The next few minutes are about the eight stages inside a basic RAG chatbot, written as eight lines of code. RAG is retrieval augmented generation. If you know the word but cannot yet list what actually runs, this is for you.By the end, you can name the eight stages in order, and you can say where a chatbot keeps the conversation.Picture one help page you trust, the page that explains how to reset a password. You want a chatbot that answers from that page, not from whatever it absorbed in training.The first four stages run before anyone asks. Gather is first. Your help page, or a P D F, a scan, or a recording, becomes clean text.Split takes that text and breaks it into chunks, short pieces of a few sentences each.Embed turns each chunk into a vector. A vector is a list of numbers, and chunks with similar meaning get similar numbers.Store puts those vectors in a database, ready for a question.The last four stages run every time someone asks. Say a user types, how do I reset my password?Embed the question runs that same embed step on the question. Because it is the same step, the question's numbers can be compared with the chunks' numbers.Find similar asks the database for the chunks closest to the question. Those chunks are not the answer. They are the reading the model gets.Messages puts the question and those chunks together, in one request for the model.Then the model reads the messages and returns raw text. Your app turns that raw text into the reply the user sees.So here are all eight as one chain. Gather, split, embed, store, embed the question, find similar, messages, raw. Each line hands its result to a later line. Last tip in this series we called RAG retrieve, augment, generate. Embed the question and find similar are retrieve, messages is augment, and the model is generate.But these eight lines have a gap. They start over on every question. Watch what happens when the same user sends a follow-up. And what if the email never arrives? The chatbot has no idea they mean the reset email.One wrong turn is to push the old chat into Gather. Gather is for your documents, and it runs before anyone asks. Another wrong turn is to retrain the model so it remembers. Training is slow, and the chat changes on every turn.Memory, meaning the earlier turns of this chat, goes in Messages. Next to the new question and the chunks, you add those turns, the password question and the reply it got. Now the model can tell they mean the reset email. The other seven lines can stay as they are.So leave with this. A basic RAG chatbot is eight stages, the closest chunks are reading, not the answer, and the conversation lives in Messages.Now give the quiz under this tip a try. It names a stage and asks for its job. Then open the practice. It hands you the eight lines and a follow-up question, and you change Messages so the chatbot keeps up.Until the next tip, happy building.

A basic RAG chatbot in eight lines of code, from gather and embed to find similar and the model. Why chunks are not the answer, and where chat memory goes.

RAG in 8 lines, and where memory goes | Tech Tips by Moh

Embed the question, find similar, messages, the model

Gather, find similar, messages, the model

A vector, a list of numbers where similar meaning gets similar numbers

So the question's numbers can be compared with the chunks' numbers

So the question is saved in the database for next time

Into a new training run, so the model remembers

Into Messages, beside the new question and the chunks

RAG in 8 lines, and where memory goes

<div class="course-slide cs-center">
 This tip
 <h2 class="cs-title cs-at-1-5">RAG in 8 lines</h2>
 Retrieval augmented generation
 You know the word. Now the stages.
</div>

<div class="course-slide">
 By the end
 <h2 class="cs-title">Two things you can say</h2>
 <ul class="cs-checks">
 <li class="cs-at-0-9">The 8 stages, in order</li>
 <li class="cs-at-3-1">Where a chatbot keeps the conversation</li>
 </ul>
 <div class="cs-card cs-card-accent cs-at-5-6">The exampleOne help page: how to reset a password</div>
 Answer from that page, not from training
</div>

<div class="course-slide">
 One line per stage
 <h2 class="cs-title">RAG in 8 lines</h2>
 <div class="cs-code cs-no-copy">
 <div class="cs-code-bar">rag_chatbot.py · pseudocode</div>
 <pre class="cs-code-body cs-lines cs-focus-at-69-5 cs-focus-at-82 cs-focus-at-123">text = gather(sources)
chunks = split(text)
vectors = embed(chunks)
db.add(vectors)
q = embed(question)
hits = find_similar(q)
messages = build(question, hits)history, question, hits)
raw = llm(messages)</pre>
 </div>
 <div class="cs-band cs-band-note">
 Before anyone asks: lines 1–4
 Gather: any source becomes clean text
 Split: chunks of a few sentences
 Embed: a vector, a list of numbers
 Similar meaning, similar numbers
 Store: in the database, ready
 Every time someone asks: lines 5–8
 “How do I reset my password?”
 Same embed, so the numbers compare
 Find similar: the closest chunks
 Not the answer. The reading.
 Messages: question and chunks, one request
 The model returns raw text
 Your app turns it into the reply
 All eight, as one chain
 Each line hands its result on
 Retrieve, augment, generate
 Retrieve: lines 5–6
 Augment: line 7. Generate: line 8.
 The gap: no memory
 Every question starts over
 “And what if the email never arrives?”
 Which email? It never saw the first question
 Wrong turn: old chat into Gather
 Gather is your documents, before anyone asks
 Wrong turn: retrain the model
 Slow, and the chat changes every turn
 Memory goes in Messages
 history: the earlier turns
 The password question and its reply
 Now it knows: the reset email
 Only line 7 changed. The other 7 stay.
 </div>
</div>

<div class="course-slide cs-center">
 Leave with this
 A RAG chatbot is 8 stages
 <ul class="cs-checks">
 <li class="cs-at-4-1">Closest chunks are reading, not the answer</li>
 <li class="cs-at-6-7">The conversation lives in Messages</li>
 </ul>
</div>

<div class="course-slide cs-center">
 On this tip
 <h2 class="cs-title cs-at-0-4">Quiz, then practice</h2>
 Quiz: name each stage's job
 Practice: change Messages for a follow-up
 Until the next tip, happy building
</div>