Chapter 17: Advanced RAG & Knowledge Systems¶

Go beyond top-k: hybrid sparse+dense retrieval, Reciprocal Rank Fusion, reranking, chunking strategies, and retrieval evaluation for production knowledge systems.

Metadata¶

Learning Objectives¶

• Implement BM25 from scratch and explain term frequency, IDF, and length normalization
• Build dense retrieval with cosine similarity over embeddings
• Fuse rankings with Reciprocal Rank Fusion (RRF) to combine sparse and dense signals
• Rerank a candidate set with a cross-encoder-style scorer
• Chunk documents with size/overlap trade-offs and attach metadata for filtering
• Evaluate retrieval with Recall@k, MRR, and nDCG
• Reason about hybrid search — when sparse beats dense and vice versa
• Design a production knowledge system with caching, indexing, and quality monitoring

What's Included¶

Notebooks¶

Scripts¶

• config.py — Chapter config: seeds, toy corpus, fusion constant k
• hybrid_search.py — BM25, dense cosine retrieval, Reciprocal Rank Fusion, reranker
• evaluation.py — Retrieval metrics: Recall@k, MRR, nDCG

Exercises¶

• Problem Set 1: BM25 & Dense — Implement IDF, score a document with BM25, and rank with cosine similarity
• Problem Set 2: Fusion & Metrics — Fuse rankings with RRF and compute nDCG over a labelled query
• Solutions — in exercises/solutions/ (notebooks and solutions.py for CI)

Diagrams (Mermaid)¶

• hybrid_retrieval.mermaid, chunking.mermaid

Read Online¶

• 17.1 Introduction — BM25 from scratch, dense cosine retrieval, and where each wins
• 17.2 Intermediate — Reciprocal Rank Fusion and a phrase-aware reranker
• 17.3 Advanced — Chunking strategies, metadata filtering, and retrieval metrics

Or try the code in the Playground.

How to Use This Chapter¶

git clone https://github.com/luigipascal/berta-chapters.git
cd berta-chapters/chapters/chapter-17-advanced-rag-and-knowledge-systems
pip install -r requirements.txt
jupyter notebook notebooks/01_sparse_and_dense.ipynb

Created by Luigi Pascal Rondanini | Generated by Berta AI