The first configuration of a strategy comparison. Every other configuration is compared with it: metric differences, and the questions it fixed or broke.
A classic keyword-ranking method: chunks containing the question’s rarer words, more often, rank higher. Used by hybrid search.
A passage of a document, typically a few paragraphs. Documents are split into chunks, and retrieval returns chunks. See Chunking.
The method used to split documents into chunks (Recursive, Character, Token, Markdown, Code or JSON), with its settings such as size and overlap.
An embedding model with its retrieval pipeline, for example
text-embedding-3-small - hybrid - rerank cohere/rerank-v3.5. Evaluations and comparisons run configurations.The measure of how close two embeddings are. Higher means more similar. Scores are only comparable within one model.
A list of numbers (a vector) that represents the meaning of a text. Texts with similar meanings have close embeddings.
The chunks the first search (dense or hybrid) returns for a reranker to re-score: 30 by default, from 1 to 100, and at least top k.
A set of test questions with their expected chunks. See Golden datasets.
Vector search and BM25 keyword search run together, their rankings merged with reciprocal rank fusion. Helps on exact terms such as product names and error codes. See Hybrid search and reranking.
Large language model, such as GPT or Claude. In RAG, it writes the answer from the retrieved chunks.
A technique where a system retrieves relevant passages from your documents and gives them to an LLM to answer a question. See How RAG retrieval works.
How hybrid search merges the dense and keyword rankings: each chunk scores by its rank in each list, weighted by the Keyword / Dense setting.
A model (a cross-encoder) that reads the question with each first-stage candidate and re-scores them, keeping the top k. Useful when the right chunk is retrieved but not ranked first. Truvec offers Cohere Rerank 3.5 and Voyage rerank-2.5 and rerank-2.5-lite. See Hybrid search and reranking.
What runs around the embedding model in a configuration: dense or hybrid search, and an optional reranker. Chunking isn’t part of it, because it changes the chunks themselves.
The similarity of the best expected chunk minus the similarity of the best distractor. A larger gap means the model separates right from wrong chunks more clearly.