AI & ML / README.md

RAG and embeddings

Updated 2 min read index source
On this page6
  1. Foundations
  2. The pipeline, stage by stage
  3. Storage
  4. Making it good
  5. The three answers worth having
  6. The order to improve things

RAG and embeddings

The largest folder here, because it is what most applied-AI roles actually build. Read 01 and 02 first; everything else refines one stage of the pipeline they describe.

Foundations

# File Covers
01 What is RAG? two pipelines, RAG vs fine-tuning, where it fails
02 What are embeddings? static vs contextual, normalisation, the re-embed trap
03 RAG Architecture Patterns naive, advanced, modular, agentic

The pipeline, stage by stage

# File Covers
05 Chunking Strategies and Retrieval Techniques chunk size, overlap, strategies
07 Structured Document Pipelines — Parsing, Tables, OCR, Extraction PDFs, tables, OCR, vision fallback
08 Hybrid search and reranking BM25 fusion, cross-encoders, RRF
11 Fine-tuning embeddings when the embedder itself is the problem

Storage

# File Covers
04 Vector Databases index algorithms, the landscape, filtering
10 pgvector in production HNSW vs IVFFlat, ef_search, the filtered-ANN cliff
06 Knowledge Graphs and GraphRAG when graphs beat vectors

Making it good

# File Covers
09 Agentic RAG and evaluating retrieval recall@k, MRR, faithfulness, agentic retrieval

The three answers worth having

RAG turns a knowledge problem into a retrieval problem. It is two pipelines, not one: ingest runs per document, query runs per request. Naming rewrite and rerank as distinct stages is what separates a production answer from a demo.

RAG systems fail at retrieval, not generation. Measure recall@k before touching the prompt — if the evidence was never in the candidate set, no amount of prompt engineering recovers it.

Changing the embedding model means re-indexing everything. Vectors from two models are not comparable and there is no conversion, so it is a migration rather than a config change. Version the index and cut over.

The order to improve things

  1. Chunking — usually the largest win, and free.
  2. Hybrid search plus a reranker — the second largest, and cheap.
  3. Query rewriting — when queries are conversational or underspecified.
  4. Fine-tuning the embedder — last, and only for genuinely unusual vocabulary.

Contents 12