Retrieval-Augmented Generation

Knowledge cut-offs, hallucination, attribution and cost — the four problems retrieval actually solves.

Assumes: Hallucination

The full pipeline from query to grounded answer, with the indexing and serving paths separated.

Assumes: Why RAG Exists

PDFs, HTML, tables and scanned documents; layout-aware extraction and metadata capture.

Assumes: RAG Architecture

Fixed, recursive, semantic, structural and late chunking; overlap, size selection and the trade-offs.

Assumes: Document Ingestion and Parsing

Bi-encoders, dimensionality, Matryoshka embeddings, domain fit, and how to benchmark a choice.

Assumes: Chunking Strategies · word2vec: CBOW and Skip-Gram

Cosine, dot product and Euclidean; normalisation effects and metric/model compatibility.

Assumes: Embedding Models

Exact flat search, IVF partitioning, tree and LSH approaches, and the recall/latency frontier.

Assumes: Similarity Metrics for Retrieval

Navigable small-world graphs, the layered skip-list structure, and the M and efSearch parameters.

Assumes: Vector Index Structures

Compressing vectors into codebooks, asymmetric distance computation, and memory/recall trade-offs.

Assumes: HNSW

Metadata filtering, hybrid storage, upserts and deletes, multi-tenancy and access control.

Assumes: HNSW

The BM25 formula derived and computed by hand, inverted indexes, and where lexical beats semantic.

Assumes: TF-IDF

Combining dense and sparse results, reciprocal rank fusion, score normalisation and weighting.

Assumes: BM25 and Sparse Retrieval · Vector Databases

Rewriting, expansion, decomposition of multi-hop questions, and routing across sources.

Assumes: Hybrid Search and Fusion

Hypothetical document embeddings, multi-query generation, and step-back prompting.

Assumes: Query Transformation

Why bi-encoders lose precision, cross-encoder scoring, the retrieve-then-rerank budget and latency cost.

Assumes: Hybrid Search and Fusion

Token-level matching with MaxSim, the storage cost, and PLAID-style optimisations.

Assumes: Reranking with Cross-Encoders

Ordering retrieved chunks, deduplication, contextual compression and budgeting the window.

Assumes: Reranking with Cross-Encoders

Positional bias in long contexts, the empirical U-curve, and reordering to exploit it.

Assumes: Context Assembly and Compression

Prompt patterns that force reliance on context, span-level citation, and refusal when evidence is absent.

Assumes: Lost in the Middle

Recall@k, precision@k, MRR, NDCG and MAP, each computed by hand, plus building a golden set.

Assumes: Grounding Prompts and Citations

Faithfulness, answer relevance, context precision and recall, and detecting unsupported claims.

Assumes: Evaluating Retrieval

Automated RAG evaluation, the metric definitions, and calibrating automated scores against human judgement.

Assumes: Evaluating Generation: Faithfulness

Searching small chunks but returning larger context, sentence windows and summary indexes.

Assumes: Context Assembly and Compression

Models deciding when to retrieve, grading their own retrievals, and fallback to web search.

Assumes: RAGAS and Evaluation Frameworks

Multi-step retrieval loops, tool selection, query planning, and termination conditions.

Assumes: Self-RAG and Corrective RAG · LLM Agents

Extracting entities and relations into a knowledge graph, community summarisation, and global questions.

Assumes: Agentic RAG · Knowledge Graphs and Embeddings

Retrieving over images, tables and charts; unified versus separate embedding spaces.

Assumes: GraphRAG · Multimodal LLMs

Index freshness and incremental updates, caching, cost control, monitoring and access control.

Assumes: Multimodal RAG