Addressing Abrupt Topic Switches in Conversational AI

At a glance

The community member is using condenseQuestionChatEngine for conversations, but when the user abruptly switches topics, it sometimes condenses the question into an incoherent or nonsensical standalone question. The community members discuss using CondensePlusContext chat engine as an alternative, which condenses and retrieves context for every user message. However, the community member has a custom query engine and is unsure how to integrate it with CondensePlusContext, as it requires a response synthesizer instead of a query engine. The community members suggest customizing the condense prompt and using CondensePlusContextChatEngine.from_defaults(custom_retriever, node_postprocessors=node_postprocessors), but there is no explicitly marked answer on how to add the response synthesizer to the provided code.

rranjanj4

I am using condenseQuestionChatEngine for conversations. However, when the user abruptly switches topics, it sometimes condenses the question into an incoherent or nonsensical standalone question. How can I address this issue effectively?

9 comments

LLogan M

customizing the condense prompt?

LLogan M

tbh this is why i prefer the CondensePlusContext chat engine (condenses and calls retrieve for every user message, inserts the retrieved text as context into the system prompt)

rranjanj4

What do you mean by 'every user message'? I only want it to retrieve the context for latest user message. I already have a query engine that I’m passing to condenseQuestionChatEngine, but the CondensePlusContext function doesn’t accept a query engine as a parameter. Instead, it requires a response synthesizer. How can I handle this situation?

Plain Text

 def initialize_query_engine(vector_index, docstore, top_k_context):
    try:
        custom_retriever = QueryFusionRetriever(
            [
                vector_index.as_retriever(similarity_top_k=top_k_context),
                BM25Retriever.from_defaults(docstore=docstore, similarity_top_k=int(top_k_context))
            ],
            similarity_top_k=12,
            mode='reciprocal_rerank',
            use_async=True,
            verbose=True,
            query_gen_prompt=query_gen_str,
        )
        
        response_synthesizer = get_response_synthesizer(
            response_mode="compact", 
            text_qa_template=custom_qa_prompt, 
            verbose=True,
            use_async=True,
            streaming=False
            
        )
        
        custom_query_engine = RetrieverQueryEngine.from_args(
            retriever=custom_retriever,
            node_postprocessors=[
                MetadataReplacementPostProcessor(target_metadata_key="table_df"),
                reranker,
                PrevNextNodePostprocessor(docstore=docstore, num_nodes=1)
            ],
            response_synthesizer=response_synthesizer,
            verbose=True,
            use_async = True
        )
        return custom_query_engine

LLogan M

Yea latest user message is what I meant, same thing 🙂

LLogan M

CondensePlusContext uses a retriver

LLogan M

CondensePlusContextChatEngine.from_defaults(custom_retriever, node_postprocessors=node_postprocessors)

LLogan M

In your case, itd be the above

rranjanj4

how do i add the response synthesizer in the above code ?

LLogan M

There is no response synthesizer (or there is, but it's not meant to be set by the user)

Just set the condense and context prompts as needed

Add a reply

Find answers from the community

Addressing Abrupt Topic Switches in Conversational AI