Is this a video course?

No. This is an interactive, slide-based learning platform. Each lesson has rich text, animated diagrams, live code editors, and quizzes. You learn by reading, interacting, and doing, not by watching videos passively.

How long do I have access?

Forever. Both pricing tiers are one-time payments with lifetime access. This includes all current 766 lessons and any future content we add.

What level of experience do I need?

None. We start from absolute basics like 'What is latency?' and build up to distributed consensus protocols. The Foundation level assumes zero prior knowledge of system design.

How much does the system design course cost?

7.99 US dollars for lifetime access globally, or 499 Indian rupees for lifetime access in India. One-time payment, no subscription, no hidden fees. 11 lessons are free with no signup required.

What technologies are covered?

Everything from DNS and load balancers to Kubernetes, Kafka, distributed databases, consensus protocols, stream processing, security architecture, and observability. We cover principles and real-world implementations used at Netflix, Google, Amazon, Uber, Stripe, and more.

Is this useful for system design interview preparation?

Yes. The lessons are structured around the exact topics asked in system design interviews at FAANG and top-tier companies. Interactive diagrams help you practice whiteboard-style explanations. Covers everything from URL shortener design to distributed payment systems.

How is this different from ByteByteGo or Educative?

766 interactive lessons (4x more than most competitors), 16 different diagram types that build step by step, real production examples from Netflix, Google, Amazon, Uber, and Stripe, and lifetime access for a one-time payment of $7.99 instead of annual subscriptions costing 100 to 200 dollars per year.

What is hybrid retrieval and why use it?

Hybrid retrieval runs a dense vector search and a sparse keyword search at the same time and merges their results, so it covers both methods' blind spots. Dense search matches meaning but blurs exact rare tokens, keyword search matches exact terms but misses paraphrases, and fusing them catches queries that either one alone would drop.

What is the difference between dense and sparse retrieval?

Dense retrieval turns text into an embedding and finds documents whose meaning is close to the query, so it answers questions phrased in fresh words. Sparse retrieval, classic keyword search like BM25, scores a document by how many of the query's exact terms it contains, so it matches spelling rather than sense and catches exact identifiers that a model smooths away.

What is BM25 and how does it score documents?

BM25 scores a document by adding a weight for each query term it contains, set by three levers: term frequency with saturation so repeats matter less over time, inverse document frequency so rare terms count more, and length normalization so long documents do not win by size alone. A rare token like an error code or function name gets a large weight, which is exactly the signal an embedding blurs.

Why can't you just add a BM25 score and a cosine similarity together?

You cannot add the raw scores because they live on incomparable scales: a cosine similarity sits roughly between 0 and 1, while a BM25 score is unbounded and varies by corpus, so adding a 0.8 to a 22 is meaningless. Reciprocal Rank Fusion sidesteps this by throwing away raw scores and combining only rank positions, scoring each document by the sum of 1 divided by a constant plus its rank in each list.

When is hybrid retrieval worth running a second index?

Hybrid is worth the second index when exact tokens like error codes, SKUs, part numbers, or code identifiers matter in your corpus and users also paraphrase in their own words, which is common on technical and enterprise corpora. If your content is plain prose with few identifiers, dense-only is simpler and probably enough, so measure the recall gap on your own data before committing to running and syncing two indexes.

A user searches your RAG system for the exact error string 'ORA-01555' and the one runbook that names that error does not appear in the top results, even though your vector search ran fine. What is the most likely cause?

Dense embeddings capture meaning but blur exact rare tokens, so the literal error code is a weak signal for vector search, This is the dense blind spot. An embedding places an exact rare token near vaguely related text and loses the precise match. Error codes, SKUs, names, and code identifiers are exact strings, not concepts, so vector search ranks them poorly. Keyword search, which matches the literal token, is what catches them, which is the case for hybrid retrieval.

Why does Reciprocal Rank Fusion combine a BM25 list and a vector list without rescaling their scores?

It ignores raw scores and uses only each document's rank position, scoring by the sum of 1 / (k + rank) across lists, RRF discards the raw scores, which live on different and incomparable scales (bounded cosine versus unbounded BM25), and keeps only rank position. Each document scores the sum over lists of 1 / (k + rank), with k defaulting to 60. Because it looks only at ranks, no calibration between the two systems is needed, which is why it is the common default.

In BM25, why does a rare term like an error code or a function name get a high weight?

Because inverse document frequency gives a large weight to terms that appear in few documents, so the one document containing the rare exact term rises to the top, Inverse document frequency is the lever. A term in almost every document carries little weight; a term in only a handful carries a lot. An error code, SKU, or function name is rare across the corpus, so BM25 weights it heavily and surfaces the document that contains it. This is exactly the signal an embedding smooths away.

You add hybrid retrieval in front of your cross-encoder reranker. Why does the order matter?

A reranker can only reorder the candidates retrieval gave it, so hybrid widens recall to make sure both the paraphrase match and the exact match are in the candidate set before reranking, Retrieval controls recall, which documents enter the candidate set. Reranking controls precision, the ordering of that set. A reranker cannot promote a document that never made it into the candidates. Hybrid retrieval widens recall so both the exact and paraphrase matches are present, then the reranker orders them. Hybrid feeds rerank; it does not compete with it.

Hybrid Retrieval: When Keyword Search Beats Your Embeddings

Name: System Design Masterclass
Availability: InStock

in Retrieval Rag

This lesson is part of the AI Engineering course. Unlock the full course, a separate one-time payment, for lifetime access.

One-time payment. Lifetime access. No subscription.

Global

$20/lifetime

One-time payment via PayPal

Get Lifetime Access

5 Foundation lessons free, try before you buy

India Special

India

₹999/lifetime

One-time via UPI, cards, or net banking

Get Lifetime Access

5 Foundation lessons free, try before you buy

“The interactive diagrams alone are worth way more than the price.”

Anish K., Senior Engineer at a Bengaluru startup

Every ML Platform lesson, Foundation to Advanced

Real architecture diagrams with real tool logos

MLOps, LLMOps, and data engineering in depth

Quizzes with instant feedback

Built and reviewed by senior engineers

Lifetime access, all future content included

The first 5 Foundation lessons of this course are free, no payment needed. Try them now