What is RAG: 25 Different Types of RAG Architectures

Ever had one of those days when you’re standing in line at a store, waiting for a sales assistant to help you find a product? You know the drill: waiting... waiting... and waiting! As an AI engineer, you start thinking, "Wait a minute—why don’t they just have an AI assistant to speed this up?"

But then, it hits you. AI Assistants are the future, and the secret to making them super efficient lies in one mysterious force. This is what RAG is. RAG stands for Retrieval-Augmented Generation, and it's the ultimate behind-the-scenes technology that powers many AI assistants today.

RAG stands for Retrieval-Augmented Generation, and it's the ultimate behind-the-scenes technology that powers many AI assistants today.

What is RAG?

Now, imagine my surprise when I discovered that RAG isn't just one thing— there are 25 different types of RAG! Yes, you read that right—25 flavors of RAG, each designed to solve specific problems and to handle user queries in unique ways. It's like an AI buffet, with each RAG offering a different solution to different problems, and trust me, there's something for everyone.

In this blog, we’ll take a whirlwind tour through this AI mall of RAGs, exploring 25 different types of RAG that can be your AI assistant's best friend. From the Agentic RAG to the Auto RAG, let's explore which one might just be your new favorite assistant. And don't worry, we promise this isn't just a techy lecture—we’re going full-on quirky!

So, let's take a short tour through the "Mall of RAG"!

Store 1: Corrective RAG – The Self-Critic

“The guy who speaks bad about himself so that others don’t get a chance to.”

This type of RAG goes into the response it generates and looks for errors, corrects them, and generates the corrected output to the user. Basically, it performs a factual check before giving the output.

Pros: Fixes errors before they happen.
Cons: A bit of a perfectionist, can be slow.

Best for: Healthcare and finance—where a small error could cost big.

Store 2: Speculative RAG – The Over-Achiever

“The guy who preps up well before a meeting to present the un-demanded.”

This type of RAG understands the context and the need of the user and thus keeps what the user might ask next ready. It preemptively answers questions to save time and ensure user satisfaction.

Pros: Ready for any question, even before you ask it.
Cons: Always a step ahead, but sometimes doesn't get it quite right.

Best for: Time-sensitive platforms like customer service (think Amazon) where speed matters.

Store 3: Agentic RAG – The Caring Sibling

“The Understanding SIBLING RAG, who cares for you, grows up with you and suggests the best for you.”

This RAG grows with every interaction you have in the form of a query. With each interaction, it tries to understand the user more and make responses more explicit. It doesn’t update, but it evolves.

Pros: Grows with every interaction, learns your needs.
Cons: Doesn't update but evolves with time.

Best for: McDonald's self-order systems, always suggesting that extra portion of fries.

Store 4: Refeed Retrieval Feedback RAG – The Employee Who Listens

“The obedient employee who takes the feedback to improve.”

This type of RAG learns from user interactions. When the user points out what's wrong, it takes that feedback to improve future responses.

Pros: Learns from user feedback and improves.
Cons: Might get carried away, focusing too much on small feedback.

Best for: Customer support bots that get better with each chat.

Store 5: Raptor RAG – The Speed Demon

“No, it’s not what you are thinking, it’s the speed-of-light guy.”

This RAG organizes the information in the form of tree branches, so whenever a user asks a query, it follows the branch of information and quickly retrieves relevant answers. It’s fast, no time to waste!

Pros: Quick as lightning, organizes info like a pro.
Cons: Sometimes too fast for its own good, leaves out details.

Best for: Medical diagnoses, where speed is crucial in saving lives.

Store 6: Memo RAG – The Forget-Me-Not

“The girlfriend, who remembers all the dates.”

This RAG keeps track of all past conversations. It remembers the key points and aligns new queries with them to ensure that the context is never lost.

Pros: Remembers everything, even the smallest details.
Cons: Sometimes it remembers too much.

Best for: Amazon customer support bots that recall your previous complaints.

Store 7: GenerativeAI RAG – The Smart Cheater

“The smart cheater, who changes the language after copying.”

This RAG generates responses by pulling out past information and creatively shaping it into a new response while keeping the core meaning intact.

Pros: Masters the art of recycling information creatively.
Cons: Could start sounding like a copycat after a while.

Best for: Marketing and advertising, creating new content from the same old.

Store 8: Replug RAG – The Social Butterfly

“The extrovert connecting child.”

This RAG connects to external databases to retrieve real-time information when needed. It doesn’t have access to everything but taps into other sources to get the info.

Pros: Connects to external sources for real-time info.
Cons: A little too reliant on others; needs to be plugged in.

Best for: Real-time stock price predictions, or weather forecasting apps.

Store 9: Replug Retrieval Feedback – The Overly-Cautious Child

“The extrovert child who checks with his mom if he did right.”

This RAG connects with external sources and constantly takes feedback on which sources were useful and which weren't, refining its results over time.

Pros: Connects to external databases and checks its sources every time.
Cons: Too much feedback, can get stuck in the loop of 'was that right?'

Best for: Market insights tools that need constant refinement.

Store 10: Adaptive RAG – The Chameleon

“The adjusting kid.”

This type of RAG analyses the environment of the user, and as the context of the query changes, it adjusts its responses and the information it retrieves.

Pros: Changes its responses based on your environment.
Cons: Hard to predict, sometimes takes a while to adjust.

Best for: Event-based industries like stand-up comedy, where context matters!

