Being-Creative-with-AI/C4/Understanding-the-RAG-pipeline/English

Title of the Script: Understanding the RAG Pipeline.

Author: EdyPyramids Team.

Keywords: RAG, Retrieval Augmented Generation, Artificial Intelligence, Embeddings, Vector Database, Vector Search, AI Chatbot, Document Chunking, AI Pipeline, Query Processing, Information Retrieval, Generative AI, EduPyramids, video tutorial.

Visual Cue	Narration
Slide 1 Title Slide	Welcome to this Spoken Tutorial on Understanding the RAG Pipeline.
Slide 2 Learning Objectives	In this tutorial, we will learn: The stages of a RAG pipeline. How a user query is processed. How retrieval improves the generated response.
Slide 3 Disclaimer Slide As AI tools constantly evolve, if you are unable to locate any icon or encounter difficulty at any step, you may use any conversational AI Chatbot for guidance.	As AI tools constantly evolve, if you are unable to locate any icon or encounter difficulty at any step, you may use any conversational AI Chatbot for guidance.
Slide 4 System Requirements	To record this tutorial, I am using: Ubuntu 24.04 LTS, Firefox version 148.0.2 Learners will also need a working internet connection
Slide 5 Prerequisites https://EduPyramids.org	To follow this tutorial, Learners should know basic computer and internet usage. Basic AI knowledge is required. For the Prerequisite of this tutorial, visit the website shown on your screen
	Let us get started.
RAG pipeline diagram with Retrieval + Generation blocks	RAG is known as Retrieval Augmented Generation.
Flowchart showing sequence of stages	It is a sequence of steps used to answer a question.
AI system connected to external documents/database	It uses external data to generate accurate answers.
AI brain with “Stored Knowledge” and “Retrieved Knowledge” labels	The system does not rely only on stored knowledge.
Search and retrieval animation from documents	It retrieves relevant information to generate answers.
“Step-by-step process” illustration	Let us understand this step by step.
Grocery shopping chatbot interface	Consider a grocery app chatbot.
User typing a question into chatbot	Ask this question: “Can I return vegetables after 2 days?”
Human brain vs computer comparison graphic	Now the system cannot understand words directly like humans
Flowchart showing processing stages	So it follows a series of steps to find the answer.
Text converting into vectors/numbers animation	First, the system converts the question into a machine-understandable form.
Highlight the term “Embedding”	This process is called embedding.
Binary numbers or vectors beside words	Computers work with numbers, not words.
Words converting into numeric arrays	So every word is represented as a list of numbers.
Example vector representation of “vegetables”	For example, The word ‘vegetables’ may be converted into a list of numbers like: [0.21, 0.45, 0.78, …]
Side-by-side vectors for vegetables and fruits	Similarly, The word ‘fruits’ will have a different set of numbers.
Highlight closeness between vectors	But the values will be close to ‘vegetables’ because both are related.
Vector comparison illustration	Now, instead of comparing words, The system compares these numbers.
Similar vectors connected visually	If two sentences have similar vectors, their meanings are similar.
Question matched with stored documents	This lets the system compare the question with stored documents
Similarity measurement graphic	It helps measure sentence similarity.
Company documents or policy PDF shown	Next, the system searches company documents such as the return policy.
Large document splitting into blocks	These documents are often divided into smaller parts called chunks.
Three text boxes labeled Chunk 1, Chunk 2, Chunk 3	A return policy document may be split into chunks like this
Three text boxes labeled Chunk 1, Chunk 2, Chunk 3	Chunk 1: Return rules for fruits and vegetables Chunk 2: Return rules for packaged items Chunk 3: Refund process and timelines
Chunks converting into embeddings	Now, each chunk is converted into numbers using the same embedding process.
Vector database illustration with stored vectors	These number representations are stored in a database called a vector database.
Search icon over vector database	A vector database stores data in the form of numbers. It can quickly find similar meanings.
User query converting into vectors	When you ask: ‘Can I return vegetables after 2 days?’ Your question is also converted into numbers.
Query vector compared with stored chunk vectors	Then, the system compares your question with chunks in the vector database.
Best matching chunk highlighted	It finds the most similar chunk. For example: ‘Fresh vegetables can be returned within 24 hours only.’
Retrieved chunk sent to AI model	This relevant chunk is then used to generate the final answer.
Multiple retrieved policy lines displayed	The system compares the question with chunks and retrieves relevant chunks.
Highlight retrieved policy statements	The system may retrieve statements such as: “Perishable items cannot be returned.” “Items can be returned within 24 hours.”
Good retrieval vs bad retrieval comparison	Note that the quality of the final answer depends on the retrieved information.
Relevant chunk highlighted among many results	The system selects the most relevant information from the retrieved results.
Question and retrieved chunk combined visually	This information is combined with the original question.
Two blocks labeled Question and Policy Line	The system now has two things: Your question and the relevant policy line.
Context box formed from question + retrieved text	Together, these form the context The AI uses the background information to generate an answer.
Policy statement highlighted	For example, from the return policy, it may find this line: ‘Fresh vegetables can be returned within 24 hours only.’
Original question shown beside retrieved statement	Now, this information is combined with your original question:
Question displayed again prominently	‘Can I return vegetables after 2 days?’
Combined input entering AI model	So the system now has both, your question and the relevant policy information
Label “Context” shown clearly	This combined input is called context.
Definition text animation Context → AI → Response flowchart	In simple terms, context is the background information given to the system.
Final chatbot answer displayed on screen	Using this context, the system can now generate a correct response, like this: ‘No, vegetables cannot be returned after 2 days. As the policy allows returns only within 24 hours.
AI model reading retrieved chunk and question	The AI model reads the question and the retrieved information.
Generated answer appearing	It then generates an answer based on this information.
Alternative generated answer example	For example: “Vegetables are perishable items and cannot be returned after delivery.”
Incorrect guessing crossed out, retrieved answer ticked	The system reduces guess work and bases its answer on retrieved information.
Relevant answer highlighted	This leads to a more relevant response. Without retrieval, the answer may be generic or incorrect.
Three-step pipeline animation: Question → Retrieval → Answer	The process is as follows: A question is asked. Relevant information is retrieved. An answer is generated.
Wrong retrieval leading to incorrect answer illustration	Wrong or unrelated retrieval may lead to an incorrect answer.
	With this we come to the end of this tutorial.
Slide 6 Summary In this tutorial, we learnt: The stages of a RAG pipeline. How a user query is processed. How retrieval improves the generated response.	In this tutorial, we learnt: The stages of a RAG pipeline. How a user query is processed. How retrieval improves the generated response.
Slide 7 Assignment Create a simple grocery return policy with at least three rules. Ask the question: “Can I return vegetables after 2 days?” Then ask the same question again using your policy as context. Compare both responses and observe how retrieval improves the final answer. Also identify: the user question, the retrieved information, and the final answer.	We encourage you to do this assignment.
Slide 8 Acknowledgement Domain Inputs: Bhavani Shankar R and Saisudha Sugavanam Script Writer: Ketki Naina Admin Reviewer: Arthi Varadarajan Quality Reviewer: Sakina Sidhwa Novice Reviewer: Misbah Samir AI Narration: Debosmita Mukherjee AI Graphics: Arvind Pillai Video Editor: Arvind Pillai Web Developer: Ankita Singhal	Thank you for joining.
Slide 9 Acknowledgement This Spoken Tutorial is brought to you by EduPyramids Educational Services Private Limited at SINE, IIT Bombay.

Contributors and Content Editors

Ketkinaina