A reranker cannot rescue a chunk it never receives.
Calculate recall@10 = number of shared results / 10.
I observed that Retrieval‑Augmented Generation pipelines can miss the right chunk even when it exists in the vector store, because approximate nearest‑neighbor search may not return the true nearest vector.
Exact nearest‑neighbor search compares the query with every vector, which costs roughly O(Nd) distance work per query.
HNSW connects vectors in a layered graph. Search starts in a sparse upper layer, follows promising links, then explores a denser layer; recall may be below 100 %.
The search budget, often called efSearch (hnsw_ef in Qdrant), controls how widely the graph is explored. Raising it can improve recall, but costs latency.
It does not change how many results you ask for; it changes how hard the index looks for them.
If exact search finds the needed chunk but HNSW doesn't, investigate the index and search settings.
If exact search doesn't find it either, a wider graph search won't fix your embeddings, chunking, or missing source content.
When recall remains low after raising efSearch, I will inspect upstream steps such as embedding generation, chunk size, and source ingestion before blaming the reranker.