New nodes

At a glance

The post describes a bug where the response is empty when new_nodes is empty. Community members discuss this issue and suggest potential solutions:

- If no nodes are added to the index, there is nothing to query.

- The community member needs GPT to provide meaningful responses even when no reference nodes are found.

- They suggest using second-stage processing and filtering nodes based on a custom logic, but encounter an issue where the query engine returns None when all nodes are filtered out.

- Another community member points out a bug in the documentation, where the example code is missing a required parameter for the node_postprocessor.postprocess_nodes function.

- The community members discuss using LlamaIndex as a custom tool in a framework like Langchain to achieve a more chatbot-like experience, as LlamaIndex is currently focused on being a search tool.

There is no explicitly marked answer in the comments.

Useful resources

aaejoaj

BUG！
reference: (https://gpt-index.readthedocs.io/en/latest/how_to/query/second_stage.html#using-as-independent-module-lower-level-usage),
when new_nodes is empty,response is emtpy.

Plain Text

    list_index = GPTListIndex(
        new_nodes,
        service_context=service_context
    )
    query_engine = list_index.as_query_engine(
        similarity_top_k=1, response_mode="compact",
        text_qa_template=TEXT_QA_TEMPLATE,
        refine_template=REFINE_TEMPLATE
    )
    response = query_engine.query(query_str)
    print(response)

8 comments

LLogan M

If you didn't add nodes to the index, then there is nothing to query ...

aaejoaj

I know that, but I need to have GPT answer user questions whether or not a reference node is found. So what show I do?

LLogan M

I'm not sure I understand the full use case here 🤔

aaejoaj

Ok, let's think step by step:

I need filter nodes(custom logic, node.score > 0.5 or something else), so I use second-stage processing low level usage(https://gpt-index.readthedocs.io/en/latest/how_to/query/second_stage.html#using-as-independent-module-lower-level-usage)
Sometime all nodes are filtered, at this time, query_engine.query("hello!") return None, BUT, I need query_engine return meaningful results (just like "hello I'm chatbot") instead of None
what should I do?

aaejoaj

Also I found a bug in document (https://gpt-index.readthedocs.io/en/latest/how_to/query/second_stage.html#using-as-independent-module-lower-level-usage)
your example does not pass (query bundle) required by node_postprocessor.postprocess_nodes

Plain Text

new_nodes = node_postprocessor.postprocess_nodes(resp_nodes)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
...site-packages/llama_index/indices/postprocessor/node_recency.py", line 70, in postprocess_nodes
    raise ValueError("Missing query bundle in extra info.")
ValueError: Missing query bundle in extra info.

LLogan M

If you want a more chatbot experience, I would use llama index as a custom tool in something like langchain

Right now, llama index is focused on being a good search tool.

In the future, we have plans to add more native chat controls though 🙏🙏

aaejoaj

Maybe It's not experience problem but a bug. When the search does not find the corresponding node, the second-stage engine should at least be able to reply according to the user's question, instead of None

LLogan M

At the core of llama index, every llm response needs to be grounded by context (hence, it being a search tool)

Very welcome to PRs to improve this 💪

Add a reply

Find answers from the community

New nodes