Retrieve-wide then rerank-narrow

Exercise: Retrieve-wide then Rerank-narrow

In this exercise, you will implement a function that demonstrates the retrieve-wide then rerank-narrow pattern used in the LlamaIndex + Qdrant grounding pipeline. You will:

1. Retrieve a wide candidate set of nodes (e.g., top_k=20) from a retriever.
2. Rerank those nodes using a FastEmbedRerank cross-encoder node-postprocessor, keeping only the top_n=6 most relevant nodes.
3. Explain in a comment why a cross-encoder (which reads each (query, chunk) pair together) ranks better than the bi-encoder used for the initial retrieval.

Provided

• A retriever object (already configured with similarity_top_k=20).
• The build_reranker helper from the repository (from .rerank import build_reranker).
• The QueryBundle class from llama_index.core.schema.

Your Task

Fill in the # TODO: gaps in the starter code below. The function should:

• Use the retriever to get a list of NodeWithScore objects.
• Build a reranker (with top_n=6) using build_reranker.
• Apply the reranker via postprocess_nodes, passing the nodes and a QueryBundle created from the query string.
• Return the reranked nodes.
• Add an inline comment after the rerank call explaining why a cross-encoder is more precise for ranking than a bi-encoder.