Relevance

Relevance is how well a retrieved document passage or search result matches a user’s query. In DocLD, vector search returns chunks ranked by similarity score (e.g. cosine similarity); optional reranking can further improve relevance before the LLM generates an answer.

How Relevance Is Used

• Retrieval — RAG fetches the top-k most similar chunks; higher relevance means better context for the model.
• Ranking — Results are ordered by similarity (and optionally by a reranker) so the most relevant content appears first.
• Filtering — You can set a confidence threshold or minimum score to exclude low-relevance chunks.

Improving relevance often involves chunking strategy, embedding model choice, and hybrid search or reranking.

Related Concepts

Similarity score and cosine similarity quantify relevance in embedding space. Reranking refines the order of retrieved chunks. Semantic search retrieves by meaning, which directly affects relevance.