Meet 20 RAG Family Members

Retrieval Augmented Generation, better known as RAG, gives a language model an open book. Instead of answering from memory alone, the model first looks up relevant information, then writes its answer from what it found. This reduces hallucinations and keeps answers grounded in your own data. Simple idea, but the family has grown fast. Today there are many RAG strategies, each built to fix a specific weakness of the one before it.

In this article, we meet 20 members of the RAG family. We group them into four branches: Better Retrieval, Structured Knowledge, Self-Correcting Loops, and Agentic Systems. For each member, you get a one-line definition and two real-world use cases.

Branch 1: Better Retrieval

This branch focuses on one question: how do we find the right chunks?

1. Naive RAG
Embed your documents, retrieve the closest matches to the question, and pass them to the LLM.
Use cases: an HR chatbot answering leave-policy questions; a product FAQ assistant on a company website.

2. Hybrid RAG
Combine keyword search (BM25) with vector search to catch both exact terms and meaning.
Use cases: bank support search for codes like "IFRS 9 Stage 2"; IT helpdesk search for error codes and vague complaints.

3. Re-ranking RAG
Retrieve many candidates, then use a cross-encoder to reorder them by true relevance.
Use cases: legal research surfacing the strongest precedent first; enterprise search ranking the most useful policy page on top.

4. Metadata-filtered RAG
Filter by tags such as date, region, or document type before searching.
Use cases: finding only Ontario contracts signed after 2024; retrieving only the latest version of a compliance manual.

5. Multi-query RAG (RAG-Fusion)
Rewrite the question several ways, search with each, and merge the results.
Use cases: customer support where "card not working" also means "card declined"; academic search across different terminology.

6. HyDE (Hypothetical Document Embeddings)
Let the LLM draft a hypothetical answer, then search using that draft.
Use cases: medical Q&A where patient words differ from clinical terms; searching technical docs with plain-language questions.

7. Parent-document RAG
Match on small chunks but return the larger section they belong to.
Use cases: policy manuals where one sentence needs its full section; textbooks where a definition needs its surrounding explanation.

8. Contextual Retrieval
Add document-level context to every chunk before embedding it.
Use cases: quarterly financial reports where "revenue grew 3%" needs company and quarter attached; multi-client contract libraries.

9. Multi-vector RAG (ColBERT)
Store many vectors per document so fine details still match.
Use cases: e-commerce search for "waterproof hiking boots, wide fit"; patent search where small technical details matter.

Branch 2: Structured Knowledge

This branch changes how knowledge is organized before retrieval.

10. GraphRAG
Build a knowledge graph of entities and relationships, then retrieve over it.
Use cases: fraud investigation linking accounts and transactions; tracing citations between court cases.

11. Hierarchical RAG (RAPTOR)
Build a tree of summaries so questions can be answered at any level of detail.
Use cases: summarizing themes across 500 customer complaints; answering big-picture questions about a long book.

12. Multimodal RAG
Retrieve across text, images, tables, and PDFs together.
Use cases: insurance claims combining damage photos with policy documents; manufacturing support using diagrams and manuals.

13. SQL RAG
Convert the question into SQL and answer straight from the database.
Use cases: "top 5 branches by loan volume last quarter"; sales teams querying CRM numbers in plain English.

Branch 3: Self-Correcting Loops

This branch teaches the system to check and fix its own work.

14. Self-RAG
The model decides when to retrieve and critiques its own answer.
Use cases: a medical assistant that skips retrieval for greetings but fact-checks dosage questions; a research assistant flagging unsupported claims.

15. Corrective RAG (CRAG)
Grade the retrieved documents and fall back to web search when they are weak.
Use cases: an internal knowledge bot whose docs may be outdated; a news assistant needing current facts.

16. Adaptive RAG
Route each query to a simple or deep retrieval path based on complexity.
Use cases: support bots answering FAQs instantly while handling billing disputes in depth; cost control in high-traffic apps.

17. Multi-hop RAG
Chain several retrieval steps to connect facts across documents.
Use cases: "Which director of the acquiring company sits on another board?"; due diligence across linked filings.

18. Speculative RAG
A small model drafts answers from retrieved documents and a large model verifies them.
Use cases: high-volume customer chatbots that need speed; reducing inference costs in production.

Branch 4: Agentic Systems

This branch hands the retrieval strategy to an agent.

19. Agentic RAG
An agent chooses which tools, sources, and steps to use.
Use cases: an analyst assistant checking SQL, policy docs, and the web in one task; a coding agent searching repos and documentation.

20. Memory RAG
Retrieve from past conversations and user history to personalize answers.
Use cases: a tutoring bot remembering last week's weak topic; a support agent recalling a customer's earlier tickets.

Choosing Your Family Member

You don't need all 20. Most production systems start with Naive RAG, then add Hybrid search, Re-ranking, and Metadata filters. Those four solve most real problems. Before adding a new member, measure first. Tools like RAGAs or DeepEval show whether retrieval or generation is the weak link, so you fix the right problem instead of adding complexity for its own sake. Move to GraphRAG or Multi-hop when questions need connected facts, and to Self-correcting or Agentic designs when accuracy and flexibility matter more than speed.

The RAG family keeps growing, and new members appear every few months. But almost every newcomer fits into one of these four branches. Learn the branches, and you will always know where a new member belongs.