We don't publish
your competitive advantage.

IFAzureChatOpenAI with streaming enabled fails with AssertionError when the model endpoint does not support streaming or when Azure content filter is triggered.

THENDetect whether the model supports streaming before calling stream(). If not, fall back to a non-streaming generation method like invoke(). Alternatively, provide a parameter to force streaming off when instantiating AzureChatOpenAI, e.g., streaming=False. Also handle content filter errors (openai.BadRequestError with code content_filter) gracefully instead of raising an assertion.

IFKeyError for the configured output key (e.g., 'answer') appears in console when streaming from a RunnableWithMessageHistory chain, even if the stream works correctly.

THENEnsure robust access to streamed chunks by using .get() with a fallback; upgrade langchain-core to the latest version if possible, as this is a known callback tracer issue in versions ≤0.2.23. For a quick workaround, wrap the stream loop in a try-except for KeyError to suppress the log noise.

IFUsing AzureChatOpenAI with an endpoint that does not support streaming (or when Azure content filter triggers) causes stream() to raise an AssertionError because generation is None.

THENSet streaming=False when instantiating AzureChatOpenAI to disable streaming. If the parameter is not available in your version, upgrade langchain-openai. Alternatively, catch the AssertionError and fall back to non-streaming generation using .invoke() or .generate().

IFUsing LiteLLM with Groq (via Databricks adapter) and streaming with stream_options={'include_usage': True} causes IndexError: list index out of range.

THENRemove 'include_usage' from stream_options or set stream_options to an empty dictionary (stream_options: {}) in the model's litellm_params configuration. This prevents the malformed chunk from being parsed.