Free pilot · Limited seats · 30 days
Understand retrieval by building it from scratch
You work with AI systems every day. But search, the retrieval layer that everything depends on, is still a black box. IndexZero fixes that. You will build a working search engine from the ground up: tokenizer, index, BM25 ranker, evaluation harness, vector search, and a FastAPI endpoint. Every piece written by you, in Python. Purpose-built for engineers who want real understanding, not configuration recipes.
What you'll build
One codebase. Every layer of search.
You will know how BM25 actually works
Not just that it scores documents by relevance. You will have implemented it from the formula: IDF weighting, document length normalization, and term frequency saturation. When your Elasticsearch or Algolia rankings behave unexpectedly, you will know why.
You will understand retrieval evaluation
Most engineers cannot tell you whether their search is getting better or worse. You will build a proper evaluation harness with nDCG, MRR, and precision@k. Then you can measure whether a change actually improves results, rather than eyeballing a few queries.
You will see RAG retrieval failures clearly
When your RAG pipeline returns wrong answers, you will know whether the problem is chunking, embedding, top-k threshold, or prompt. This course makes every retrieval stage observable, so debugging becomes a process instead of a guess.
You will understand vector search deeply
Not just embed and cosine similarity. You will build the embedding pipeline and implement cosine similarity scoring yourself. Then you will combine lexical and semantic retrieval in a hybrid pipeline and understand exactly what reciprocal rank fusion does.
You will have a working FastAPI search endpoint
The capstone is a complete search API running your full retrieval pipeline over HTTP. It is an artifact that proves you understand the entire stack and is useful to show in any technical discussion.
You will think about search differently afterward
The shift is hard to describe but easy to recognize. You stop thinking of search as a service to configure and start thinking of it as a system to design. That is the thing you cannot unlearn.
The course
Three parts. Each one changes the system.
Language into structure
M0–M1 — 5–8 hours
Start with raw text. Build a tokenizer that handles lowercasing, punctuation, stop words, and stemming. Every normalization choice is explicit in code you write. M0 begins with a ranking audit on a real movie corpus. By the end of M1, you have a working tokenizer and vocabulary builder.
Structure into ranking
M2–M4 — 12–16 hours
Build the inverted index that makes search fast. Implement BM25 from the math. Wire up an evaluation harness with labeled queries. By the end of Part 2, you have a working search engine you can measure with nDCG, MRR, and precision@k.
Ranking into production
M5–M9 — 15–20 hours
Add structured queries (AND/OR/NOT, phrase, proximity), vector search, and hybrid retrieval. Build an incremental index that handles updates without full rebuilds. Wire the whole thing into a FastAPI endpoint. The capstone is a complete search API running on your own code.
M0 and M1 are always free. No application needed. M2–M9 are included in the 30-day pilot access. Total: 20–30 hours of focused work.
Who this is for
You should take this course if
You build RAG pipelines
Retrieval is the most failure-prone part of RAG. When results are bad, you need to know whether to fix the chunking, the embedding, the ranker, or the prompt. After this course, you will be able to isolate each component and fix it systematically.
You are adding search to a product
Backend or full-stack engineers building search into an app. You have read the docs, but you do not have an internal model for what BM25 is doing when it ranks your results. This course builds that model in code you wrote yourself.
You are a senior engineer who wants the real understanding
You have used hosted search for years. You know how to configure it well. But you have never built the pieces from scratch, so the internals are still opaque. After this, they are not.
You want a portfolio piece that shows real depth
A FastAPI search endpoint backed by a search engine you built from scratch is a different signal than configuring Elasticsearch. This course gives you something to show that demonstrates understanding, not just familiarity.
Free pilot access
30 days free. 20–50 seats.
Full course access, M0 through M9, free for 30 days. No credit card, no commitment. In exchange, we ask for honest feedback after 4 weeks: what broke, what was confusing, what you would change. We are building this course iteratively and pilot students are part of that process.
We are selective about who gets in. Not because we want to exclude, but because we want people who will actually engage with it and give us real feedback. If you are the kind of person who wants to really understand retrieval, apply.
About
Built by Sumit Garg
Sumit spent years building search infrastructure at Microsoft, working on the systems behind Azure AI Search. He built this course because most engineers use search without understanding what it does underneath. IndexZero makes retrieval visible in code you write yourself. So you understand it the way you understand systems you have built, not systems you have only configured.