AI Systems Research guide

What is GRAIN? The inspectable retrieval loop

RAG names the architecture. GRAIN names the honesty check.

Five-station mail sorting belt: incoming crate, pigeonholes, assembly tray, inspection lamp, a switch sending old mail aside
Five moves. Inspect fails the run. Navigate is freshness, not another lookup.

People search “GRAIN” after a slide used the letters and nobody wrote down the moves. Some of them meant cereal. Some meant a vector database. The useful object is smaller.

What is GRAIN? A retrieval loop: Gather corpora, Rank and rerank, Assemble context, Inspect failures, Navigate freshness. If you cannot inspect the retrieved and supplied passages, you do not have inspectable retrieval. You have a fluent guess.

What is RAG? owns the architecture. This page owns the letters and the collision.

The five moves

Not a product pack. Five jobs with owners.

Gather. Corpora: sources, permissions, structure, versions. Wrong splits and stripped metadata fail here, before any embedding argument starts.

Rank. Retrieve candidates and rerank them for the question actually asked. In the index but not in the top-k is a Rank miss, not a model miss.

Assemble. Context budget, ordering, citations. Right chunks retrieved, then dropped, buried, or duplicated, is an Assemble miss.

Inspect. Traces, evals, bad-chunk analysis. If the team cannot name which boundary failed, the regression ships undetected. Evaluate the reasoning, not the fluency is the same demand in other clothes.

Navigate. Freshness: updates, staleness, TTL. Confident answers from superseded documents fail here, not at the embedding.

The lamp is Inspect. Demo quality dies there. Freshness is Navigate. Do not smuggle “the next lookup” into the N.

Collision control

Not grain size. Not a milling metaphor. Not a cereal brand.

Not “we added a vector store.” A store covers part of retrieval. It does not supply reranking, context assembly, failure inspection, or freshness control.

Not RAG itself. RAG, retrieval-augmented generation, is the method: look up a named passage, then write. GRAIN is how you keep that method honest. You can have RAG theatre without GRAIN. You cannot have GRAIN without retrieval.

If someone says GRAIN and means a vendor feature named Grain, ask which of the five moves they own. Empty answer, empty loop.

What this page is not

It is not the book-length build: chunking, hybrid retrieval, rerankers, grounding, freshness, evals, drift. It is not a chunking tutorial. It is not a bake-off.

The RAG Engineer owns the expansion. It is live in English and Dutch paperback, hardcover, and Kindle. You do not need the hardcover to fail a demo that cannot open its own citation.

Write the five moves on a card. Run one failure trace. If you cannot locate the failed boundary, the retrieval system is not inspectable.

Terminology

Sources

  1. GRAIN (glossary)
  2. What is RAG?
  3. Evaluate the reasoning, not the fluency
  4. The RAG Engineer

Further reading

Markdown for LLMs