Reranking
Summary
I kept seeing reranking in RAG discussions and finally clicked that it's a specific technique, not a vague concept. Your fast initial retrieval grabs a batch of candidate documents, then a slower, sharper model re-scores just those and reorders them before the LLM ever sees them. That second pass is where you can factor in recency or source authority that would melt your servers if you ran it across the whole corpus.
On Reranking
I've been seeing the term reranking everywhere in RAG retrieval discussions, and it finally clicked that this isn't just a general concept. It's a specialized technical term with a specific job. After your initial retrieval grabs documents, they get re-scored and reshuffled using more sophisticated scoring than whatever yanked them in the first place.
Here's how reranking typically works:
- Stage 1: Grab the top-K results (your best initial matches) using something fast. Usually embedding similarity, which measures how close documents are in meaning space
- Stage 2: Apply a beefier scoring model to just these candidates; more accurate but too expensive to run on everything
- Reorder based on the refined scores
- Ship it to the LLM for final response generation
In RAG this technique lets you play different games at each stage. Initial retrieval might prioritize semantic similarity, but reranking can factor in recency, source authority, or even task-specific relevance that would melt your servers if you ran it against your entire document corpus.
In more depth, this one is pretty good: Rerankers - Pinecone