12 Essential RAG Types

(A Comprehensive Guide)

Karn Singh
 What is Retrieval-Augmented Generation (RAG)
Retrieval-Augmented Generation (RAG) is a powerful technique that enhances the capabilities of large language models

(LLMs) by integrating external knowledge sources. There are various types of RAG models, each designed to address specific

needs and improve the performance of information retrieval and generation tasks.
Types of RAG Model

Agentic
RAG
RAG
Naive RAG
Definition: Naive RAG enhances large language models (LLMs) by integrating external knowledge into their responses.

Purpose: Addresses LLM limitations, particularly the inability to access real-time or updated information.
Naive RAG

Key Steps in Naive RAG

• Document Chunking:

• Large documents are divided into smaller, manageable chunks for efficient retrieval.

• Embedding Model:

• Both document chunks and user queries are converted into numerical representations (embeddings) for semantic

comparison.

• Retrieval:

• Relevant document chunks are retrieved from an indexed database based on user query embeddings.

• Response Generation:

• The LLM generates coherent responses using the retrieved chunks and the original query, ensuring relevance and

context.
Advanced RAG
Definition: Advanced RAG enhances Naive RAG by integrating sophisticated techniques for better retrieval and generation.

Techniques Used

• Re-ranking: Prioritizes retrieved documents

based on relevance.

• Dynamic Embeddings: Adjusts embeddings for

specific tasks or domains.

• Hierarchical Indexing: Organizes data into a

structured hierarchy for improved retrieval.

• Corrective RAG (CRAG): Scores and filters

documents for relevance and accuracy.

Advanced RAG

Goals:

• Enhance efficiency, accuracy, and relevance of information retrieval.

• Tackle complex queries and handle diverse data sources effectively.

Advantages Over Naive RAG:

• Better relevance and coherence in responses due to advanced filtering and re-ranking.

• Enhanced query optimization through methods like query rewriting.

• Improved scalability for handling larger datasets efficiently.

Applications: Suitable for complex applications requiring higher precision, such as advanced question-answering systems and AI

chatbots.

Advanced RAG represents a significant evolution in the capabilities of RAG systems, addressing key challenges faced by Naive

RAG implementations and providing a more robust framework for generating accurate, contextually rich responses.
Modular RAG
Definition: Modular RAG enhances traditional RAG systems by introducing modularity for improved flexibility and performance.

Key Components

• Customizable Retrievers: Tailored retrieval

mechanisms for specific use cases, enhancing

efficiency and relevance.

• Adaptive Generators: Generative models that

integrate with various retrievers for better coherence

and accuracy.

• Plug-and-Play Modules: Components that can be

easily added or replaced for system customization.

Modular RAG

Applications : Suitable for diverse applications, including customer support chatbots and advanced question-answering

systems, where tailored solutions are essential.

Modular RAG represents a significant evolution in retrieval-augmented techniques, addressing the limitations of Naive RAG

and providing a robust framework for building adaptable AI systems.

Query-Based Retrieval-Augmented Generation (QB-RAG)
Definition: QB-RAG optimizes retrieval by pre-computing a database of potential queries, improving alignment between user

questions and relevant content.

Key Features

• Query Pre-computation: Generates a comprehensive

set of potential queries from the knowledge base to

facilitate efficient retrieval.

• Vector Search: Utilizes vector search techniques to

match incoming user queries against the pre-

generated query database, enhancing retrieval

accuracy.

• Semantic Alignment: Focuses on aligning user queries

with content across distinct semantic representations,

addressing gaps in traditional retrieval methods.

Query-based RAG

Advantages

• Improved Accuracy: Empirical evaluations show that QB-RAG significantly enhances the accuracy of

responses, particularly in healthcare question-answering applications.

• Robustness: Provides a more reliable framework for applications requiring trustworthy responses from

LLMs.

Applications

• Particularly effective in digital health chatbots and other domains where accurate, real-time

information is critical.

QB-RAG represents a significant advancement in retrieval techniques for RAG systems, addressing existing

challenges in aligning user queries with relevant knowledge effectively.

Logit-based RAG
Definition: Combines retrieval information with generative models using logits (raw output values before softmax) during the

decoding process.

Key Features

• Logit Integration: Integrates relevant retrieved

information into generation through logits,

enabling nuanced decision-making.

• Augmentation Methodology: Allows retrieved

results to influence generation stages, either as

input or by modifying logits directly.

Logit-based RAG

Advantages

• Enhanced Relevance: By using logits, this approach can better determine which retrieved information is most relevant to

the current query.

• Improved Output Quality: The integration of retrieval data through logits helps generate more accurate and contextually

appropriate responses.

Applications

• Particularly useful in scenarios requiring high accuracy and contextual relevance, such as advanced question-answering

systems and AI-generated content.

Logit-Based RAG represents a significant advancement in retrieval techniques, effectively leveraging the strengths of

generative models to improve response quality and relevance.

Latent Representation-based RAG
Definition: Incorporates retrieved data as latent representations within generative models to enhance comprehension and

output quality.

Key Features

• Latent Representation Integration: Integrates

retrieved objects at a deeper level, influencing the

model's hidden states during generation.

• Enhanced Comprehension: Provides a nuanced

understanding of context through the use of latent

representations.
Latent Representation-based RAG

Advantages

• Improved Output Quality: Enhances the relevance and quality of generated outputs.

• Adaptability Across Modalities: Suitable for various applications, including text, code, image, and audio generation.

Applications

• Effective in fields such as natural language processing, computer vision, and audio processing, where accurate and

contextually relevant outputs are essential.

