In brief
- Standard RAG searches once for relevant passages and passes them to a language model.
- Hybrid RAG usually combines semantic vector search with exact keyword search and additional ranking.
- Multi-hop RAG connects several retrieval steps; GraphRAG exposes relationships and corpus-wide patterns.
- Multimodal RAG can include tables, images, audio or video.
- Adaptive and agentic approaches decide dynamically whether, where and how often to retrieve.
- The simplest architecture that demonstrably meets the need is usually the best starting point.
Why there is no single official list
The original term retrieval-augmented generation describes a principle: a generative model receives information relevant to a query from an external knowledge source. Many architectures, retrieval methods and quality mechanisms have since been developed.
A frequently cited research taxonomy distinguishes naive, advanced and modular RAG. For an enterprise project, a second practical view is useful: how are sources represented, how does retrieval proceed and how is the workflow controlled?
- Form of the knowledge source
- Text passages and vectors, structured data, knowledge graphs and multiple media types require different ingestion and retrieval methods.
- Retrieval flow
- A query may be answered with one retrieval step or require several dependent queries and sources.
- Orchestration and quality control
- The workflow may be fixed or decide dynamically which retrieval, source, correction or additional check is required.
Standard RAG and hybrid RAG
In a simple RAG workflow, a question is translated into a search query. The system retrieves relevant passages, provides them to the language model as context and generates an answer. This is often a sensible starting point for a bounded document collection.
Vector search is good at recognising similar meaning. Classic keyword search often handles exact identifiers, product names and rare technical terms more reliably. Hybrid RAG combines both methods and usually reranks the resulting candidates.
| Approach | How it works | Typical strength | Typical limitation |
|---|---|---|---|
| Standard RAG | One query, one retrieval step, relevant passages and an answer based on them. | Bounded knowledge collections and direct questions | Complex questions spanning several sources may receive an incomplete answer. |
| Vector search | Finds content with similar meaning even when different words are used. | Naturally phrased questions and semantically related passages | Exact codes, names or number sequences may receive too little weight. |
| Keyword search | Weights matching words and character sequences. | Product numbers, error codes, proper names and specialist terms | May miss a relevant passage that expresses the same idea differently. |
| Hybrid RAG | Combines semantic and exact retrieval, filters and often a reranker. | Heterogeneous enterprise documents and mixed query types | More components must be tuned and tested with real questions. |
Multi-hop RAG and GraphRAG solve different problems
Multi-hop RAG describes the workflow: a complex question is split into several dependent retrieval steps. An intermediate result determines what to retrieve next. This helps when an answer only emerges by connecting multiple pieces of evidence.
GraphRAG primarily describes how knowledge is represented and made retrievable. Entities such as people, systems, products or contracts are connected through relationships. This can reveal dependencies, networks and corpus-wide themes that are difficult to see in isolated passages.
The two can work together: a multi-hop workflow can follow several paths through a knowledge graph. Conversely, not every GraphRAG query requires multiple retrieval steps.
- Multi-hop RAG
- Useful for a question such as: Which affected plants belong to suppliers whose certification expires soon and that have open critical deviations?
- GraphRAG – local view
- Retrieves the neighbourhood of an entity or relationship, such as systems that depend on a particular interface.
- GraphRAG – global view
- Summarises structures and recurring themes across a corpus, such as major risk clusters in a large collection of project reports.
- Additional effort
- Entities, relationships, updates and permissions must be generated, maintained and evaluated reliably.
Multimodal, adaptive and agentic RAG
Other terms describe which media are included or how flexibly the system controls retrieval. They are not automatically better maturity levels; they address different requirements.
- Multimodal RAG
- Includes tables, diagrams, images, audio or video as well as text. Recognition, linking and source presentation must be evaluated for each medium.
- Adaptive RAG
- Chooses between no retrieval, a simple search or a multi-step workflow according to the question. The aim is to match effort and quality to actual complexity.
- Agentic RAG
- An agent plans retrieval steps, selects tools or sources, assesses intermediate results and may trigger further queries. This increases flexibility as well as latency, cost and the need for control.
