05 / AI Explained · Field briefing
Ask your documents. Follow the evidence.
An answer should lead somewhere.
A language model's general knowledge is not a current copy of your organisation's policies. Retrieval-augmented generation, or RAG, adds relevant document material to a request. The model then produces an answer using that material as context. This is a way to supply evidence, not a guarantee of correctness.
How can an answer stay connected to information I can inspect?
An explanatory schematic—not a live system or measured performance.
KAF World Digital · Prepared 4 October 2026 · First editorial edition. Worked examples are invented for teaching.
Try the example02 / Where retrieval can fail
Where retrieval can fail
The workflow starts before the question. Documents need readable text, ownership, version information and access controls. They may be split into chunks and indexed for keyword or semantic search. A chunk should preserve enough surrounding context to interpret a statement: a heading, date or exception can change the meaning of a nearby paragraph.
In an invented expenses library, the current policy allows certain rail travel, while an old memo discusses a temporary exception. A query about the exception can retrieve both. The answer needs to distinguish dates and authority rather than combine the documents into one apparently coherent rule.
Retrieval and generation have different failure modes. Search may miss the applicable section. It may find an obsolete but similar passage. The model may misread a correct passage, overstate it or cite a source that does not support the claim. Checking only whether an answer has a citation misses those problems.
A useful citation lets the reader reach a specific document location and inspect the relevant passage. Its presence does not settle whether the source is authoritative, complete or current. Some APIs return structured citation locations; others require an application to track source IDs and locations. The interface should make the difference clear.
Permissions must be enforced before the model receives restricted material. Hiding a citation from the reader does not undo disclosure through a generated summary. Indexing, retrieval, logging and exports all need the same access boundary.
Evaluate with questions that should have answers and questions that should not. Include conflicting versions, missing pages and users with different permissions. A useful refusal explains which evidence is missing and how to obtain it, rather than making up the absent rule.
Evidence for the technical mechanisms: Microsoft · Retrieval Augmented Generation, Anthropic · Citations. The worked examples and practical framework are KAF editorial recommendations.
03 / Worked example
Inspect the answer
Follow a question from search to a passage you can inspect. This interactive lesson uses prepared synthetic examples and local calculations.
Use the companion worksheet and the worked example. All assumptions are illustrative.
No account, model API or external service is connected. Inputs are not submitted or stored by this page.
04 / Go deeper
Useful words. Important details.
- Retrieval
- Finding candidate material relevant to a question.
- Chunk
- A document segment indexed for search with enough context to interpret it.
- Groundedness
- Whether an answer’s claims are supported by the supplied evidence.
- Citation
- A pointer to a source location; its presence alone does not prove the claim.
Retrieval and answering are separate tests
Check whether retrieval finds the applicable passage, then whether the answer faithfully represents it. Fluency, retrieval similarity and citation presence are not equivalent to correctness.
Permission precedes context
Enforce access rules while selecting documents and passages. Post-generation hiding of a source cannot undo a restricted passage that was already supplied to the model.
Handle versions explicitly
Include source identifier, location, effective date and supersession status where available. A newer file timestamp alone does not establish a policy’s authority.
The technical notes qualify the explanation; they are not product-specific assurance or a certification.
05 / Evidence & scope
Inspect the source.
Keep the context.
Sources were accessed on 4 October 2026. These technical and security references support the mechanisms described; they do not validate our fictional examples or establish comparative product performance.
- Microsoft · Retrieval Augmented Generation
Retrieval, document chunking, search and evaluation of grounded answers.
- Anthropic · Citations
Structured citation locations in a provider implementation. Citations do not establish source authority or factual correctness.
- OWASP · Sensitive Information Disclosure, 2026
Information can leak through retrieval, logs, derived representations and outputs. This briefing uses the general risk framing, not legal conclusions or incident statistics.
- OWASP · Misinformation, 2026
Unsupported outputs, stale context, omissions and claim-check-act patterns. A risk framework, not a research accuracy benchmark.
What this briefing does—and does not—establish
The worked examples, suggested questions and worksheets are original KAF teaching material. Numbers are explicitly illustrative. No customer data, observed adoption results, market-size forecast or live provider pricing is presented.
Check the exact product, release, configuration and policy before using a workflow. Source documents can change; this is a dated educational briefing.
Keep a useful copy.
Read the full briefing offline, or use the companion worksheet to apply it.