What is a knowledge base?

A knowledge base is an organised collection of articles, policies, procedures and answers that a firm maintains so staff and customers can find accurate information. In support it holds how-to guides, product terms and troubleshooting steps. Once an AI agent answers from it, the knowledge base becomes the agent's source of truth, and every error in it is repeated.

Why does a knowledge base matter once agents act?

A person reading an outdated article might notice something is off, ask a colleague or check the system. An agent tends to take the article at its word and repeat it confidently. If two articles contradict each other, the agent may quote either. If an article is missing, the agent may fill the gap from general knowledge, which is how support agents end up hallucinating answers.

That makes knowledge base quality a control, not a housekeeping task. The articles an agent may use, who owns them, when they were last reviewed and which version applied to a given conversation all become part of the evidence of what the agent did and why.

How does a knowledge base work for an agent in practice?

The agent retrieves the articles relevant to a question and answers from them. To work well, the knowledge base needs:

• One answer per question: duplicates and near-duplicates cause contradictions.
• An owner and a review date for each article, with stale articles excluded automatically.
• Clear scope: who the article applies to, such as which product, customer type or market.
• Explicit limits: what the article does not cover, so the agent knows when to stop.
• Plain structure: short sections that answer one thing each.

Our guide on how to write a knowledge base an AI agent can use covers the writing in detail.

An example from a pensions administrator

Take a pensions administrator whose support agent answers member questions. Two articles describe the transfer-out process: one written before a policy change, one after. Both are live. The agent answers some members from the old article, telling them a form that is no longer accepted.

The fix is not in the agent. The old article is retired, each article gets an owner and a review date, and articles past review are excluded from the agent's sources until checked. The agent's certification is rerun against transfer questions to confirm it now cites only the current process.

How is a knowledge base different from a knowledge graph?

A knowledge base stores documents written for people to read: articles and policies in prose. Retrieval finds the right text and the agent reads it.

A knowledge graph stores facts as connected entities and relationships: this customer holds this product, which has this fee, which changed on this date. It answers precise questions about how things relate. Many agents use both, the knowledge base for explanations and the graph for facts that must be exact. Whether you need a knowledge graph for your AI agents depends on how much of the work turns on those relationships.

Which terms sit next to knowledge base?

• Retrieval: finding the passages relevant to a question before the agent answers.
• Grounding: tying every claim in an answer to a source the agent was allowed to use.
• Source of truth: the record that wins when two sources disagree.
• An AI support agent that knows when to say no refuses when the knowledge base has no answer.

Where Pryme fits

In Pryme, knowledge base articles reach an agent with their provenance: an article published from the firm's policy library is sealed, one owned by a named team is signed, and a pasted note is declared. Attest's grounded gate checks that answers rest on permitted sources.