Building with Retrieval

The book

Most RAG systems that fail in production don't fail because of the model — they fail because of chunking decisions made in a notebook, retrieval pipelines that looked fine on twenty documents, or indexes that quietly go stale. Building with Retrieval works through the full stack: embedding models, vector stores, hybrid search, reranking, and the prompt patterns that keep answers grounded in what was actually retrieved. Marcus Hale draws on real deployment decisions — comparing costs in dollars at production query volumes across Pinecone, pgvector, and Qdrant, weighing LlamaIndex against LangChain for orchestration, and treating provenance as a first-class concern through the Anthropic citations API. It also tells you, with specifics, when RAG is the wrong solution entirely.

What you'll learn

• Why chunking strategy has more impact on retrieval quality than any other single decision, and which patterns hold up under production load
• The operational and cost tradeoffs between pgvector, Pinecone, Weaviate, and Qdrant — what each buys you and what each costs you at scale
• When hybrid search (BM25 combined with dense vector retrieval) outperforms either approach alone, and how a reranking pass sharpens results further
• How to write prompts for grounded answers, surface citations using the Anthropic citations API, and make provenance something users can actually verify
• Update cadences, expiration policies, and the failure modes that follow when index freshness falls out of sync with your data
• Evaluation methods that go beyond manual review: measuring recall, faithfulness, and answer relevance in a repeatable, automated way

Who it's for

Backend and ML engineers who have a RAG prototype working and are now facing the problems tutorials skip: retrieval precision at scale, keeping the index current, measuring whether the system is actually good, and deciding when long context or fine-tuning would serve better than retrieval.