Intermediate

Vector Databases — Intermediate

Production indexing, chunking, hybrid retrieval, multi-modal embeddings, and the evaluation discipline that keeps retrieval honest.

5 lessons ~231 min total
  1. 1

    Index Types: HNSW, IVF, and Quantization

    Pull back the curtain on vectorDB.search(). Understand how HNSW graphs, IVF clustering, and quantization actually work — and which combination to pick when your dataset grows from a million vectors to a billion.

    45 minHNSWIVFProduct Quantization
  2. 2

    Chunking Strategies That Don't Ruin Retrieval

    The single highest-leverage decision in RAG retrieval quality. Fixed-size vs token-aware, recursive splitting, overlap and the boundary problem, parent-document retrieval, and layout-aware chunking for markdown, HTML, PDF, code, and tables.

    46 minChunkingToken-Aware SplittingLayout-Aware
  3. 3

    Hybrid Search: Dense + Sparse

    Where dense retrieval fails, BM25 wins, and vice versa. The fundamentals of keyword search, Reciprocal Rank Fusion (the production standard for combining the two), when hybrid actually beats pure dense, and how Pinecone / Weaviate / Qdrant / Elasticsearch each implement it.

    47 minBM25Reciprocal Rank FusionHybrid Retrieval
  4. 4

    Multi-Modal Embeddings: CLIP and Beyond

    Text and images in the same vector space. How CLIP works (contrastive training on 400M internet pairs), when to reach for SigLIP / ImageBind / managed APIs, where it breaks (compositional reasoning, OCR), and the production patterns for unified multi-modal indexes.

    47 minCLIPSigLIPMulti-Modal Indexes
  5. 5

    Retrieval Evaluation: The Measurement Loop

    The measurement infrastructure that ties the whole course together. Build a gold set, pick the right metric (recall@K / MRR / NDCG) for your workload, navigate the offline-vs-online chasm, and run the full tuning loop in the right cost order. Intermediate capstone.

    46 minGold SetsRecall@KTuning Loop