Lab
Index bake-off
Same corpus, three indexes, recall and p95 on a labelled set.
Vector Database Deep Dive
Indexes, HNSW, filtering, hybrid search, and operations: the store under your RAG, not a logo on a slide.
Created by Baljeet Dogra
Course Objectives
- Choose an index because of recall, latency and filter needs — not a vendor page.
- Explain HNSW well enough to debug a bad neighbour set.
- Combine BM25 and vectors without cargo-cult fusion.
- Operate backups, reindexes and dimension changes without downtime theatre.
What you will produce
Ops
Runbook
Reindex, backup, and a dimension-change drill.
4 weeks curriculum
Expand a part for the syllabus. Content stays searchable when closed.
01 Embeddings and distance Week 1
What embeddings capture and what they do not. Metrics, dimensions, normalisation.
- Metrics
- Normalisation
- Models
- Failure modes
02 Indexes Week 2
Flat, IVF, HNSW. Parameters you can defend. Filtered search.
- HNSW
- IVF
- Filters
- Recall vs latency
03 Hybrid search Week 3
BM25, fusion, rerankers. When keywords win outright.
- BM25
- Fusion
- Rerank
- Metadata
04 Operate it Week 4
Capacity, reindex, backup, multi-tenancy. Capstone: bake-off plus runbook.
- Capacity
- Reindex
- Tenancy
- Capstone
Who this is for
RAG builders
You called similarity_search. You cannot explain the index.
Data engineers
You own stores. Vectors are a new access pattern, not a new religion.
Prerequisites
- Python
- You have built or used a RAG demo
- SQL or another datastore at a basic level
Not a fit if
- A survey of twenty vendor logos
- People who have never retrieved a document
Related: Hands-on LangChain · Applied GenAI Engineering
Questions
Which product?
One open store you can run locally and one managed option. The ideas transfer.
Do I need a GPU?
No. Embedding APIs or a small CPU model are enough.
Ready to start?
Four weeks to an index you can explain and operate. Create an account to enrol.
Enrol now