Latent Representation-Based RAG signifies a notable advancement in retrieval techniques, leveraging sophisticated

algorithms to incorporate external information into generative processes effectively.

Speculative RAG
Definition: Speculative RAG enhances retrieval-augmented generation by using a larger generalist language model (LM) to verify

multiple drafts generated in parallel by a smaller, specialized LM.

Speculative RAG

Key Features

• Drafting and Verification: Separates drafting (specialist LM) from verification (generalist LM) to improve efficiency.

• Parallel Draft Generation: Creates multiple drafts from distinct subsets of retrieved documents, allowing for diverse

perspectives.

• Efficiency: Offloads drafting to a smaller model, accelerating response generation while maintaining accuracy.

Advantages

• Enhanced Accuracy: Improves accuracy by up to 12.97% on benchmarks like TriviaQA and PubHealth.

• Reduced Latency: Lowers response times by 51% compared to traditional RAG systems.

Applications

• Particularly effective in knowledge-intensive tasks such as question answering, where timely and accurate information

retrieval is crucial.
Speculative RAG
Self Reflective RAG
Definition: Enhances language models (LMs) by enabling on-demand retrieval of relevant passages and self-reflection on

generated outputs using reflection tokens.

Self RAG

Key Features

• Reflection Tokens: These tokens signal the need for retrieval or assess the quality of generated outputs, allowing the model to adapt its behavior

during inference.

• Adaptive Retrieval: The framework allows LMs to determine when to retrieve additional information based on the context of the input and previous

generations.

• End-to-End Training: Self-RAG trains a single LM to generate text informed by retrieved passages while also critiquing its own outputs.

Advantages

• Improved Factuality: Significantly enhances the accuracy of responses, outperforming state-of-the-art LMs in tasks like open-domain QA and fact

verification.

• Versatility: Maintains the original creativity and versatility of LMs while improving their factual accuracy.

Applications

• Effective in various tasks requiring high-quality, factual responses, such as question answering, reasoning, and content generation.
 Branched RAG
An advanced framework that enhances standard RAG through a structured, multi-step approach for retrieval and response

generation..
Branched RAG
Key Features

•Multiple Retrieval Steps: Conducts sequential retrievals to gather information progressively.

•Hierarchical Structure: Uses a branching pattern where each retrieval informs the next, allowing deeper topic exploration.

•Specialized Knowledge Bases: Different branches can query distinct knowledge bases tailored to specific sub-topics.

•Dynamic Query Refinement: Refines queries based on intermediate results for focused and relevant retrievals.

Workflow

1.Initial Broad Retrieval: Captures potentially relevant information for context.

2.Intermediate Retrievals: Narrows the search space based on initial results.

3.Final Focused Retrieval: Yields highly relevant information, improving precision.

4.Generation Step: Synthesizes information from multiple retrieval steps for comprehensive responses.

Applications

•Effective for complex queries requiring multi-step reasoning or synthesis of information, such as in legal tools or

multidisciplinary research.
Agentic RAG
Definition: Integrates AI agents into the RAG framework to enhance information retrieval and processing capabilities for complex

tasks.
Agentic RAG
Key Features

• Agent-Based Architecture: Uses agents to orchestrate retrieval processes and make sourcing decisions.

• Tool Integration: Agents access various tools (e.g., vector search engines, web searches, APIs) for information gathering.

• Dynamic Decision-Making: Agents evaluate when to retrieve information and which tools to use based on context.

Advantages

• Enhanced Flexibility: Supports multi-step retrieval processes and adaptive responses to complex queries.

• Improved Accuracy: Agents validate retrieved information, leading to more robust outputs.

Applications

• Effective in real-time adaptive responses, such as automated customer support, internal knowledge management, and

research assistance.
Adaptive RAG
Definition: Adaptive RAG dynamically adjusts its retrieval strategy based on the complexity or nature of the query, enhancing the

accuracy and relevance of responses.

Adaptive RAG
Key Features

•Dynamic Retrieval Strategy: Alters retrieval methods in real-time, using a single source for simple queries and multiple

sources for complex ones.

•Query Classification: Determines the type of query (factual, analytical, opinion-based, contextual) to apply appropriate

retrieval strategies.

•LLM Integration: Utilizes language models at different stages to optimize document ranking and response generation.

Advantages

•Tailored Responses: Ensures customized responses for diverse query types, bridging the gap between precision and

breadth.

•Enhanced Efficiency: Improves information retrieval speed and relevance by adapting to query characteristics.

Applications

•Suitable for environments with varied query types, such as search engines and AI assistants, where dynamic adjustments

are crucial for effective information retrieval.

Corrective RAG
CRAG is a strategy that incorporates self-reflection and self-grading on retrieved documents to enhance the accuracy and

relevance of generated responses.

Corrective RAG

Key Features

•Self-Reflection: Evaluates the relevance of retrieved documents before generation.

•Knowledge Refinement: Partitions documents into "knowledge strips" and grades each for relevance.

•Supplementary Retrieval: If documents fall below a relevance threshold, CRAG performs additional retrievals, such as web searches.

Workflow

1.Initial Retrieval: Retrieves documents based on the input query.

2.Grading Documents: Evaluates each document's relevance to the query.

3.Knowledge Refinement: Filters out irrelevant knowledge strips.

4.Supplemental Search: Uses web searches if necessary to find additional relevant information.

5.Response Generation: Generates a response using refined and relevant information.

Applications

•Ideal for scenarios requiring high factual accuracy, such as legal document generation, medical diagnosis support.
WAS THIS POST USEFUL?

FOLLOW FOR
MORE!