What are the RAG methods to minimize AI confabulation?

Found on google.com after search for "What are the different RAG methods to minimize confabulation?"

Before model context protocol MCP and other agentic context enhancement protocols retrieval augmented generation or RAG was the most important tool to add and extend context on order to minimize confabulations within LLM AI systems.

Following kinds of RAG were or are used so far:

• Vanilla RAG
  In the simplest form of RAG is a search for documents similar to the user prompt and the injection or the relevant documents into the resulting enhanced prompt in a straightforward pipeline.

• Standard RAG
  Here the vanilla RAG is enhanced with query rewriting and document search result reranking for improved retrieval quality.

• CAG (Cache-Augmented Generation)
  Instead of retrieving data every time during for each prompt, CAG loads relevant information into the model's context beforehand. This speeds things up and works well when the data is stable and fits in the model's context window.

• Hybrid RAG
  Combines multiple retrieval methods (dense, sparse, knowledge graphs) to collect comprehensive information.

• HyDe (Hypothetical Dokument Embeddings) takes a three step approach:

  1. First the system lets an LLM generate a "fake" or "hypothetical" answer to the user's query. This answer may not be factually accurate, but it will most probably contain the relevant keywords and context that match the users intention and the searched target information. The key assumption here is that the hypothetical document captures the intent and context of the query better than the raw query itself, leading to more complete or accurate document retrieval.
  2. Then the generated document is converted into a vector (embedding) which is then used to find real documents in a vector database that are semantically similar to the hypothetical answer.
  3. After all the returned selected relevant documents are injected into the resulting enhanced prompt

• Agentic RAG
  AI agents decide when, what, and how to retrieve information, with dynamic planning and multi-step reasoning.