Reranking
Re-scoring and reordering retrieved passages with a stronger model after initial retrieval.
In plain language
Vector search excels at recall, but similarity is not relevance. A reranker (cross-encoder or LLM) reads the question and each candidate together, judges whether it actually answers the question, and sends the best passages to the generator. It improves RAG precision at the cost of latency and compute. A common pattern retrieves 20-50 candidates, reranks and keeps the top 3-8.
Example
For “how long do refunds take,” vector search may retrieve the refund policy; reranking puts the “3-5 business days” passage first.
Recommended tools
Related terms
Embedding
Converting text, images or other content into vectors whose similarity can be computed.
Hybrid Search
Combining keyword search and vector search, then fusing the two result sets.
Retrieval-Augmented Generation
Retrieve relevant passages from a knowledge base, then have the model answer from them.
Agent Memory
A mechanism for an agent to store and retrieve user preferences, past tasks and key facts across sessions.