26 types of RAG systems

Briefing on Retrieval-Augmented Generation (RAG) Architectures

The landscape of Retrieval-Augmented Generation (RAG) has evolved significantly from a single foundational concept into a diverse ecosystem of at least 26 distinct architectures. This expansion reflects a growing need for specialized AI systems capable of addressing complex, domain-specific challenges. These architectures can be systematically categorized along seven key dimensions: architectural patterns, process flow, retrieval methods, context and memory handling, intelligence and autonomy, data modality, and application scope.

The selection of an appropriate RAG architecture is a critical strategic decision, governed by a series of trade-offs between complexity, latency, accuracy, and cost. High-accuracy models tailored for critical applications, such as Self-RAG, Speculative RAG, and Recursive RAG, inherently demand higher complexity, cost, and latency. Conversely, foundational models like Standard RAG offer lower cost and complexity, making them suitable for simpler Q&A tasks.

Specialized RAG variants have been developed to meet specific operational requirements. Federated RAG addresses privacy-first use cases in regulated industries like healthcare. Streaming RAG and Temporal RAG are designed for real-time data feeds common in finance and news. Conversational applications are best served by Conversational RAG, Contextual Retrieval RAG, and Memory-Augmented RAG, which maintain state and context across interactions. Systems requiring reasoning over structured data can leverage Graph RAG or Knowledge-Enhanced RAG. A structured decision-making framework, based on primary project constraints such as speed, accuracy, or data type, is essential for navigating this complex landscape and selecting the optimal architecture.

Overview of 26 RAG Architectures

The following section details the 26 identified RAG architectures, each designed for distinct purposes and operational contexts.

1. Standard RAG: The foundational model combining retrieval from a vector database with LLM-based generation to provide grounded responses.
2. Agentic RAG: Empowers AI agents with autonomous retrieval and tool-use capabilities, enabling them to decide when to retrieve and how to chain actions for complex workflows.
3. Graph RAG: Utilizes knowledge graphs to perform relational reasoning by traversing entities and relationships, ideal for expert systems in fields like law and medicine.
4. Modular RAG: Decomposes the RAG pipeline into independent, swappable components for retrieval, reasoning, and generation, facilitating scalable and collaborative development.
5. Memory-Augmented RAG: Incorporates persistent external memory to retain context across sessions, enabling long-term conversational memory and personalization.
6. Multi-Modal RAG: Processes and retrieves from multiple data types simultaneously, including text, images, and audio, for applications like video summarization.
7. Federated RAG: Facilitates privacy-preserving retrieval from decentralized data sources, where data remains at its origin and only query results are aggregated.
8. Streaming RAG: Performs real-time retrieval and generation on continuously updating data streams, suitable for live dashboards and monitoring systems.
9. ODQA RAG (Open-Domain QA): Designed to handle large, diverse, and unconstrained datasets, making it ideal for general-purpose search engines and virtual assistants.
10. Contextual Retrieval RAG: Maintains session-level awareness, using the history of a conversation to inform subsequent retrievals and maintain coherence.
11. Knowledge-Enhanced RAG: Integrates structured data sources (databases, ontologies) with unstructured text retrieval for more informed generation.
12. Domain-Specific RAG: Tailored for a particular industry (e.g., finance, healthcare) by optimizing embeddings, chunking strategies, and prompts for that vertical.
13. Hybrid RAG: Combines multiple retrieval methods, typically sparse (keyword-based) and dense (vector-based), to bridge structured and unstructured data for higher precision.
14. Self-RAG: Introduces a self-reflection loop where the model critiques its own retrieved documents and generated answers, iterating to improve factuality and reasoning.
15. HyDE RAG (Hypothetical Document Embeddings): The LLM first generates a hypothetical ideal answer document and then uses its embedding to find similar real documents, bridging semantic gaps for abstract queries.
16. Recursive / Multi-Step RAG: Executes multiple cycles of retrieval and generation, where the output of one step becomes the input for the next, enabling complex problem-solving.
17. Corrective RAG (CRAG): Implements a quality gate that evaluates the relevance of retrieved documents. If they are insufficient, it triggers alternative retrievals or web searches.
18. Adaptive RAG: Dynamically adjusts its retrieval strategy based on the query's complexity, routing simple questions to a standard pipeline and complex ones to a multi-step process.
19. Speculative RAG: Generates multiple draft answers in parallel using different subsets of retrieved context, then verifies and selects the best one, trading compute for higher accuracy.
20. Hierarchical RAG: Organizes information in layers (e.g., summaries, sections, details) and retrieves at the appropriate level of granularity, improving efficiency in large document collections.
21. Temporal RAG: Prioritizes retrieval based on time, weighting recent information more heavily, which is critical for evolving domains like news and financial markets.
22. Query Decomposition RAG: Breaks a complex, multi-part question into simpler sub-questions, retrieves information for each, and then synthesizes a comprehensive final answer.
23. Reranking RAG: Adds a second, more precise reranking stage after the initial retrieval to refine the set of documents passed to the generator, improving precision.
24. Ensemble RAG: Combines results from multiple, diverse retrieval methods simultaneously (e.g., dense, sparse, semantic) to achieve more robust and comprehensive recall.
25. Conversational RAG: Reformulates user queries by incorporating dialogue history to resolve pronouns and implicit references, ensuring coherent multi-turn conversations.
26. Compressed RAG: Summarizes or compresses retrieved documents before sending them to the LLM, allowing more information to fit within the model's context window while preserving key details.

