RAG concepts and terminology

Retrieval-Augmented Generation (RAG) is a design pattern in which an application retrieves information relevant to a question and supplies that information to a language model. The goal is to make generation depend on selected source material rather than on the model's built-in knowledge alone.

Ingestion path #

A document loader identifies or reads source material. A chunker divides content into smaller pieces. An embedder represents those pieces as vectors. A vector store holds those representations for retrieval. These are conceptual runtime responsibilities; the current Ragify frontend represents them as configurable node types but does not execute ingestion.

Query path #

A retriever finds pieces likely to match a user's question. A hybrid retriever combines semantic similarity with keyword matching, often using rank fusion to combine result lists. A reranker evaluates the retrieved candidates again and adjusts their order. An LLM receives selected context and prepares an answer. Answer / Output represents the final response stage.

RAG is more than a model call #

Quality depends on source selection, chunking, embeddings, retrieval, reranking, model behavior, and evaluation. Ragify makes the shape of the pipeline visible so those decisions can be discussed and adjusted individually.

Design versus execution #

A pipeline graph is a design artifact. Its shape alone does not show that documents exist in storage, API credentials are authorized, or a model can produce an answer. Read Product availability before interpreting preview screens as results.