Do I need a vector database?

Not for this scale. The sprint uses an in-memory index so you can inspect every stored vector, and the final session explains what a hosted vector database adds once you outgrow that, which is later than most tutorials imply.

Which embedding model?

A small open one that runs locally, so there is no API cost and no rate limit while you are iterating on chunking. Swapping in a hosted embedding model is a one-line change and the sprint has you measure whether it actually helps your data.

Why so much emphasis on refusal?

Because a RAG system that always answers is worse than no system: it produces confident, well-cited, wrong answers about the gaps in your documents. Measuring refusal is the only way to know whether yours is trustworthy, and almost no tutorial does it.

Build sprints₹499

RAG that refuses to guess

Take this if you want an assistant that answers from your own documents and can be trusted when the answer is not in them.

What you’ll be able to do

You can chunk and embed a document set, retrieve relevant passages, ground an answer in them with citations, and measure your system's refusal behaviour on questions it should not be able to answer.

What it assumes

Confident Python and some familiarity with embeddings — the free recommendations course covers enough.

Price: ₹499
Length: 3 weeks
Sessions: 12
Effort: ~6 h/week · 18 h total
For: College · Working professional

Adds it to My courses on your dashboard so you can find it again. Payment is not open yet — nothing is charged.

What’s taught, in order

01
Chunking is the whole ballgame
Splitting documents four different ways and seeing retrieval quality change more than any later choice does. Most disappointing RAG systems are lost here, before any model is involved.
02
Embedding and retrieving
Building the index, running nearest-neighbour search, and reading the retrieved passages yourself before any generation happens. If the right passage is not retrieved, no prompt can save the answer.
03
Grounding, and citing
Constructing the prompt so the answer must quote its source, and making the citation checkable. An answer you cannot trace is not a grounded answer, whatever the model claims.
04
Teaching it to refuse
Writing the negative test set — questions whose answers are genuinely absent — and tuning until the system declines instead of confabulating. This is the module the sprint is named after.

What you build

Build RAG over a document set of your own, such as your course notes or a set of public reports. Ship an evaluation set of at least twenty questions, of which at least six are unanswerable from the documents, and report retrieval accuracy, citation correctness and refusal rate on the unanswerable ones.

Check it yourself against this

Your document set is real and at least twenty pages in total.
Twenty or more evaluation questions, with six or more deliberately unanswerable.
Retrieval accuracy reported separately from answer quality.
Every generated answer carries a citation you verified by hand on a sample.
Refusal rate on unanswerable questions is reported honestly, whatever it is.
You can name the chunking choice that most affected your results.

Then answer this

Your system retrieves the right passage and still answers wrongly — what are the two distinct failures that could produce that, and how would you tell them apart?

Two or three sentences, in your own words. If you cannot, the course is not finished — go back to the module it came from.

What this course does not do

A single-document-set, single-turn system with no reranking, no hybrid keyword search, no conversation memory and no production concerns. It is enough to build something genuinely useful for yourself and to evaluate it honestly, and not enough to serve other people.

Where this leads

This sprint is three weeks of the fourteen-week Generative AI phase, which surrounds it with LLM internals, alignment, multimodal models and the inference optimisation that makes a retrieval system fast enough to use.

See the Generative AI phase Try a real lesson free

FAQ

Do I need a vector database?: Not for this scale. The sprint uses an in-memory index so you can inspect every stored vector, and the final session explains what a hosted vector database adds once you outgrow that, which is later than most tutorials imply.
Which embedding model?: A small open one that runs locally, so there is no API cost and no rate limit while you are iterating on chunking. Swapping in a hosted embedding model is a one-line change and the sprint has you measure whether it actually helps your data.
Why so much emphasis on refusal?: Because a RAG system that always answers is worse than no system: it produces confident, well-cited, wrong answers about the gaps in your documents. Measuring refusal is the only way to know whether yours is trustworthy, and almost no tutorial does it.

Other courses

This is a standalone course. It does not add weeks to the 78-week curriculum or change your roadmap. If you want the whole path instead, the full programme is ₹399/month · ₹2,000/year — see what it covers.