Star on GitHub
DocsRoadmap

Vector DB learning path

A sequenced curriculum covering every topic on this site — 25 lessons grouped into 8 stages, from first embedding to production RAG.

How to use this page
Work top-to-bottom. Each stage builds on the previous one. Skip a stage only if you already ship the things it describes — the later stages assume the earlier ones as background.

1. Foundations

Beginner2 lessons

Understand what a vector database is and run your first query.

  1. Step 1Introduction10 min

    Mental model: why vectors beat keyword search for semantic data.

  2. Step 2Quickstart15 min

    Embed text, insert it, and run a similarity query end-to-end.

2. Core concepts

Beginner4 lessons

Learn the math and data structures that power vector search.

  1. Step 1Embeddings20 min

    How models map text/images into a dense vector space.

  2. Step 2Similarity search15 min

    Cosine vs. dot vs. L2 — when each distance metric is correct.

  3. Step 3Indexing (HNSW, IVF)30 min

    ANN structures that make billion-scale search fast.

  4. Step 4Hybrid search20 min

    Combine BM25 keyword and dense vectors with score fusion.

3. Pick a database

Intermediate7 lessons

Survey the ecosystem and choose the right engine for your workload.

  1. Step 1Pinecone15 min

    Managed serverless — fastest path to production.

  2. Step 2Weaviate15 min

    Schema-first with built-in modules for vectorization.

  3. Step 3Qdrant15 min

    Rust-based, strong filtering and payload indexing.

  4. Step 4Chroma10 min

    Embedded DB ideal for prototyping LLM apps locally.

  5. Step 5Milvus20 min

    Distributed, GPU-accelerated, designed for tens of billions of vectors.

  6. Step 6pgvector15 min

    Keep vectors next to relational data inside Postgres.

  7. Step 7FAISS20 min

    The reference library — every other engine borrows from it.

4. Internals & scale

Advanced3 lessons

Understand how production systems compress, distribute, and filter.

  1. Step 1Quantization (PQ / SQ / BQ)25 min

    Shrink vectors 4-64× with minimal recall loss.

  2. Step 2Sharding & replication20 min

    Scatter-gather, the 2k rule, semantic vs. random sharding.

  3. Step 3Metadata filtering15 min

    Integrated filters inside HNSW vs. naive post-filtering.

5. Build RAG

Intermediate2 lessons

Wire embeddings + retrieval + an LLM into a working RAG app.

  1. Step 1Build RAG30 min

    The canonical retrieve → augment → generate pipeline.

  2. Step 2Chunking strategies25 min

    Fixed, recursive, semantic, and structural chunking with code.

6. Advanced retrieval

Advanced4 lessons

Move beyond naive RAG with reranking, query rewriting, and agents.

  1. Step 1Advanced RAG (HyDE, RRF)25 min

    Hypothetical document embeddings and Reciprocal Rank Fusion.

  2. Step 2Reranking25 min

    Cross-encoders, ColBERT late interaction, MMR diversity.

  3. Step 3Agentic RAG30 min

    CRAG and Self-RAG — let the model decide when to retrieve.

  4. Step 4Multimodal RAG25 min

    CLIP shared spaces and ColPali for OCR-free PDF retrieval.

7. Production

Expert2 lessons

Measure quality, optimize cost, and run reliably at scale.

  1. Step 1Evaluation (RAGAS)25 min

    Faithfulness, answer relevance, recall@k, MRR, nDCG.

  2. Step 2Scaling & performance30 min

    Latency budgets, caching, batching, and capacity planning.

8. Bonus — how LLMs work

Advanced1 lesson

Open the black box so you understand what RAG is feeding.

  1. Step 1nanoGPT explained45 min

    Transformer internals, KV-cache, and decoding from scratch.

What you'll be able to build

  • Semantic search over your own documents, code, or images.
  • A production-grade RAG chatbot with reranking and evaluation.
  • Hybrid search combining BM25 keyword + dense vectors.
  • Multimodal retrieval over PDFs, screenshots, and figures.
  • An agentic system that decides when to retrieve and when to answer.