Find the shelf before writing the answer

Imagine answering a question in a library. First you locate relevant books, then read selected passages, then compose an answer. Retrieval-augmented generation, often shortened to RAG, combines a retrieval step with a generative model. The retrieved material becomes part of the information available for answering.

In a help center, the retrieval step might select documentation related to a particular feature. The model can then explain that material in a conversational format rather than relying only on patterns learned during training.

Two stages create two kinds of failure

The retriever may select the wrong document, miss an important exception, or return an outdated version. Even when the right passage is available, the generator may misread it or extend the claim beyond what the source says. A citation does not prove that both stages worked.

Test a simple question whose answer is explicitly stated in one document. Then try a question that requires combining two sections. Finally, ask something the collection does not answer. A dependable system should identify that gap rather than manufacture certainty.

Grounding needs maintenance and boundaries

Document permissions, versioning, and update schedules are part of the design. Private material should not become available to someone merely because a search component can index it. Obsolete documents need a clear retirement process.

For a small collection, begin with a manually curated source set and clear references. Do not assume a complex retrieval pipeline is necessary. Measure whether it improves answer quality and review effort before investing in additional infrastructure.

Further reading

Original sources and related references. These links lead to independent websites; inclusion does not imply a partnership or endorsement.

Browse the complete resource shelf →