Adding knowledge

Imagine hiring an experienced customer-service adviser. They know how to explain a return politely, but they do not know your company’s return policy. Experience is not a substitute for your handbook. The same distinction applies to an agent: a language model can write a convincing answer about refunds without knowing what your business actually promises.

Knowledge, in this unit, means reference material the agent can consult: approved policies, product catalogues, operating manuals, and other documents. Giving the agent this material is different from teaching it a new writing style. You are providing evidence for particular answers, not assuming that the model’s general training contains your company’s facts.

Why general knowledge is not your company knowledge #

A model learned patterns from material used during its training. That may help it explain common concepts, but it is not a dependable record of your current policies. Your delivery regions may have changed. Your catalogue may include a service launched after the model was trained. An internal exception might never have appeared in public at all.

Even familiar questions need local evidence. “Can I exchange this item?” sounds simple, yet the answer can depend on the product category, condition, sales channel, and applicable policy. A fluent explanation of how exchanges usually work could be completely wrong for this customer. The agent needs the relevant rule and enough information about the situation to apply it.

Knowledge also creates accountability. A supervisor can compare an answer with an approved document rather than debating whether it sounds reasonable. This does not make every answer correct automatically. The agent can still choose the wrong passage or misunderstand an exception. It does, however, give your team something concrete to inspect and improve.

Retrieve first, answer second #

Retrieval-augmented generation, usually shortened to RAG, means finding relevant documents before asking the model to answer with their help. Think of an adviser opening the right handbook page rather than reading the entire filing cabinet before every call. The system searches the available material, places useful passages alongside the question, and asks the model to produce an answer supported by those passages. This normally does not retrain the model or permanently change what it learned. The detailed mechanics belong in What is a RAG; for now, remember the order: find evidence, then explain it.

Searching is not the same as answering. A search can find a document about returns without finding the clause that covers a damaged item. Before relying on a passage, the agent must consider whether it addresses the actual question, applies to the right product or audience, and is still valid. If the needed evidence is missing, asking a follow-up question or contacting a person is better than completing the policy from imagination.

Try a small knowledge collection #

The demonstration below uses invented documents for a fictional shop. It illustrates selecting reference material, not the behaviour or accuracy of a production search service. Try terms such as “damaged,” “delivery,” or “manual,” then think about what the matching document does and does not establish.

A useful exercise is to ask a question the collection cannot answer, such as whether a particular parcel has arrived. None of these documents is a live shipment record. A good knowledge design makes that boundary visible. The agent should use an authorised tracking tool or explain that it cannot confirm the current status from these documents.

Choose the first documents deliberately #

Start with a recurring task, not every file the company owns. For a support agent, that task might be explaining returns and warranty coverage. For an internal assistant, it might be helping employees request equipment. For sales, it might be explaining product differences without inventing contractual promises.

Ask the people doing that work which questions recur, where the authoritative answers live, and which mistakes cause the most rework. Select documents that answer those questions directly. A short, approved policy is often a better starting point than a large folder of presentations containing contradictory drafts. More material is useful only when it adds relevant, trustworthy evidence.

Before adding a document, read it as a newcomer would. Does “standard package” refer to a named offer? Does “contact the usual team” identify an actual responsibility? Does the exception appear beside the rule it modifies? Material that depends on unwritten background knowledge is difficult for both a new employee and an agent to use reliably.

Keep private information out unless the use case genuinely requires it and access is appropriately restricted. An internal salary policy and a public product guide do not belong in the same unrestricted answer pool. The fact that a document is searchable must not mean that every user is entitled to receive its contents.

Make freshness somebody’s responsibility #

Freshness means how current the information is for its purpose. A cleaning guide may remain useful for a long time, while a promotion can become obsolete overnight. Upload date alone is not enough: a file added today may contain a policy withdrawn last year.

Assign a business owner to each subject. That owner decides which version is authoritative, when it takes effect, and what should happen to the previous version. Keep those decisions visible through a title, owner, scope, review date, and source reference. These details are often called metadata: information describing a document rather than the main document text.

A review calendar helps, but event-driven updates matter too. When the business changes a policy, updating the agent’s sources should be part of the same release checklist. Otherwise the website and the agent can give different answers even though both teams believe they published the new rules.

On AIVAX, groups of searchable documents are called collections. Semantic search finds relevant material by meaning rather than requiring an exact wording match. These features support retrieval; your team still owns document quality, access decisions, and the business meaning of the answer.

What’s next: learn when an agent should leave its stored documents and connect to the world for current information.