All programsRAG — Retrieval-Augmented Generation

Make a model answer from your documents, and admit when it cannot.

Every organisation eventually wants the same thing: a model that answers from our documents, not from the internet. The demo takes an afternoon. The version you can put in front of a customer takes rather longer, because the failure is quiet — it retrieves the wrong passage and answers confidently anyway. Four weeks on the difference between the two, built on your own material.

Duration
4 weeks · around 16 hours
Mode
Live online — real sessions, not recordings
Timing
Evenings, designed around a full-time job
Cohort size
20–25
Material
Build on your own documents, or ours
Recordings
Every session recorded, yours to keep

Who this is for

  • Developers asked to build a chatbot over company documents
  • Teams whose first RAG attempt works in the demo and embarrasses them in production
  • Data and knowledge-management people sitting on a large document estate
  • Anyone who has seen a retrieval system answer fluently and wrongly

Complete beginners — this assumes you can write a script and call an API. AI Foundations covers RAG at an introductory level in week 3.

Prerequisites

  • Working Python, and comfort calling an HTTP API.

No machine-learning background is required. You will not train a model in this program; you will make one useful.

Modules

Module 1Week 1

Why the model needs your documents at all

What a context window is and where it runs out. Why fine-tuning is usually the wrong answer to “it does not know our products”. Embeddings as meaning-with-coordinates, and the honest limits of similarity search. A first end-to-end pipeline, running by the end of the session.

IN THE SESSION

Build a first end-to-end pipeline over a document set in the session — rough, working, and yours.

You leave with: A working retrieval pipeline over a document set, however rough.

Module 2Week 2

Chunking, which decides everything downstream

The least discussed and most consequential choice in the stack. Fixed windows, overlap, structure-aware splitting, and what each does to a table, a contract clause, a code file and a scanned PDF. Metadata that makes filtering possible later. We measure the damage rather than guess at it.

IN THE SESSION

Chunk the same documents four ways, run identical questions against each, and pick your strategy from the numbers rather than the advice.

You leave with: A chunking strategy chosen from evidence about your own documents.

Module 3Week 3

Retrieval that actually retrieves

Where pure vector search fails — names, numbers, exact phrases, acronyms — and why hybrid search with keyword matching fixes most of it. Reranking. Filtering by metadata. Multi-query and query rewriting for the questions users really type.

IN THE SESSION

Add keyword search and reranking alongside vectors, then measure how much more often the right passage reaches the model.

You leave with: A measurable jump in how often the right passage is in front of the model.

Module 4Week 4

Grounding, citation, refusal — and proving it works

Making the answer cite the passage it came from, and making it say “not in these documents” instead of inventing. Then evaluation: a test set of real questions, retrieval and answer scored separately, so you can tell which half is broken. Cost and latency per query, at your volume.

IN THE SESSION

Make it cite the passage it used and refuse when the answer is not in your documents — then score retrieval and answers separately to find which half is weak.

You leave with: A system that cites its sources, refuses when it should, and has a number attached.

What you walk away with

  • A working RAG system over documents you actually care about
  • An evaluation set, so “better” stops being a feeling
  • Answers that carry citations, and refusals where refusal is correct
  • A cost-per-query figure you can take to whoever pays for it
  • Codnov Certificate of Completion, with a unique ID and a public verification link

Sample projects

  • A policy and handbook assistant for an internal team
  • A support assistant grounded in real product documentation
  • A research assistant over a library of papers or reports, with citations
  • A contract and tender search tool that quotes the clause it found
The guarantee

You finish with a deployed, reviewed AI project, or we re-run you free.

Requires minimum 75% attendance and on-time milestone submission. Full conditions at codnov.ai/training/terms.

Taught by Codnov engineers who build and run AI systems in production every day.

Price

Cohort pricing is confirmed per intake — enquire for the current fee and dates. All prices exclusive of 18% GST.

Nothing is paid on this website. Register your interest and we will confirm your seat with you directly.

Register your interest

We will come back to you with the full syllabus, the schedule, and whether there is a seat in the next cohort.

No payment is taken on this website. Registering interest does not enrol you or commit you to anything.

Codnov Certificate of Completion. Not a degree or diploma recognised under the UGC Act, 1956 or the AICTE Act, 1987. All prices exclusive of 18% GST. Project completion guarantee subject to minimum 75% attendance and on-time milestone submission; full terms at codnov.ai/training/terms. Open to participants aged 16 and above.