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 reranking in a RAG pipeline?

Reranking is a second stage that takes the candidates a fast retriever returned, rescores each one against the query with a slower and more accurate model, and keeps only the best few for the prompt. It is often the single highest-return change to a RAG stack because it is a few lines of code, needs no re-indexing, and directly fixes the common case where the right chunk was retrieved but ranked too low to be used.

What is a cross-encoder in retrieval?

A cross-encoder joins the query and one document into a single input and runs them through the model together, so attention flows across both texts and it outputs one relevance score. This joint reading makes it more accurate than a bi-encoder, which embeds the query and each document separately and compares two summaries that never saw each other, but it is also slower because the score cannot be precomputed.

Should you use a hosted reranker like Cohere or self-host one like BGE?

A hosted reranker like Cohere Rerank is a network call with a per-query fee and no serving to run, while a self-hosted open cross-encoder like a BGE reranker is free to download and keeps your data on your own infrastructure but requires a GPU and the usual work of running a model. A common progression is to start hosted to prove the gain, then move to self-host once volume grows to remove the per-query fee.

How much latency and cost does reranking add?

Reranking adds one model pass over the candidate set per query, which can be tens of milliseconds to a few hundred depending on the model, candidate count, and hardware, plus either a per-call fee or the compute to self-host. It is often small next to the time the model spends generating the answer, and you can tune it by reranking fewer candidates, since latency rises linearly with the candidate count while quality flattens after a knee.

What kind of RAG failure does reranking not fix?

Reranking cannot fix a recall problem, because it can only reorder the candidates the first stage returned and has nothing to promote if the answer chunk was never retrieved. Before adding one, pull the raw candidate set for a few failing queries and read it: if the right chunk is there but ranked low, reranking is the fix, and if it is missing entirely, the fix lives upstream in chunking, the embedding model, or the query.

What is MMR and when should you use diversity reranking?

Maximal Marginal Relevance, or MMR, reranks so that each chosen chunk is both relevant to the query and different from what has already been selected, controlled by a lambda knob that trades relevance for variety. It is a second, optional pass on top of the relevance rerank, most useful when your retrieved set is full of near-duplicates so the top results would otherwise be several copies of the same fact.

Why is a cross-encoder more accurate at judging relevance than the bi-encoder used for vector search?

It reads the query and the document together in one pass, so attention runs across both, instead of comparing two separately made vectors, A bi-encoder embeds the query and each document separately and compares the finished vectors, so it never sees them together and captures topic overlap more than true relevance. A cross-encoder joins the query and one document into a single input and runs them through the model together, letting attention span both texts. That joint reading is what makes it accurate, and because the score depends on the query it cannot be precomputed, which is why it is also slow.

In a retrieve-then-rerank pipeline, why does the first stage return around 50 candidates instead of just the 5 you will finally use?

To give the reranker a wide, high-recall set to reorder, since it can only promote chunks that were actually retrieved, The first stage optimizes for recall: cast a wide net so the right chunk is almost certainly inside the candidate set, even if it is ranked low. The reranker can only reorder what it is given, so if the right chunk is not in the 50, no reranking recovers it. Pulling more candidates and then keeping fewer, pass more and keep fewer, is what lets the accurate-but-slow reranker do its job on a set small enough to afford.

Your RAG answers are weak. You pull the raw top 50 for the failing queries and the correct chunk is not in the list at all. What should you do?

Fix the first stage: revisit chunking, the embedding model, or the query, because reranking cannot promote a chunk that was never retrieved, This is a recall problem, not a ranking problem. A reranker only reorders the candidates the first stage returned, so if the right chunk is not in the 50 it has nothing to promote. The fix lives upstream in retrieval: chunk size and boundaries, the embedding model, or the query. Reading the raw candidates is the cheap check that tells you whether you have a ranking failure (reranking helps) or a recall failure (it does not).

Which metric does adding a reranker primarily improve, and which stays roughly the same?

It improves precision@5 while recall@50 stays roughly the same, because the reranker reorders the retrieved set rather than adding to it, Recall@50 is decided when the first stage runs, and a reranker adds nothing to the candidate set, so recall stays roughly flat. Precision@5 is what the reranker moves, because it pulls the genuinely relevant chunks into the top few you actually send to the model. High recall with low precision is a reranking problem; low recall is a retrieval problem that lives upstream, which is why measuring both tells you which stage to tune.

Which statement about the cost of reranking is accurate?

It adds one model pass over the candidate set per query, so latency rises with the candidate count, though the added time is often small next to generation, Reranking sits on top of your existing retriever with no re-indexing, but it is not free. It adds one cross-encoder pass over the candidates for every query, and because it does one forward pass per candidate, latency rises with the candidate count while quality flattens past a knee. The exact latency depends on the model, candidate count, and hardware, and it is often small compared with the model's generation time, which is part of why the trade is usually worth it.

Reranking: The Highest-Return Change in Your RAG Stack

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