要解决您在使用 streaming=True 功能时遇到的问题以及记录 DataFrame 中缺少的列，请参考以下步骤和注意事项：

1. 确保流式传输的正确配置:
   确保 RetrieverQueryEngine 已正确配置以处理流式传输响应。在构建 Response Synthesizer 时需要设置 streaming=True ,然后将其传递给 RetrieverQueryEngine 。

from llama_index.legacy.response_synthesizers import get_response_synthesizer
from llama_index.core.query_engine import RetrieverQueryEngine

synth = get_response_synthesizer(streaming=True, ...)
query_engine = RetrieverQueryEngine(response_synthesizer=synth, ...)

1. 处理记录 DataFrame 中缺少的列:
   TruLlama 类使用 get_eval_results_df 函数在记录 DataFrame 中聚合和显示反馈。如果像 "groundedness" 这样缺少列，请确保反馈函数已正确定义并应用。

from trulens_eval import Feedback, TruLlama

# Define feedback functions
f_groundedness = Feedback(provider.groundedness_measure_with_cot_reasons).on(context.collect()).on_output()
f_answer_relevance = Feedback(provider.relevance).on_input_output()
f_context_relevance = Feedback(provider.context_relevance_with_cot_reasons).on_input().on(context).aggregate(np.mean)

# Initialize TruLlama with feedbacks
tru_query_engine_recorder = TruLlama(query_engine, app_id='LlamaIndex_App1', feedbacks=[f_groundedness, f_answer_relevance, f_context_relevance])

1. 关于缺少列的已知问题:
   在 get_retrieval_results_df 函数中存在关于缺少列的已知问题。如果结果 DataFrame 中不存在指定的度量键，则函数会引发 ValueError 。请确保 DataFrame 中包含所有必要的度量键。

def get_retrieval_results_df(names, results_arr, metric_keys=None):
    metric_keys = metric_keys or DEFAULT_METRIC_KEYS
    avg_metrics_dict = defaultdict(list)
    for name, eval_results in zip(names, results_arr):
        metric_dicts = [eval_result.metric_vals_dict for eval_result in eval_results]
        results_df = pd.DataFrame(metric_dicts)
        for metric_key in metric_keys:
            if metric_key not in results_df.columns:
                raise ValueError(f"Metric key {metric_key} not in results_df")
            avg_metrics_dict[metric_key].append(results_df[metric_key].mean())
    return pd.DataFrame({"retrievers": names, **avg_metrics_dict})

通过确保流式传输的正确配置并解决关于缺少列的已知问题，您应该能够解决输出列为空以及记录 DataFrame 1(https://github.com/run-llama/llama_index/blob/main/docs/docs/module_guides/deploying/query_engine/streaming.md)2(https://github.com/run-llama/llama_index/blob/main/llama-index-core/llama_index/core/evaluation/notebook_utils.py) 中缺少 groundedness 列的问题。

• 要继续与 Dosu 对话，请提及 @dosu。*

我检查了所有方面，仍然面临相同的问题。如果流媒体为True, Jmeter 板输出始终为空。

To resolve the issue with the output being None when streaming=True in the query_engine , you need to ensure that the StreamingResponse object is properly handled and processed. The StreamingResponse object requires you to iterate over the tokens as they arrive, which might not be directly compatible with how TruLens expects the data.
Here is an example of how you can handle the StreamingResponse to ensure that the output is properly captured and processed:

# Assuming you have already set up the query engine with streaming enabled
query_engine = index.as_query_engine(streaming=True, similarity_top_k=1)

# Perform the query
streaming_response = query_engine.query("What did the author do growing up?")

# Initialize an empty string to collect the response
full_response = ""

# Iterate over the streaming response to collect the full output
for text in streaming_response.response_gen:
    full_response += text

# Now you can use the full_response as needed
print(full_response)

# Ensure that the full_response is passed to TruLens for evaluation
# Example: records_df = trulens.evaluate(full_response)

Additionally, ensure that the CondenseQuestionChatEngine is properly configured to handle streaming responses. The stream_chat method in the CondenseQuestionChatEngine class is designed to handle streaming responses and record them correctly:

@trace_method("chat")
def stream_chat(
    self, message: str, chat_history: Optional[List[ChatMessage]] = None
) -> StreamingAgentChatResponse:
    chat_history = chat_history or self._memory.get(input=message)

    # Generate standalone question from conversation context and last message
    condensed_question = self._condense_question(chat_history, message)

    log_str = f"Querying with: {condensed_question}"
    logger.info(log_str)
    if self._verbose:
        print(log_str)

    # Query with standalone question
    query_response = self._query_engine.query(condensed_question)

    tool_output = self._get_tool_output_from_response(
        condensed_question, query_response
    )

    # Record response
    if (
        isinstance(query_response, StreamingResponse)
        and query_response.response_gen is not None
    ):
        # override the generator to include writing to chat history
        self._memory.put(ChatMessage(role=MessageRole.USER, content=message))
        response = StreamingAgentChatResponse(
            chat_stream=response_gen_from_query_engine(query_response.response_gen),
            sources=[tool_output],
        )
        thread = Thread(
            target=response.write_response_to_history,
            args=(self._memory, True),
        )
        thread.start()
    else:
        raise ValueError("Streaming is not enabled. Please use chat() instead.")
    return response

By collecting the full response from the StreamingResponse object and ensuring that the CondenseQuestionChatEngine is properly configured to handle streaming responses, you can ensure that the output is not empty and can be properly evaluated by TruLens. This should resolve the issue with the output column being empty and the groundedness column missing 1(https://github.com/run-llama/llama_index/blob/main/docs/docs/module_guides/deploying/query_engine/streaming.md)2(https://github.com/run-llama/llama_index/blob/main/llama-index-core/llama_index/core/chat_engine/condense_question.py)3(https://github.com/run-llama/llama_index/blob/main/llama-index-legacy/llama_index/legacy/chat_engine/condense_question.py).

• To continue talking to Dosu , mention @dosu.*

仍然无法工作，出现相同的问题。