What would I need to add to turn this into an actual deployed chatbot?

A persistent vector database (like Pinecone, Chroma, or pgvector) instead of an in-memory list, a real document ingestion pipeline for your actual files, and a UI layer calling this same retrieve-then-generate pattern per user message. The core logic you built in this masterclass doesn't change — only the storage and interface layers around it.

Could I swap OpenAI for another provider in this pipeline?

Yes — the chunking and retrieval logic is provider-agnostic Python. Only the embeddings call and the generation call are provider-specific, and both would just need their request/response shapes adapted to a different API.

HTML MASTER CLASS /// LEARN TAGS /// BUILD STRUCTURE /// SEMANTIC WEB /// HTML MASTER CLASS /// LEARN TAGS ///

Your Complete RAG Chatbot

Add a citation requirement to close out a complete, working RAG pipeline, and verify the final answer is both correct and traceable to its source.

Narrated Video Summary

data-composition-id="ragchatbotmasterclass-module3_lesson9"1280×720 @ 30fps3 clips0:51 total

The Last Piece: Citations

Back in Module 2 you labeled every retrieved chunk with a source number specifically so the model could cite it. This lesson finally uses that: one more instruction, and every grounded answer names exactly which source it came from — the mark of a genuinely production-ready RAG answer.

system += "Always end your answer with the source citation, like [Source 1]."

Masterclass Complete

You built a real RAG chatbot: chunking, embeddings, a vector store, retrieval, grounded generation, graceful refusal, and citations — every piece implemented and run for real, not just watched. The exact hallucination from Lesson 1 is fixed. This is the actual architecture behind production RAG systems.

/* RAG Chatbot: Complete */
.pipeline { status: 'production-ready'; }

0:00 / 0:51

Scene 1 / 3 — The Last Piece: Citations

⚡ Total XP: 0|💻 ragchatbotmasterclass XP: 0

Skill Matrix

UNLOCK NODES BY LEARNING NEW TAGS.

Complete RAG Pipeline

Grounded, refused, and cited.

Quick Quiz //

What does adding citations to a grounded answer actually verify?

🚀 LEVEL UP TO SENIOR:Unlock 500+ Advanced Practical Challenges & Exercises.

🎓 COURSERA PARTNER:Earn professional Google, Meta, and IBM certificates to supercharge your resume.

The full chain, assembled: chunking, embeddings, retrieval, grounded generation, refusal handling, and citations — all built and run for real.

1Why Citations Close the Loop

Every piece before this lesson made the answer correct. Citations make it verifiable — a user (or a developer debugging a wrong answer) can check exactly which source backs a claim, rather than trusting the model's word for it. It's the difference between 'probably grounded' and 'demonstrably grounded', and it only cost one more line in the system prompt because Module 2 already labeled every source.

2The Full Pipeline You Built

Chunk real documents (Module 1) → embed chunks and queries with a real API (Module 1) → search and format the most relevant chunks (Module 2) → generate a strictly grounded, cited answer (Module 3). Every one of those four stages is a real function or a real API call you wrote and ran yourself — not a diagram you watched get explained.

3Step-by-Step Breakdown

The Last Piece: Citations. Back in Module 2 you labeled every retrieved chunk with a source number specifically so the model could cite it. This lesson finally uses that: one more instruction, and every grounded answer names exactly which source it came from — the mark of a genuinely production-ready RAG answer.

Your Complete RAG Chatbot. This is the full chain: strict grounding, real retrieved context with numbered sources, and now a citation requirement. Run it and confirm the answer both states the correct policy AND ends with the exact source that backs it — a fully traceable, grounded, cited answer.

What does this masterclass's complete pipeline actually consist of, end to end?

→Chunk documents, embed the chunks and the query, retrieve and format the most relevant chunks, then generate a grounded, cited answer using a strict system prompt.
→A single API call that automatically knows your private company data.

Masterclass Complete. You built a real RAG chatbot: chunking, embeddings, a vector store, retrieval, grounded generation, graceful refusal, and citations — every piece implemented and run for real, not just watched. The exact hallucination from Lesson 1 is fixed. This is the actual architecture behind production RAG systems.

Level Up 🚀

Advanced cheat sheets, SEO tricks, and interview prep for this topic.

Browser Support

ChromeSupported

Fully supported.

FirefoxSupported

Fully supported.

SafariSupported

Fully supported.

EdgeSupported

Fully supported.

Accessibility (A11y)

1Make Citations Navigable, Not Just Visible

A production RAG UI should let users click or focus a citation like [Source 1] to jump to or preview the actual source document, not just display it as static text.

<a href="#source-1" aria-label="View Source 1: PTO Policy">[Source 1]</a>

SEO Implications

1
Target 'build a RAG chatbot from scratch' as the primary search for this masterclass
This is the exact high-intent phrase developers search for once they've decided to build rather than just read about RAG.

Best Practices

Ship Citations Alongside Every Grounded Answer

A cited answer costs almost nothing extra to generate once sources are labeled, and it converts an opaque 'trust the model' UX into a verifiable one — always include citations in production RAG systems.

Frequent Bugs

THE BUG

Adding a citation instruction but not testing that it survives alongside the refusal instruction for out-of-scope questions.

THE FIX

Test citation behavior across both the answerable and unanswerable cases — a citation requirement can sometimes cause a model to fabricate a source number even when refusing, which needs its own explicit instruction to prevent.

Real-World Examples

Production HR Chatbot, Complete

The Nexora HR chatbot now answers 'You can roll over up to 5 PTO days [Source 1]' — correct, grounded, and traceable back to the exact handbook section, the complete production pattern this masterclass built piece by piece.

"You can roll over up to 5 unused PTO days. [Source 1]"

Interview Prep

?Frequently Asked Questions

Pascual Vila

Frontend Instructor // Code Syllabus

Common Pitfalls & Errors

The Error //

Hardcoding sensitive credentials

// Wrong
const API_KEY = 'sk-123456789';

// Correct
const API_KEY = process.env.API_KEY;

The Solution //

Never hardcode API keys, passwords, or secrets in your source code. Use environment variables (.env files) to keep them secure and out of version control.