Retrieval typically works in two phases. A fast retriever pulls a broad pool of candidate passages using embeddings, then a heavier cross-encoder reranker scores each candidate against the query with far more nuance, promoting the truly relevant passages to the top so the language model works from the best possible context.
Because reranking decides the final shortlist a model reasons over, it strongly influences which sources get cited. Acromatico's approach is to make each candidate passage unambiguously answer the target query, so that when a reranker weighs relevance closely, the client's content earns the top positions that actually feed the generated response.
Related terms
Questions people ask
First-pass retrieval favors speed, casting a wide net that includes loosely relevant passages. Reranking applies a slower, more accurate model to reorder that pool so the most relevant content rises to the top, giving the language model a cleaner, higher-quality context to answer from.
The reranker sets the final shortlist the model actually reads, so a passage buried by low relevance rarely gets cited. Content that directly and clearly answers the query scores higher in reranking, earning the top positions that feed the generated answer and its sources.
See if AI actually names your brand
Acromatico runs live AI-visibility audits across ChatGPT, Perplexity, Gemini, and Claude.
Get a free AI Visibility Audit →