A Taxonomy of RAG Systems

The 26 RAG architectures can be organized into a taxonomy based on seven primary dimensions, providing a framework for understanding their core characteristics and differentiators.

1. Architectural Patterns

Describes how the RAG system is structurally organized.

• Monolithic: Standard RAG, ODQA RAG
• Modular: Modular RAG, Ensemble RAG, Hybrid RAG
• Hierarchical: Hierarchical RAG, Graph RAG
• Distributed: Federated RAG

2. Process Flow Patterns

Defines how the retrieval and generation components interact over time.

• Single-Pass: Standard RAG, Reranking RAG
• Iterative: Recursive RAG, Multi-Step RAG, Query Decomposition RAG
• Self-Refining: Self-RAG, Corrective RAG (CRAG)
• Speculative: Speculative RAG, HyDE RAG

3. Retrieval Method

Categorizes the underlying technique used to find relevant information.

• Dense Vector: Standard RAG and most other variants (as a default)
• Sparse/Hybrid: Hybrid RAG, Ensemble RAG
• Graph-Based: Graph RAG, Knowledge-Enhanced RAG
• Hypothetical: HyDE RAG

4. Context & Memory

Outlines how the system handles temporal and conversational state.

• Stateless: Standard RAG, ODQA RAG
• Session-Aware: Contextual Retrieval RAG, Conversational RAG
• Persistent: Memory-Augmented RAG
• Time-Sensitive: Temporal RAG, Streaming RAG

5. Intelligence & Autonomy

Refers to the system's level of autonomous decision-making.

• Passive: Standard RAG
• Adaptive: Adaptive RAG, Query Decomposition RAG
• Autonomous: Agentic RAG
• Self-Improving: Self-RAG, Corrective RAG

6. Data & Modality

Describes the types of information the system is designed to process.

• Text-Only: Standard RAG and most other variants
• Multi-Modal: Multi-Modal RAG
• Structured: Graph RAG, Knowledge-Enhanced RAG
• Compressed: Compressed RAG

7. Application Scope

Defines the breadth of the system's domain coverage.

• General-Purpose: Standard RAG, ODQA RAG, Agentic RAG, most architectural variants
• Domain-Specific: Domain-Specific RAG, Knowledge-Enhanced RAG, Graph RAG (specialized)

Strategic Decision Guide

To select the appropriate RAG architecture, identify the primary constraint or requirement of the project and consult the following recommendations.

• For Speed/Latency: Consider Streaming RAG, Compressed RAG, or Standard RAG.
• For Highest Accuracy: Consider Self-RAG, Corrective RAG (CRAG), Speculative RAG, or Recursive RAG.
• For Privacy: The primary choice is Federated RAG.
• For Multi-Modal Data: The designated architecture is Multi-Modal RAG.
• For Complex Reasoning: Consider Agentic RAG, Recursive RAG, or Query Decomposition RAG.
• For Conversational Applications: Consider Conversational RAG, Contextual Retrieval RAG, or Memory-Augmented RAG.
• For Structured Data Integration: Consider Graph RAG, Knowledge-Enhanced RAG, or Hybrid RAG.
• For Real-Time Data Feeds: Consider Streaming RAG or Temporal RAG.
• For Scalability: Consider Modular RAG or Hierarchical RAG.
• For a Starting Point: Begin with Standard RAG, then incrementally add capabilities such as Reranking RAG, and later evolve to a Hybrid RAG.