- Self-RAG and corrective RAG
- These research approaches add self-assessment or an evaluation of retrieved material. Weak evidence should be detected and retrieval corrected when necessary.
Which approach fits which task?
The appropriate architecture follows from real questions, sources and failure risks rather than a product label. The following guide is a starting point; representative tests must provide the evidence for a decision.
| Starting point | Sensible first approach | What to verify before adding complexity |
|---|---|---|
| Direct questions over a clear document collection | Standard RAG with citations | Are the right passages retrieved and sufficient for a supported answer? |
| Natural questions plus exact codes or specialist terms | Hybrid RAG with filters and reranking | Which retrieval method produces better candidates for each query type? |
| Answers combine several documents or systems | Multi-hop or iterative RAG | Are intermediate steps traceable and protected against error propagation? |
| Relationships and corpus-wide patterns matter more than passages | GraphRAG or a structured graph query | Does the benefit justify building, maintaining and controlling the graph? |
| Essential information resides in tables, images or video | Multimodal RAG | Can the system read, cite and prioritise each medium when sources conflict? |
| The retrieval path varies and uses multiple tools | Adaptive or agentic RAG | Which steps, costs, permissions and stopping conditions need limits? |
Not every RAG component is a separate type of RAG
Many important terms refer to improvements within a RAG pipeline. They can be used across different architectures and should not be treated as complete solutions in their own right.
- Chunking
- Splits documents into retrievable passages. Size, overlap and structure determine whether enough context is preserved.
- Query rewriting and HyDE
- Rephrase a query or generate a hypothetical target passage so retrieval can find better candidates.
- Reranking
- Assesses initial candidates more precisely and moves the passages most likely to be useful to the top.
- Vector database
- A possible technical component for semantic retrieval, but not a complete RAG system by itself.
- Fine-tuning
- Changes model behaviour or knowledge through training. It is a different method and does not replace current, permission-aware retrieval.
Add complexity only for an evidenced reason
A more sophisticated architecture is not automatically more accurate. Every additional retrieval step can introduce new errors, latency and cost. A simple reference solution should first show which questions it already answers reliably and where it fails.
- Step 1
Define query classes and quality targets
Collect real direct, exact, multi-step and, where relevant, multimodal questions together with expected evidence.
- Step 2
Clarify authoritative sources and permissions
Determine which information governs, how current it must be and who may see it.
- Step 3
Start with a simple baseline
Measure retrieval, evidence coverage, answer faithfulness, latency and cost with a traceable reference architecture.
- Step 4
Extend deliberately
Add hybrid, multi-hop, graph, multimodal or agentic components only for demonstrated gaps.
- Step 5
Evaluate retrieval and generation separately
Distinguish whether the correct information was not retrieved or whether the model mishandled evidence that was available.
One platform can combine several RAG patterns
A manufacturer starts with hybrid RAG for work instructions so that natural questions and exact equipment codes both work. A question connecting suppliers, expiring certificates and open quality cases uses several sources in sequence. A graph only becomes worthwhile when dependencies between suppliers, parts and plants are analysed regularly. Technical drawings are added through a separately evaluated multimodal path.
What to remember
Treat RAG terminology as combinable architecture properties, not a ranking. Start from real questions and a simple baseline; extend retrieval, data representation and orchestration only where tests demonstrate a concrete need.
Sources and further reading
These primary sources provide further detail on definitions, technical foundations or responsible use.
Content reviewed
Research uses overlapping taxonomies. The terms in this article provide a practical classification of established architecture patterns and do not represent a binding standard.
- Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
- Gao et al.: Retrieval-Augmented Generation for Large Language Models – A Survey
- Microsoft Research: From Local to Global – A Graph RAG Approach
- Tang and Yang: MultiHop-RAG – Benchmarking RAG for Multi-Hop Queries
- Jeong et al.: Adaptive-RAG – Adapting Retrieval to Question Complexity
- Chen et al.: MuRAG – Multimodal Retrieval-Augmented Generator
- Asai et al.: Self-RAG – Retrieve, Generate and Critique through Self-Reflection
- Yan et al.: Corrective Retrieval Augmented Generation