Resource Library · November 2026

Building Production-Ready RAG Systems

Production-ready, in this guide, means a system that passes twelve design tests — not a guarantee, not a bake-off, and not a claim that an answer is hallucination-free.

19 pages30 min readAdvancedCompanion workbook included

By AsymiLink Meta · Published · Updated

Request the guide and workbook

Summary

Production-ready, in this guide, means a system that passes twelve design tests — not a guarantee, not a bake-off, and not a claim that an answer is hallucination-free. Frameworks excerpted on this page: The twelve design tests; The anatomy of the system. A 19 pages November 2026 document, with a companion workbook.

  • Production-ready means the twelve tests pass — not a guarantee, and not a claim that answers are hallucination-free.
  • Citations are required, not a courtesy. A draft without a source is not finished.
  • Permissions travel with the chunk: a user cannot retrieve what they could not open in the source system.
  • Quality is measured against a frozen question set, scored by a person, including after a change.

A crawlable plain-text copy of this page is also at /resources/rag-architecture-guide.txt. The canonical URL is https://asymilinkmeta.com/resources/rag-architecture-guide.

Written for Technology, knowledge, risk and operations leaders who have to put a retrieval-augmented system into production, and the sponsors who have to fund one.

What's in this guide

This 19-page guide is a design method for putting a retrieval-augmented generation (RAG) system into production. “Production-ready,” wherever it appears here, means the system passes twelve design tests. It is not a guarantee, not a certification, and not a claim that an answer is hallucination-free. The guide will not rank embedding models, publish a best chunk size, or invent an accuracy figure. Those are things it has not measured, and will not invent.

It is written for the people who have to fund, design and operate a retrieval-augmented system: technology, knowledge, risk and operations leaders, and the sponsors who have to stand behind one. A companion workbook holds a test register, a corpus sheet and a permissions sheet, so a team can apply the tests to a named case rather than only read about them.

The rest of the library is the method around this one. The agentic-workflows paper is the companion for when an internal, read-only research assistant is the right first pattern. The offline deployment guide is the companion when the corpus cannot leave. The data-strategy guide is the companion for the named source of truth the corpus is built from. This page is a summary of what is inside the guide, written from the guide itself. The PDF and workbook are the working documents.

Who it is for, and what a first case looks like

The honest first case in this guide is read-only and internal. A later project can be a write-back or a customer-facing use; that is a different case, with the tests applied again. The first case has a named operator who is not only the builder, and a switch-off that can be used without a project. Isolation, or “just a search box,” does not make the operator optional.

If the corpus contains restricted data — protected health information, privilege, or other restricted material — that is a design and a counsel question, not something the guide will assume away. The HIPAA design framework and the governance framework in this library are the companions for those questions. This guide is not legal advice, and it will not close a counsel question.

The companion workbook is a working register, not a certificate. For each test you write a sentence, a named owner and a piece of evidence. A missing answer is a fail, not a pass. A completed workbook means the tests were asked. It does not mean the system is certified, and it does not mean an answer will never be fluent and wrong.

The twelve design tests

Section 2 applies the same twelve tests to every case — a document assistant and a system that would route work. A failed test on the source of truth, citations, the frozen set, permissions, or a fluent draft treated as finished is usually a reason to pause.

  1. T1

    The source of truth is named and approved. What is in the corpus, what is out, and who approved it. A shared drive is not a source of truth.

  2. T2

    Citations are required, not a courtesy. A draft without a source is not finished. The reader can go to the passage.

  3. T3

    Evaluation is against a frozen question set: questions the organization actually asks, written down before the first measure, scored by a person.

  4. T4

    Permissions travel with the chunk. A user cannot retrieve what they could not open in the source system.

  5. T5

    Chunking is a design choice, and it is tested. There is no “best chunk size.” The frozen set is how a choice is judged.

  6. T6

    Retrieval choices are tested, not crowned. Keyword, dense, hybrid, re-ranking: options to try, not a league table.

  7. T7

    An operator, and a switch-off. A named operator who is not only the builder. Isolation does not make this optional.

  8. T8

    No write-back, and no customer-facing use, unless a later project. The honest first case is read-only and internal.

  9. T9

    Restricted data in the corpus is asked about, not assumed away. That is a design and a counsel question.

  10. T10

    If the corpus cannot leave, the offline guide is the companion. A local front end on a hosted service is not “on-premise RAG.”

  11. T11

    Training and reuse are written down. A silence is a fail.

  12. T12

    A fluent answer is not a finished one. “Hallucination-free” is not a claim this guide will make.

The full wording, and a register you can fill in, are in the PDF and the companion workbook.

The anatomy of the system

Around the tests, the guide walks the path an answer actually takes: the corpus, the chunk, the index, the draft, the person, and the measure. Three of those are where a convenient build quietly fails.

  1. Corpus

    T1 lives here. What is in, what is out, and who approved it. A drive that was never reviewed is not a source of truth.

  2. Draft and citations

    An answer assembled from retrieved chunks, with citations (T2). A draft without a source is not finished.

  3. Permissions

    A login in front of the assistant is not T4. Permissions travel with the chunk, including after a document is added or removed.

  4. The frozen set

    A change to the model, the chunking, or the retrieval is not finished until it is re-measured against the same frozen questions, by a person who can score (T3).

Get the guide and workbook

Enter your name, work email and company. The files appear on this page immediately. A backup download link is also emailed when email delivery is available.

By requesting this resource you agree that AsymiLink Meta may contact you about it. See our Privacy Policy and Terms of Use.

Frequently asked questions

What does “production-ready” mean in this guide?

It means the twelve design tests pass. It is not a guarantee, not a certification, and not a claim that a system is hallucination-free. A fluent answer is not, on its own, a finished one.

Does the guide recommend a chunk size or an embedding model?

No. Chunking and retrieval are design choices, tested against a frozen question set for this corpus. The guide will not crown a method or publish a recommended size.

What is in the companion workbook?

A design-test register, a corpus sheet and a permissions sheet. For each test: a sentence, a named owner and a piece of evidence. A completed workbook is not a certificate.

What does a first RAG case look like here?

Read-only and internal, with a named operator and a switch-off. A write-back or a customer-facing use is a later project, with the tests applied again.