What Is Context Relevance in RAG?

Context relevance is the degree to which a retrieved passage actually addresses the user's query, as scored by the retrieval and reranking stages of a RAG pipeline. It is the number that decides whether your content enters the model's context window at all: passages are embedded, compared against the query vector (typically via cosine similarity), then re-scored by a cross-encoder reranker — and only the top handful reach answer synthesis.

How relevance scoring works

1. Chunking. The engine splits your page into passages, often 100-300 tokens each; a passage inherits only the meaning it contains, not the page's.
2. Embedding retrieval. Query and passages are mapped into vector space by an embedding model; nearest neighbors by cosine similarity form the candidate pool, often blended with keyword (BM25) matches in hybrid search.
3. Reranking. A cross-encoder — models like Cohere Rerank or open-source BGE rerankers — reads query and passage jointly and outputs a relevance score, reordering the pool.
4. Cutoff. Only the top-k passages (frequently 5-10) enter the context window.

Writing to maximize relevance honestly

• One question per passage. A 40-80 word paragraph that fully answers one query scores higher than a compound paragraph partially answering three — the principle behind passage-level optimization.
• Front-load the entity and the answer. "SOC 2 Type II audits take 3-12 months" outscores a paragraph that arrives at the same fact in sentence four.
• Mirror question vocabulary naturally. Rerankers reward semantic alignment; a heading phrased as the user's question ("How long does a SOC 2 audit take?") anchors the match.
• Make passages self-contained. Pronouns pointing at previous paragraphs ("this approach") lose their referent after chunking and score as vague.

Example

Two security vendors answer the same question. One writes "timelines depend on many factors" across three paragraphs; the other writes a single passage with the range, the audit stages, and the variable that moves it most. The reranker scores the second dramatically higher — and it becomes the quoted source. Relevance is earned at the paragraph level, which is why GEO content audits grade pages passage by passage.