Connecting agents to external knowledge so they can retrieve facts rather than hallucinate them.
Retrieval-Augmented Generation (RAG) addresses the fundamental limitation of language models: their knowledge is frozen at training time. RAG connects agents to live, organisation-specific knowledge stores so they can retrieve accurate, current, authorised information before generating responses.
A RAG system has three core components. The knowledge store: a collection of documents, records, or data points that the agent can query. The retrieval mechanism: typically a vector similarity search that finds documents most relevant to the current query. The generation step: a language model that uses the retrieved documents as context to generate an accurate, grounded response. The critical architecture decisions are: how is the knowledge store maintained and updated? How is access to documents controlled? How does the system handle retrieval failure — when the query returns no relevant documents? And how does the model signal when its response is grounded in retrieved context versus falling back on parametric knowledge?
A large retailer deploys a RAG-based agent for its store operations team. The knowledge base contains product specifications, supplier contracts, store procedures, and regulatory compliance guides — updated nightly. When a store manager asks about the lead time for a specific product, the agent retrieves the current supplier contract and answers from it. When asked about a procedure that was updated last week, the agent retrieves the new version. Access controls ensure that agents serving store staff cannot retrieve documents marked as executive-only. Retrieved source documents are cited in every response, enabling the user to verify the answer.
Without RAG, agents answer questions about your organisation's current policies, products, and procedures from training data that may be months or years out of date — or from hallucinated content that sounds plausible but is wrong. RAG is the pattern that makes agents reliable for knowledge work: it grounds responses in authoritative, current, organisation-owned information.
How this pattern fails in practice - and what to watch for.
The agent retrieves a document and then hallucinates a paraphrase of it that says something the document does not say. The response looks grounded because it cites a source, but the cited content doesn't support the claim. This is more insidious than pure hallucination because the citation creates false confidence.
The retrieval system returns a document that has been superseded by an update. The indexing pipeline failed to process the updated version. The agent confidently cites the outdated policy. In a regulated environment, this can constitute a compliance failure.
The retrieval system does not enforce document-level access controls at query time. An agent serving a customer retrieves an internal pricing strategy document because it is semantically similar to the customer's query. The agent incorporates internal information into its customer-facing response.
Seven things to verify before deploying this pattern in production.
Use the checklist and failure modes above as review prompts. Record the evidence that each mapped PSF domain and PAI-8 control is operating, name the owner for unresolved gaps, and set a date to review the pattern again after material system changes.
Use this pattern library as an implementation reference for deployment safety, governance, and audit evidence. Pair each pattern with named owners and evidence from the system you are reviewing.