Presentations in This Series
- Embedding Models →Dense vs sparse; bi-encoder vs cross-encoder; modern families (BGE, GTE, E5, Voyage, OpenAI text-embedding-3, Cohere, Nomic, Jina); ColBERT late interaction; Matryoshka; domain adaptation.
- Vector Databases →pgvector, Qdrant, Weaviate, LanceDB, Pinecone, Milvus, Chroma; HNSW, IVF, PQ; recall vs latency; updates and durability; decision matrix.
- Hybrid Search & Reranking →BM25 mechanics; RRF and weighted fusion; cross-encoder rerankers (BGE-Reranker, Cohere Rerank, mxbai); ColBERT-as-reranker (PLAID); cascades; latency budgets.
- Chunking & Ingestion →Strategies (fixed, semantic, structural, contextual); recursive splitting failure modes; layout-aware (Unstructured, MinerU, LlamaParse, Docling); ColPali for vision; deduplication.
- Agentic RAG Patterns →HyDE, query decomposition, Self-RAG, Corrective RAG (CRAG), iterative retrieval, query routing, tool-using RAG.
- GraphRAG & Knowledge Graphs →Microsoft GraphRAG, KG construction, entity resolution; Neo4j, Memgraph, KuzuDB; hybrid graph+vector retrieval; Cypher in agent loops.
- Production RAG →RAGAS, ARES, Trulens; drift & freshness; latency budgets; multi-tenancy and ACLs; citation & grounding; cost; failure modes.
- Two-Step Retrieval — Architecture & Cost-Quality Mathematics →Companion deep-dive to deck 03. The bi-encoder / cross-encoder asymmetry; the indexability barrier (why MIPS scales and cross-attention doesn't); ANN math (HNSW, IVF, PQ recall–latency curves); the recall-ceiling theorem; RRF derivation; latency composition under tails; the (k, quality, latency) Pareto frontier; Platt / isotonic / temperature calibration; worked example at 10 M docs / 200 ms.
- Reranker Mathematics →Stage-2 maths in depth. Cross-encoder attention & token interactions; ColBERT MaxSim formal definition; learning-to-rank families (pointwise / pairwise / listwise); RankNet sigmoid loss and gradient; LambdaRank weighted by
|ΔNDCG|; LambdaMART GBDT; InfoNCE contrastive loss with hard negatives; MarginMSE cross-encoder → bi-encoder distillation; DCG / NDCG / MRR / MAP derivations.