AzureOpenAIGenerator not passing meta (used tokens, model...) to Langfuse

[LackesLab]

Describe the bug
When intializing a pipeline including the LangFuse tracer and an AzureOpenAIGenerator, costs and tokens cannot be tracked properly like with the "original" OpenAIGenerator

Error message

Expected behavior
Below, you see 2 Screenshots. One is frome the pipeline using anOpenAIGenerator, the second one using the AzureOpenAIGenerator. Tokens and costs are not displayed as intended.

Additional context

Add any other context about the problem here, like document types / preprocessing steps / settings of reader etc.

To Reproduce

Initialization

I initialized all components. But for the current issues, only the generator is the problem, so i just packed the initialization into the sample code.

_generator = AzureOpenAIGenerator(
        azure_endpoint="https://<endpoint>.openai.azure.com",
        azure_deployment="<deployment_name>")



extractive_qa_pipeline = Pipeline()
#Add the tracer
extractive_qa_pipeline.add_component("tracer", LangfuseConnector("Basic RAG Pipeline"))

extractive_qa_pipeline.add_component(name="text_embedder", instance=text_embedder)
extractive_qa_pipeline.add_component(name="retriever", instance=retriever )
extractive_qa_pipeline.add_component(name="prompt_builder", instance=PromptBuilder(template=self.prompt_template))
extractive_qa_pipeline.add_component(name="generator", instance=_generator)

extractive_qa_pipeline.connect( sender="text_embedder.embedding", receiver="retriever.query_embedding")
extractive_qa_pipeline.connect( sender="retriever.documents", receiver="prompt_builder.documents")
extractive_qa_pipeline.connect(sender="prompt_builder", receiver="generator")

FAQ Check

• Have you had a look at our new FAQ page?

System:

• OS: MacOs
• GPU/CPU: M2 Pro
• Haystack version (commit or version number): 2.1
• DocumentStore: Qdrant
• Reader: -
• Retriever: QdrantEmbeddingRetriever

[anakin87]

The problem probably lies in this part of the code. It seems that we are only considering OpenAI generators.

We need to investigate and see if we can generalize the behavior of the tracer.

[masci]

@anakin87 is correct, we only support OpenAIGenerator. I would expect AzureOpenAIGenerator to be a drop-in replacement for OpenAIGenerator, in that case the fix would be easy.