What is the difference between RAG and fine-tuning?

RAG retrieves relevant text at query time and puts it in the prompt, so the model answers from documents it was never trained on. Fine-tuning adjusts model weights to change style, format, or task behavior. Use RAG when the answer depends on facts that change or that the model has never seen; use fine-tuning when you need a consistent output format or domain tone. Most production systems use RAG first, and fine-tune later if at all.

Do I need to know machine learning to take this course?

No. The course is written for software and data engineers. You need solid Python and comfort with APIs. Embeddings and vector similarity are taught from scratch, and no model training is involved.

Which vector database does the course use?

Labs run on pgvector and FAISS because they are free and require no vendor account, and the retrieval concepts transfer directly. The course also covers how managed options such as Pinecone, Weaviate, and Azure AI Search differ on indexing, filtering, and cost, so you can choose deliberately.

How long is the course and how is it delivered?

Two days, roughly half instruction and half hands-on labs, delivered onsite, virtually, or hybrid. Labs run in a browser-based environment, so there is nothing to install on corporate laptops.

Can the course use our own documents?

Yes. For private cohorts we commonly swap the sample corpus for a representative set of your own documents so the chunking and retrieval problems your team hits are the ones you practice on.

BeginnerTechnical AI/ML

Introduction to RAG

Name: Introduction To RAG
Availability: InStock

Learn Retrieval-Augmented Generation from first principles: chunking, embeddings, vector search, and grounding LLM answers in your own documents.

Course Overview

Large language models are trained on public data and frozen at a point in time. They know nothing about your contracts, your ticket history, or the policy document your legal team revised last week, and when asked they will produce a confident answer anyway. Retrieval-Augmented Generation is the standard remedy: retrieve the relevant passages from your own corpus first, then ask the model to answer using only what was retrieved.

This course builds a working RAG pipeline end to end, one stage at a time. You start with document loading and chunking, because chunk boundaries determine what the retriever can find at all. You then move to embeddings and vector stores, learn why cosine similarity retrieves plausible-looking but wrong passages, and add the reranking and hybrid keyword search that fix it. Every stage is implemented in a lab rather than described on a slide.

The final third of the course is about knowing whether your pipeline actually works. Teams routinely ship RAG systems with no measure of retrieval quality and discover the problem only when a user is given a wrong answer with a citation attached. You will build a small evaluation set, measure retrieval precision and answer groundedness, and learn to trace a bad answer back to the specific stage that caused it.

Duration: 2 days|Delivery: onsite, virtual, or hybrid

Prerequisites

Python programming at an intermediate level (functions, classes, virtual environments)
Familiarity with calling a REST API and handling JSON
Having used an LLM chat interface such as ChatGPT or Claude
No prior machine learning or NLP experience required

Who Should Attend

Software engineers who need to ground an LLM feature in company documents
Data engineers responsible for the ingestion pipeline behind an AI product
Technical leads evaluating whether RAG or fine-tuning fits their use case
Solutions architects who must size and cost a retrieval system before committing

Course Outline

1Why LLMs hallucinate: context windows, training cutoffs, and parametric memory
2RAG architecture end to end: ingest, chunk, embed, index, retrieve, generate
3Document loading and parsing: PDFs, HTML, Office formats, and their failure modes
4Chunking strategies: fixed-size, recursive, semantic, and document-structure-aware
5Embeddings explained: what a vector represents and how models differ
6Vector databases: indexing, distance metrics, and choosing between pgvector, FAISS, and managed services
7Retrieval quality: top-k tuning, metadata filtering, and why pure similarity search fails
8Hybrid search and reranking: combining BM25 keyword search with dense retrieval
9Prompt construction for grounded answers, citations, and refusal behavior
10Evaluating RAG: retrieval precision and recall, answer groundedness, and building a golden test set

Learning Outcomes

Explain when RAG is the right approach and when fine-tuning or a longer context window is better
Build a complete RAG pipeline from raw documents to a cited, grounded answer
Choose a chunking strategy based on document structure rather than by default
Select and configure a vector store appropriate to your data volume and latency budget
Diagnose whether a wrong answer came from retrieval or from generation
Measure retrieval and answer quality with a repeatable evaluation set
Add citations and controlled refusal so the system says it does not know

What You Will Build

A working RAG pipeline over your own sample documents
A comparison of at least three chunking strategies measured on the same queries
An evaluation notebook reporting retrieval precision and answer groundedness
A reference architecture diagram you can take to your own team

Frequently Asked Questions

What is the difference between RAG and fine-tuning?: RAG retrieves relevant text at query time and puts it in the prompt, so the model answers from documents it was never trained on. Fine-tuning adjusts model weights to change style, format, or task behavior. Use RAG when the answer depends on facts that change or that the model has never seen; use fine-tuning when you need a consistent output format or domain tone. Most production systems use RAG first, and fine-tune later if at all.
Do I need to know machine learning to take this course?: No. The course is written for software and data engineers. You need solid Python and comfort with APIs. Embeddings and vector similarity are taught from scratch, and no model training is involved.
Which vector database does the course use?: Labs run on pgvector and FAISS because they are free and require no vendor account, and the retrieval concepts transfer directly. The course also covers how managed options such as Pinecone, Weaviate, and Azure AI Search differ on indexing, filtering, and cost, so you can choose deliberately.
How long is the course and how is it delivered?: Two days, roughly half instruction and half hands-on labs, delivered onsite, virtually, or hybrid. Labs run in a browser-based environment, so there is nothing to install on corporate laptops.
Can the course use our own documents?: Yes. For private cohorts we commonly swap the sample corpus for a representative set of your own documents so the chunking and retrieval problems your team hits are the ones you practice on.

Related Courses

Intermediate

Ready to Get Started?

Request Training Book a Brief Call with an Expert Browse All Courses