UNDERSTAND THE PROBLEM
Begin with the failure you need to prevent.
Retrieval-augmented generation is often described as a search feature attached to a model. In a consequential product, retrieval is an evidence boundary.
The system should not merely find text that looks relevant. It needs to know which sources are approved for this use, which version was active, whether the evidence supports the proposed action and what to do when sources disagree or say too little.
The right architecture connects content governance to runtime behaviour. A reviewed change should be publishable, traceable and testable without hiding inside a vector index no one can inspect.
DESIGN THE SYSTEM
Make the operating rules explicit.
A dependable AI system is easier to build when the team can see the decisions, evidence, boundaries and ownership around it. The following principles turn an ambiguous ambition into components that can be implemented and reviewed.
Treat knowledge as a product with owners and releases.
Every source needs an owner, intended use, review state, effective version and retirement path. Retrieval quality cannot repair ungoverned content.
Retrieve for the decision
Index and rank evidence around the task the system must complete, not only semantic similarity to the user's words.
Keep provenance with the answer
Preserve the source, passage, version and retrieval method so an operator can reconstruct why the system responded.
Design the no-evidence route
When approved knowledge is missing, the product should ask, abstain or escalate. The model must not fill the gap from confidence or memory.
IMPLEMENT IN ORDER
Build the smallest complete loop.
Do not automate every adjacent task at once. Start with one valuable journey, carry it from signal to outcome, and preserve enough evidence to know whether it worked. Expand only after that loop is dependable.
- 01
Inventory the sources
List the content the system may use, its owner, audience, review status, jurisdiction and actions it can support.
- 02
Design versioned ingestion
Create stable document identities, chunking rules, metadata and approval states before generating embeddings.
- 03
Test retrieval separately
Build queries with known relevant and misleading sources. Measure whether the right evidence enters the context before judging the final answer.
- 04
Constrain generation
Tie allowed claims and actions to retrieved evidence, and make insufficient support a first-class route.
- 05
Re-evaluate every release
Run retrieval and answer checks when content, chunking, embeddings, ranking, prompts or models change.
KNOW WHEN IT WORKS
Measure behaviour, not how impressive the demo looks.
The useful measure is whether the system creates the intended business or product outcome while staying inside its boundary. Review these checks before launch and whenever the model, data, prompt, tools or workflow changes.
Retrieval quality
Relevant approved evidence appears before irrelevant or retired material.
Grounded behaviour
Claims and actions remain within what the retrieved evidence actually supports.
Traceability
An authorised reviewer can reconstruct the source and version used for an outcome.
Fallback
Missing or conflicting evidence produces a safe, useful product response rather than a guess.