actual model name for cost tracking (deployment_name is used
+ # for LiteLLM routing but doesn't match pricing table entries)
+ if original_model:
+ cost_model = original_model
+ if not cost_model.startswith("azure/"):
+ cost_model = f"azure/{cost_model}"
+ result["cost_model"] = cost_model
+
+ return result
@staticmethod
def validate_model(adapter_metadata: dict[str, "Any"]) -> str:
diff --git a/unstract/sdk1/src/unstract/sdk1/embedding.py b/unstract/sdk1/src/unstract/sdk1/embedding.py
index 37e66872c..758aa0e1c 100644
--- a/unstract/sdk1/src/unstract/sdk1/embedding.py
+++ b/unstract/sdk1/src/unstract/sdk1/embedding.py
@@ -87,6 +87,7 @@ class Embedding:
try:
self.platform_kwargs: dict[str, object] = kwargs
self.kwargs: dict[str, object] = self.adapter.validate(self._adapter_metadata)
+ self._cost_model: str | None = self.kwargs.pop("cost_model", None)
except ValidationError as e:
raise SdkError("Invalid embedding adapter metadata: " + str(e)) from e
@@ -192,7 +193,12 @@ class EmbeddingCompat(BaseEmbedding):
self._tool = tool
# For compatibility with SDK Callback Manager.
- self.model_name = self._embedding_instance.kwargs.get("model", "")
+ # Prefer cost_model (actual model name) for pricing lookup accuracy,
+ # falling back to the model/deployment name used for routing.
+ self.model_name = (
+ self._embedding_instance._cost_model
+ or self._embedding_instance.kwargs.get("model", "")
+ )
self.callback_manager = None
if not PlatformHelper.is_public_adapter(
diff --git a/unstract/sdk1/src/unstract/sdk1/llm.py b/unstract/sdk1/src/unstract/sdk1/llm.py
index 486294565..b907cd396 100644
--- a/unstract/sdk1/src/unstract/sdk1/llm.py
+++ b/unstract/sdk1/src/unstract/sdk1/llm.py
@@ -109,6 +109,7 @@ class LLM:
self.platform_kwargs["adapter_instance_id"] = self._adapter_instance_id
self.kwargs = self.adapter.validate(self._adapter_metadata)
+ self._cost_model = self.kwargs.pop("cost_model", None)
# REF: https://docs.litellm.ai/docs/completion/input#translated-openai-params
# supported = get_supported_openai_params(model=self.kwargs["model"],
@@ -205,6 +206,7 @@ class LLM:
)
completion_kwargs = self.adapter.validate({**self.kwargs, **kwargs})
+ completion_kwargs.pop("cost_model", None)
# if hasattr(self, "model") and self.model not in O1_MODELS:
# completion_kwargs["temperature"] = 0.003
@@ -219,7 +221,10 @@ class LLM:
response_text = response["choices"][0]["message"]["content"]
self._record_usage(
- self.kwargs["model"], messages, response.get("usage"), "complete"
+ self._cost_model or self.kwargs["model"],
+ messages,
+ response.get("usage"),
+ "complete",
)
# NOTE:
@@ -289,6 +294,7 @@ class LLM:
)
completion_kwargs = self.adapter.validate({**self.kwargs, **kwargs})
+ completion_kwargs.pop("cost_model", None)
for chunk in litellm.completion(
messages=messages,
@@ -300,7 +306,7 @@ class LLM:
):
if chunk.get("usage"):
self._record_usage(
- self.kwargs["model"],
+ self._cost_model or self.kwargs["model"],
messages,
chunk.get("usage"),
"stream_complete",
@@ -356,6 +362,7 @@ class LLM:
)
completion_kwargs = self.adapter.validate({**self.kwargs, **kwargs})
+ completion_kwargs.pop("cost_model", None)
response = await litellm.acompletion(
messages=messages,
@@ -364,7 +371,10 @@ class LLM:
response_text = response["choices"][0]["message"]["content"]
self._record_usage(
- self.kwargs["model"], messages, response.get("usage"), "acomplete"
+ self._cost_model or self.kwargs["model"],
+ messages,
+ response.get("usage"),
+ "acomplete",
)
response_object = LLMResponseCompat(response_text)