--- title: "GRAIN | The RAG Engineer" description: "GRAIN is a retrieval loop: Gather, Rank, Assemble, Inspect, Navigate. Gather governs the corpus: sources, permissions, and versioning. Rank retrieves…" image: "https://lenvanderhof.com/photos/len-portrait-1.jpg" --- Framework # GRAIN The RAG Engineer GRAIN is a retrieval loop: Gather, Rank, Assemble, Inspect, Navigate. Gather governs the corpus: sources, permissions, and versioning. Rank retrieves candidate passages and reranks them for the question actually asked. Assemble builds the context the model may see. Inspect traces a wrong answer back to the chunk and retrieval decision behind it. Navigate keeps the index fresh: TTL policy, re-ingest, and change detection. If you cannot point at the passage, you do not have retrieval. You have a fluent guess. The moves 5 ## When to use it Use it when a RAG demo cites the wrong paragraph with a straight face. If the team cannot open the chunk the model used, the loop is broken at Inspect. ## The moves 1. 01 ### Gather Source selection, permissions, versioning, and the corpus contract that decides what belongs in the index. 2. 02 ### Rank First-stage retrieval (dense, sparse, or hybrid), then a reranker that orders candidates before assembly. 3. 03 ### Assemble Token budget, chunk order, deduplication, instruction layout, and citations that keep evidence usable. 4. 04 ### Inspect Failure traces, bad-chunk workflows, and eval suites that connect a wrong answer to a retrieval decision. 5. 05 ### Navigate TTL policy, re-ingest jobs, change detection, and signals that show operators how old an answer could be. Failure modes ## Why the demo fails in production If you cannot open the source behind the sentence, you do not have RAG. You have a chat window with extra text pasted in. GRAIN is the inspectable loop. Failure is almost always a named letter, not a weaker model. 1. 01 ### Gather No corpus contract: permissions, versioning, and what belongs in the index were never written down. Stale files stay retrievable. 2. 02 ### Rank Candidates are ordered for a different question, or never reranked. The first-stage hit is treated as the answer. 3. 03 ### Assemble The cited chunk is not in the window the model actually saw. Context is stuffed; evidence is not usable. 4. 04 ### Inspect Nobody can open the passage used. There is no eval that ties a wrong answer to a retrieval decision. 5. 05 ### Navigate No TTL, no re-ingest, no signal of how old the answer could be. The demo was fresh. Production is not. What is RAG? ## The book Available now ### The RAG Engineer Retrieval Systems That Agents Can Actually Trust in Production Open the bookThe articleThe series ## Related frameworks STACK ### The Agentic Codebase STACK is a five-layer architecture for repositories where humans and agents share work: Structure, Toolchain, Agent configuration, Connection, Knowledge and quality. The agent OS is versioned like the app. SENSE ### AI Agents for Startup Strategy SENSE is a five-move strategic intelligence layer: Scan, Evaluate, Navigate, Simulate, Execute. It turns agent work from a productivity shortcut into a sourced claim that changes a watchlist, an assumption register, or a Monday decision. CHORUS ### The Multi-Agent Organization CHORUS is a six-skill coordination protocol for teams of 2 to 15 people who share work with specialised AI workers: Charter, Handoff, Orchestrate, Review, Update, Sync. Org charts name reporting lines. CHORUS names coordination lines. ROUTE ### The Model Portfolio ROUTE is a five-skill discipline for governing a mix of language models: Register, Objective typing, Utilize policy, Track, Evolve. Register catalogs every model with mandate, cost class, privacy class, and retirement status. Objective typing names what a task actually requires before a model is chosen. Utilize policy writes the routing rules, cascades, and fallbacks. Track measures cost, latency, and quality per route. Evolve promotes, demotes, and retires on a cadence. A leaderboard pick is a collection. ROUTE is a portfolio. ## Essays on this site - [What is RAG? Retrieval, then generation](https://lenvanderhof.com/en/blog/what-is-rag/) ## Terminology - [GRAIN](https://lenvanderhof.com/glossary/grain/) - [STACK](https://lenvanderhof.com/glossary/stack/) - [SENSE](https://lenvanderhof.com/glossary/sense/) - [CHORUS](https://lenvanderhof.com/glossary/chorus/) - [ROUTE](https://lenvanderhof.com/glossary/route/) - [AI agent](https://lenvanderhof.com/glossary/ai-agent/) - [Human in the loop](https://lenvanderhof.com/glossary/human-in-the-loop/) - [Override doctrine](https://lenvanderhof.com/glossary/override-doctrine/) Not a grain-size metaphor and not “we added a vector store.” GRAIN is the inspectable loop. RAG names the architecture. The RAG Engineer is available now in English and Dutch paperback, hardcover, and Kindle. ## Citation GRAIN (The RAG Engineer). Len P. van der Hof, Systems for the Strategic Self. https://lenvanderhof.com/frameworks/grain/ ```json { "@context": "https://schema.org", "@graph": [ { "@type": "DefinedTerm", "@id": "https://lenvanderhof.com/frameworks/grain/#term", "name": "GRAIN", "description": "GRAIN is a retrieval loop: Gather, Rank, Assemble, Inspect, Navigate. Gather governs the corpus: sources, permissions, and versioning. Rank retrieves candidate passages and reranks them for the question actually asked. Assemble builds the context the model may see. Inspect traces a wrong answer back to the chunk and retrieval decision behind it. Navigate keeps the index fresh: TTL policy, re-ingest, and change detection. If you cannot point at the passage, you do not have retrieval. You have a fluent guess.", "url": "https://lenvanderhof.com/frameworks/grain/", "inDefinedTermSet": "https://lenvanderhof.com/frameworks/" }, { "@type": "WebPage", "@id": "https://lenvanderhof.com/frameworks/grain/#webpage", "url": "https://lenvanderhof.com/frameworks/grain/", "name": "GRAIN", "description": "GRAIN is a retrieval loop: Gather, Rank, Assemble, Inspect, Navigate. Gather governs the corpus: sources, permissions, and versioning. Rank retrieves candidate passages and reranks them for the question actually asked. Assemble builds the context the model may see. Inspect traces a wrong answer back to the chunk and retrieval decision behind it. Navigate keeps the index fresh: TTL policy, re-ingest, and change detection. If you cannot point at the passage, you do not have retrieval. You have a fluent guess.", "inLanguage": "en", "isPartOf": { "@id": "https://lenvanderhof.com/#website" }, "about": { "@id": "https://lenvanderhof.com/frameworks/grain/#term" }, "mainEntity": { "@id": "https://lenvanderhof.com/frameworks/grain/#term" }, "breadcrumb": { "@id": "https://lenvanderhof.com/frameworks/grain/#breadcrumb" } }, { "@type": "BreadcrumbList", "@id": "https://lenvanderhof.com/frameworks/grain/#breadcrumb", "itemListElement": [ { "@type": "ListItem", "position": 1, "name": "Home", "item": "https://lenvanderhof.com/" }, { "@type": "ListItem", "position": 2, "name": "Frameworks", "item": "https://lenvanderhof.com/frameworks/" }, { "@type": "ListItem", "position": 3, "name": "GRAIN", "item": "https://lenvanderhof.com/frameworks/grain/" } ] } ] } ```