Model Online / On-Site
Duration
Modules 6 modules

Course Objective & Scope

The ceiling of any RAG or search system is set by its embedding and retrieval layer: a poorly matched embedding model will miss Turkish morphology, domain acronyms and product codes no matter how strong the downstream LLM is. This course covers bi-encoder architectures, commercial and open-weight embedding models (BGE-M3, E5, Qwen3-Embedding, Cohere, OpenAI, Voyage), Matryoshka Representation Learning (MRL) and binary/scalar quantization.

Participants benchmark multiple embedding models on their organisation's Turkish and multilingual data using Recall@k and NDCG@k, and fine-tune an embedding model for domain terminology.

Who Should Attend?

  • AI engineers building RAG, enterprise search, recommendation or deduplication systems
  • Data scientists working on text representations, clustering and semantic matching
  • Search and platform engineers optimising retrieval accuracy and index memory footprint
  • Prerequisite: Python, basic vector/linear algebra concepts and familiarity with NLP or LLMs

Key Takeaways

  • Distinguish bi-encoder, cross-encoder, learned sparse and ColBERT (late-interaction) representations
  • Benchmark embedding models on Turkish and domain-specific corpora using custom test sets
  • Cut vector memory footprint with Matryoshka (MRL) truncation and int8/binary quantization
  • Apply instruction prefixes and asymmetric query/document encoding correctly
  • Fine-tune an open embedding model on domain data using hard negative mining and contrastive loss

Curriculum

Vector Representations and Contrastive Learning Foundations
  • From lexical matching (BM25) to dense vector spaces and InfoNCE loss
  • Cosine similarity, dot product, L2 distance and vector normalisation
Embedding Model Families and Multilingual Performance
  • Comparing OpenAI, Cohere and Voyage with BGE-M3, multilingual-E5 and Qwen3-Embedding
  • Asymmetric search (short query vs long passage), task prefixes and Turkish morphology
Dense, Learned Sparse (SPLADE) and Late Interaction (ColBERT)
  • The single-vector bottleneck and where dense embeddings miss exact codes or terms
  • Multi-representation models (BGE-M3) and ColBERT token-level MaxSim scoring
Dimensionality Reduction and Vector Quantization
  • Flexible dimension truncation with Matryoshka Representation Learning (MRL)
  • Moving from Float32 to FP16, INT8 (scalar) and 1-bit (binary) vectors with oversampled rescoring
Domain Adaptation and Embedding Fine-Tuning
  • Generating synthetic (query, positive) pairs with LLMs and mining hard negatives
  • Adapting embeddings with Sentence Transformers without catastrophic forgetting
Lab: Enterprise Embedding Benchmark
  • Measuring Recall@k, MRR and NDCG@k on internal corporate documents
  • Selecting the optimal embedding setup across memory, latency and accuracy

Other Courses in This Category

  • Vector Database Architectures: pgvector, Milvus, FAISS, Weaviate and Pinecone

    Engineering deep dive into ANN indexing (HNSW, IVF, DiskANN, PQ) and vector databases (pgvector, Milvus, Weaviate, FAISS, Pinecone), covering filtered search, multi-tenancy and capacity planning.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View
  • Data Preparation, Document Parsing and Chunking for RAG

    Parsing complex PDFs, tables and scanned enterprise documents with Docling, Unstructured and VLMs, and applying structural, semantic, parent-child and contextual chunking for high-recall RAG.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View
  • Citations, Grounding and Hallucination Control in RAG Systems

    Engineering RAG systems for strict grounding, verifiable inline citations, conflict resolution across sources, calibrated abstention when evidence is missing and automated faithfulness verification.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View
  • Advanced RAG Advanced RAG

    Enterprise RAG architectures covering BM25-dense hybrid retrieval, ColBERT, neural reranking, self-correcting Agentic RAG, GraphRAG, Text-to-SQL, and GPU-accelerated serving.

    RAG & Enterprise Knowledge Systems Corporate, Individual
    Live Course
    10 modules
    View
  • Hybrid Search and Reranking for High-Precision RAG

    Boosting RAG retrieval precision and recall by combining BM25/SPLADE lexical search with dense vectors via RRF, query rewriting (multi-query, HyDE) and cross-encoder or ColBERT reranking.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View
  • Metadata Architecture and Permission-Aware Retrieval in RAG

    Designing metadata schemas, dynamic filter extraction and RBAC/ABAC/ReBAC security-trimmed retrieval so enterprise RAG assistants never leak unauthorized documents.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View

Get in touch