Glossary

RCP is designed backwards from the state of the art in retrieval. These are the terms that recur across the docs.

Term	Meaning
Bi-encoder	Encodes query and document independently into one vector each; fast, approximate (dense retrieval).
Cross-encoder	Jointly encodes a (query, document) pair for one relevance score; accurate, used for reranking.
Late interaction	Multi-vector scoring (MaxSim) that approximates cross-encoder accuracy at bi-encoder speed (ColBERT, ColPali).
Learned-sparse	A sparse term→weight vector produced by a model (SPLADE), searchable with an inverted index.
Dense / sparse / hybrid	Retrieval over dense vectors / lexical or learned-sparse terms / a fusion of both.
RRF	Reciprocal Rank Fusion — rank-only list merging robust to incomparable scores. See Federation.
MMR	Maximal Marginal Relevance — re-ranks for the relevance–diversity trade-off.
HyDE	Hypothetical Document Embeddings — embed an LLM-generated pseudo-answer as the query. See Query Transform.
GraphRAG	Retrieval over a knowledge graph / community summaries (local, global, drift). See Graph.
Agentic RAG	A client-driven reasoning loop that interleaves transform/retrieve/graph steps, optionally threaded by `sessionId`. See Agentic RAG.
Confidence	A normalised `[0,1]` per-hit relevance/self-eval signal, comparable across calls, that corrective/adaptive systems threshold on.
Retrieval unit / level	The granularity (chunk…subgraph…page) and abstraction level (RAPTOR depth, Leiden level) of a hit.
Memory (RCP)	A compact global/session structure over a corpus that yields clues/entry-points to steer retrieval (MemoRAG, HippoRAG). See Memory.
Feedback	A client→server signal (used/cited/reward/poisonSuspected) RL-trained or corrective retrievers consume. See Feedback.
Contextual retrieval	Prepending chunk-situating context before embedding/indexing (Anthropic).
Recall stage / rerank stage	Cheap high-recall candidate generation, then expensive precise reranking. See the pipeline.
MaxSim	The late-interaction scoring function: sum over query vectors of the max similarity to any document vector.
Aggregator	A server that federates downstream engines and advertises `catalog`.
Modality	Coarse content kind: text, image, audio, code, multimodal. See Content blocks.
Capability	A feature a server advertises at `initialize`; presence ⇒ supported. See Capabilities.
Selector / Federation	Client-side routing: pick one live backend, or fan out to all and fuse.

Reference

Ecosystem