Store 11: REALM (Retrieval Augmented LLM) – The Know-It-All

“The know-it-all guy.”

This type of RAG uses LLMs to retrieve the information. Using LLM, it knows exactly where the answer to a query is and gives context-specific, in-depth information.

Pros: Uses deep insights from LLM to answer with accuracy.
Cons: Can overwhelm you with too much information.

Best for: Legal firms needing in-depth answers about regulations.

Store 12: Attention-Based RAG – The Focused One

“The attentive guy.”

This RAG focuses only on the important parts of the user’s query, discarding irrelevant details to ensure the information retrieved is appropriate and hallucination-free.

Pros: Zeroes in on the important bits.
Cons: Can be too laser-focused, missing out on the broader context.

Best for: Research industries needing precise information.

Store 13: Cost-Constrained RAG – The Bargain Hunter

“The guy who wants ZARA in ज़रा.”

This type of RAG retrieves information based on budgetary constraints. Whatever it retrieves, it ensures it fits within the cost limits.

Pros: Works within a budget—no overspending here!
Cons: Sometimes you get what you pay for—limited results.

Best for: Nonprofits and small businesses with tight budgets.

Store 14: Rule-Based RAG – The Rule Follower

“The compliant guy.”

This type of RAG retrieves information strictly by following a set of guidelines and rules, ensuring that it adheres to regulations.

Pros: Always compliant, never breaks the rules.
Cons: Too rigid, doesn’t adapt easily.

Best for: Financial advisory firms where compliance is king.

Store 15: Attention UNet RAG – The Medical Expert

“The segmenting guy.”

This RAG uses attention mechanisms to segment medical data, such as medical images, to ensure precision in diagnosis.

Pros: Perfect for segmenting medical images with precision.
Cons: Requires very specific setups.

Best for: Medical imaging—where every pixel counts.

Store 16: Self RAG – The DIY Master

“The I am my own boss guy.”

This RAG looks for ways to improve its responses on its own. It learns through every interaction, continuously enhancing its performance.

Pros: Learns and improves with every step it takes.
Cons: Sometimes too confident in its own improvement.

Best for: Stock price analysis, where the RAG can fine-tune its responses.

Store 17: RETRO (Retrieval Enhanced Transformer) – The Nostalgic One

“The souvenir guy.”

This RAG keeps the historical context of every response—whether documents, conversations, or other content. It uses this history to generate more accurate answers.

Pros: Uses past responses to improve future ones.
Cons: Can get stuck in the past sometimes.

Best for: Corporate companies needing to refer back to old project insights.

Store 18: ECO RAG – The Green Machine

“The environment-friendly guy.”

This RAG retrieves information while considering environmental factors and minimizing energy consumption.

Pros: Helps save the environment while retrieving data.
Cons: Can be slow, trying to balance efficiency and energy use.

Best for: Environmental monitoring platforms.

Store 19: Conversational RAG – The Chatterbox

“The talker guy.”

This RAG creates a bot that responds in real time, enabling engaging, back-and-forth conversations with users.

Pros: Engages in real-time conversations, just like a human.
Cons: Can get lost in endless chatter.

Best for: Chatbots in customer service.

Store 20: HybridAI RAG – The Polymath

“Polymorphism guy.”

This RAG combines strengths from various machine learning models to provide the best responses, incorporating multiple approaches for accuracy.

Pros: Combines strengths from different models for a top-tier response.
Cons: Sometimes it’s just trying too hard.

Best for: Predictive maintenance platforms that need varied responses.

Store 21: XAI (Explainable RAG) – The Justifier

“The guy who justifies his actions.”

This RAG not only provides responses but also adds explanations for why a particular response was chosen, ensuring transparency.

Pros: Explains why it gave that answer.
Cons: May over-explain and go in circles.

Best for: Healthcare, where you need to know why certain treatments are recommended.

Store 22: Grokking RAG – The Deep Thinker

“The philosopher guy.”

This RAG goes beyond the surface to explain complex concepts in detail, making it ideal for research purposes.

Pros: Breaks down complex concepts into simpler, digestible pieces.
Cons: Sometimes it goes too deep, making things harder to understand.

Best for: Research assistants that need to dive into the depths of theory.

Store 23: Context Cache in LLM – The Memory Keeper

“The Cash (Ca-che) guy.”

This RAG stores important context and uses it for future conversations, helping generate responses based on previous information.

Pros: Keeps track of important context for future conversations.
Cons: Memory overload can happen.

Best for: Virtual tutors that need context from previous lessons.

Store 24: Iterative RAG – The Perfectionist

“The complaining guy.”

This RAG learns and improves with each iteration. It doesn’t settle for "good enough"—it’s always striving for better.

Pros: Learns and improves with every iteration.
Cons: Sometimes gets stuck trying to make things perfect.

Best for: Tech bots that need to improve continuously.

Store 25: Auto RAG – The "I Can Do It" Guy

“I can do it guy.”

This RAG automates the retrieval process, adapting without any human intervention. It's a self-sufficient worker.

Pros: Automates the retrieval process without needing help.
Cons: Can adapt too quickly, sometimes without thinking things through.

Best for: News apps fetching the latest headlines.

The Grand Finale: Choose Wisely!

And with that, our mall of RAGs has come to an end. Which one will you choose for your problem statement? Will it be the Agentic RAG or the Eco RAG? Maybe you’ll even go for the Speculative RAG, ready with answers before the user even asks! Let me know which RAG fits your needs and stay tuned for the next deep dive into how to implement one of these.