The community members are discussing ways to optimize a query process without calling a large language model (LLM). They suggest using index.query(..., response_mode="no_text") to only fetch the source nodes without calling the LLM, and index.query(..., similarity_cutoff=0.5) to filter the results based on similarity. There is some uncertainty around the behavior when all nodes are filtered out. The community members also discuss increasing the number of results returned with index.query(..., similarity_top_k=4), and confirm that using similarity_cutoff without response_mode="no_text" will not call the LLM, only the embeddings model.
