Generative AI 10 Slides

Types of RAG.

Slide 1 / 10

Use ← and → to navigate

Swipe left / right on mobile

Guide Notes & Explanation

Accompanying breakdown for this slide deck

Naive RAG follows the simplest RAG pipeline: retrieve relevant information and give it to the LLM for generation.
Documents are typically chunked, embedded, stored, and retrieved using similarity search.
The retrieved chunks are added to the prompt as context for the LLM.
It is simple to build and works well for straightforward question-answering tasks.
Its main weakness is that poor retrieval can directly lead to poor answers.

Advanced RAG improves the basic pipeline with techniques before and after retrieval.
Query rewriting, query expansion, better chunking, metadata filtering, and improved embeddings can increase retrieval quality.
Retrieved documents can also be re-ranked to place the most relevant information first.
Context compression can remove unnecessary information before it reaches the LLM.
The goal is to make the same basic RAG pipeline more accurate, efficient, and reliable.

Modular RAG breaks the RAG pipeline into independent components that can be combined and replaced.
Different modules can handle retrieval, query rewriting, routing, re-ranking, generation, or memory.
The system can use different workflows depending on the type of question.
This makes RAG systems easier to customize for different business and technical requirements.
Modular RAG moves RAG from a fixed pipeline toward a more flexible architecture.

Graph RAG represents information using entities and relationships rather than relying only on isolated text chunks.
A knowledge graph can connect people, organizations, products, events, and concepts.
Retrieval can then follow relationships between entities to discover connected information.
This is particularly useful for questions requiring connections across multiple pieces of information.
Graph-based retrieval can improve performance on complex, multi-hop questions.

Multimodal RAG retrieves information from different data types such as text, images, tables, charts, audio, and video.
Instead of treating a document as text only, the system can retrieve the most relevant visual or structured information.
For example, a financial assistant could retrieve both the text explanation and the relevant chart from a report.
This makes RAG useful for documents where important information is not represented purely as text.
Multimodal RAG extends retrieval from a text-based problem to a multi-format knowledge problem.

Corrective RAG introduces mechanisms to evaluate whether the retrieved information is actually useful.
If retrieval quality is poor, the system can take corrective actions instead of blindly generating an answer.
It may refine the query, retrieve additional information, or use another knowledge source.
This reduces the risk of generating an answer from irrelevant or insufficient context.
The key idea is: retrieve first, evaluate the evidence, and correct the retrieval when necessary.

Self-RAG allows the model to determine when retrieval is needed and reflect on the retrieved information.
Instead of retrieving information for every query, the system can decide whether external knowledge is necessary.
It can evaluate retrieved passages and assess aspects of its own generated response.
This creates a more adaptive retrieval-and-generation process.
The goal is to improve factuality and relevance while avoiding unnecessary retrieval.

Agentic RAG combines retrieval with agent-style reasoning, planning, and tool use.
Instead of following one fixed retrieval pipeline, the agent can decide what information it needs and where to find it.
It can perform multiple retrieval steps, use different data sources, call tools, and refine its search based on intermediate results.
This makes it suitable for complex tasks that require multi-step reasoning and information gathering.
Agentic RAG represents a shift from "retrieve once and generate" toward "reason, retrieve, evaluate, and act."