Model Online / On-Site
Duration
Modules 6 modules

Course Objective & Scope

RAG systems built on dense vector search alone frequently miss exact SKU codes, error IDs, proper nouns and regulation clauses, while lexical search alone misses paraphrases and conceptual queries. This course covers multi-stage retrieval: combining BM25/SPLADE and dense vectors via Reciprocal Rank Fusion (RRF) or score normalisation, query transformation techniques, and second-stage reranking with cross-encoders and ColBERT.

Starting from Turkish morphological analyzers, participants build a two-stage hybrid retrieval + reranking pipeline and measure gains in NDCG@10 and MRR against latency overhead on internal data.

Who Should Attend?

  • AI engineers fixing retrieval misses in existing RAG or enterprise search systems
  • Software developers integrating search engines (Elasticsearch/OpenSearch) and vector databases
  • Data scientists working on technical support, legal/policy or catalogue search
  • Prerequisite: working knowledge of basic RAG, embeddings and vector search

Key Takeaways

  • Configure hybrid search combining Turkish-aware BM25/sparse retrieval with dense vector search
  • Tune Reciprocal Rank Fusion (RRF) and weighted score normalisation across retrieval channels
  • Apply query rewriting, multi-query expansion and HyDE where they improve recall
  • Rerank top-50 to top-100 candidates accurately using BGE-Reranker, Cohere Rerank or ColBERT
  • Balance NDCG@5 improvements against reranker latency and compute cost

Curriculum

Complementarity of Lexical (BM25/SPLADE) and Semantic Search
  • Why dense vectors struggle with exact SKUs, acronyms and proper nouns
  • BM25 parameters (`k1`, `b`), Turkish morphological analysis/stemming and learned sparse (SPLADE)
Hybrid Search and Score Fusion Techniques
  • Reciprocal Rank Fusion (RRF) and Distribution-Based Score Fusion (DBSF)
  • Hybrid query execution in Elasticsearch/OpenSearch, Weaviate, Qdrant, Milvus and pgvector
Query Transformation and Expansion
  • Conversational query decontextualisation, synonym expansion and sub-query decomposition
  • When HyDE (Hypothetical Document Embeddings) and step-back prompting help versus add latency
Second-Stage Reranking: Cross-Encoders and Late Interaction
  • Bi-encoders vs cross-encoders: full cross-attention across query and document tokens
  • Comparing BGE-Reranker, Cohere Rerank, Jina Reranker and FlashRank/ColBERT
Reranking in Production: Latency, Cut-offs and Diversity
  • Choosing candidate depth (`top_k=50` to `top_n=5`), score thresholds and MMR diversification
  • Batched inference, ONNX/TensorRT acceleration and handling long chunks in rerankers
Lab: From Single-Stage Vector Search to Multi-Stage Hybrid Retrieval
  • Benchmarking dense-only, BM25-only, hybrid and hybrid+rerank on internal queries
  • Producing a Recall@5, NDCG@5 and p95 latency comparison report

Other Courses in This Category

  • 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
  • Agentic RAG: Multi-Step, Corrective and Adaptive Retrieval Architectures

    Building Agentic RAG systems with LangGraph that plan sub-queries, route across heterogeneous data sources, grade retrieved evidence, self-correct queries (CRAG/Self-RAG) and bound latency.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View
  • GraphRAG and Knowledge-Graph-Enhanced Retrieval Systems

    Combining knowledge graphs (Neo4j, property graphs), entity-relationship extraction, community summaries (Microsoft GraphRAG) and vector search to answer multi-hop and corpus-wide enterprise queries.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View
  • RAG Evaluation and Quality Assurance: Golden Datasets, LLM-as-a-Judge and CI/CD

    Replacing ”vibe checks” in RAG with systematic evaluation: golden dataset curation, synthetic test generation, RAGAS/DeepEval/Promptfoo pipelines, CI/CD quality gates and production monitoring.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View
  • Retrieval and Generation Metrics for RAG Systems

    Hands-on programme covering information retrieval metrics (Recall@k, Precision@k, MRR, NDCG@k) and generation metrics (Faithfulness, Answer Relevance, Context Precision, Citation F1) with statistical testing.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View
  • Production RAG Architecture: Scaling, Freshness, Security and Operations

    Scaling RAG from prototype to production: event-driven incremental ingestion, zero-downtime re-indexing, multi-tier caching, security controls, latency SLAs and LLMOps observability.

    RAG & Enterprise Knowledge Systems Corporate
    Live Course
    6 modules
    View

Get in touch