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 query rewriting in RAG?

Query rewriting is a model call that turns the user's raw input into a self-contained, retrieval-friendly query before it is embedded. It resolves pronouns and follow-up references against the chat history, so a message like 'and how long does that take?' becomes 'how long does a refund take for an annual plan?'. Without it, conversational RAG breaks at the second turn, because the vector store never sees the chat history the fragment depends on.

What is HyDE and why does it work?

HyDE, Hypothetical Document Embeddings, generates a fake answer to the query and searches with the embedding of that answer instead of the question. It works because an answer, even a factually wrong one, is written in the same declarative register as the real answer chunk, so its embedding lands closer to the target than the question's embedding does. The hypothetical text is only a probe; the user only ever sees content from the real retrieved chunks. The method comes from Gao et al. 2022, arXiv 2212.10496.

What is multi-query retrieval and when does it help?

Multi-query generates several rephrasings of the question, retrieves top-k for each in parallel, and unions the deduplicated results. It raises recall because each phrasing reaches a different embedding neighborhood, so a chunk missed by one wording gets caught by another. It helps most when a corpus contains the same information in registers that differ from user language. The cost is N retrieval calls plus a mandatory reranking pass, because the union is large and noisy.

When should a query not go to the vector store at all?

When the answer must be computed rather than retrieved, or lives elsewhere. Aggregate questions like 'how many refunds in March' need a SQL tool, since vector search cannot count. Version- or tenant-scoped questions should hit the index through a metadata filter. Chitchat and acknowledgements need a direct reply with no retrieval. A router classifies the query and picks the destination before any retrieval spend.

How much latency do query transforms add, and how do I control it?

Every transform is a serial model call before retrieval: a small-model rewrite typically costs low hundreds of milliseconds, and HyDE costs a full generation, often half a second or more. Control it by gating: rewrite only when chat history exists or a standalone check fails, escalate to multi-query or HyDE only after plain retrieval comes back weak, and cache transform outputs for repeated queries. The gate itself must be nearly free, or it becomes the new cost.

A user's third message in a chat is 'and how long does that take?'. Retrieval returns generic chunks about processing times for the wrong product. What is the correct fix?

Rewrite the message into a standalone query using the chat history before embedding it, The referent of 'that' lives in the chat history, and the vector store never sees chat history, so the embedded fragment is anchored to nothing. No k, embedding model, or reranker can recover a topic that is absent from the query text. A standalone-query rewrite resolves the reference first, which is why conversational RAG systems run this transform on every follow-up turn.

Why can a factually wrong hypothetical answer still make a good HyDE search probe?

Because retrieval only uses the probe's position in embedding space, and wrong-but-plausible answer text still shares register and vocabulary with the true chunk, The hypothetical document is a probe, not an answer: the user never sees it and the generator never reads it. Its job is purely geometric, to sit closer to the answer chunk than the question does, and answer-shaped text about the right topic achieves that even when its specifics are invented. Averaging several samples dampens the influence of any one invention but does not make the text true, and nothing checks it, which is also why entity-level hallucination in the probe is HyDE's real failure mode.

Multi-query with 4 variants and top-5 each raised recall on your failure set. What cost did it necessarily introduce?

4x retrieval calls and a large noisy union that needs deduplication and reranking against the original question, The union of four top-5 lists is up to 20 chunks, most of them pulled in by the weaker variants, so multi-query converts a recall problem into a precision problem plus a fan-out cost. That is why the union is deduplicated by chunk id and then reranked against the original question, not a variant: the variants only widen the search, while the user's own phrasing stays the standard of relevance. The corpus and its embeddings are untouched.

A user asks 'how many support tickets mentioned refunds last month?'. Your RAG system retrieves ticket chunks and the model answers with a made-up number. What went wrong?

The routing layer failed: an aggregate question was sent to vector search, which retrieves text but cannot count, instead of a SQL or analytics tool, The answer to an aggregate question is not written in any chunk; it must be computed over rows. Vector search can only surface text that already exists, so the generator received a pile of individual tickets and improvised a total. The fix is upstream of retrieval entirely: a router that recognizes the aggregate shape and dispatches it to a SQL tool, which is the routing half of this lesson.

Query Transformation and Routing: Fixing Retrieval Before It Runs

Name: System Design Masterclass
Availability: InStock

in Retrieval Rag

This lesson is part of the ML Platform 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