HomeAI Agency AcademyLesson 23
Module 06 · Lesson 23

Build a knowledge base the agent can cite

Curate approved, current information with owners and retrieval boundaries.

Last updated August 5, 202615–25 minutesFree AI agent course
What you will learn

Build the knowledge source register

You will be able to prepare a knowledge base that makes answers traceable to an approved source rather than a collection of loosely related documents.

Why this matters

Good agent work is useful before it is impressive.

Uploading every document is not knowledge management. An agent needs current, scoped material; clear ownership; a way to show what it relied on; and a stop rule when the answer is absent or uncertain.

Field note 23

Connect approved sources, retrieval, answers, citations, and updates

Connect approved sources, retrieval, answers, citations, and updates
Connect approved sources, retrieval, answers, citations, and updates
Core concepts

The language that keeps the work clear.

Approved sourceContent the business has reviewed for this purpose and agrees is safe to use.
FreshnessWhether the source is current enough for the decision or customer question.
Retrieval boundaryThe limits on what collections, customer data, and document types may be searched for a job.
CitationA usable reference to the source or policy that supported an answer or recommendation.
The practical method

How to build the knowledge source register

Choose the first use case

Build knowledge around a small set of recurring questions, not every document the business owns.

Curate with owners

For each source, name the subject owner, review date, audience, and permitted use.

Make answers attributable

Require a reference or source name when the agent gives a factual answer or explain that a person must confirm it.

Retire stale material

Create a simple review rhythm so old policy, product, and pricing documents stop influencing the workflow.

Worked example

Worked case: build a knowledge base the agent can cite

A SaaS company wants an agent to answer implementation questions. It starts with current onboarding guides, supported integration documentation, and a short escalation matrix owned by product support.

The agent retrieves only the relevant approved source, answers in plain language, and adds the source title to its internal handoff. If the question asks about a beta feature or a custom security condition, it does not infer an answer from a marketing page.

A monthly owner review removes retired guidance and adds the recurring unanswered questions that should become approved material.

Build it in practice

Complete the working artifact

Source: [title]. Owner: [role]. Review date: [date]. Permitted question types: [list]. Do not use for: [list]. Citation or handoff format: [format].
Retrieval check

Retrieval quality is an operating process.

A vector database is only one part of a knowledge system. Test ingestion, parsing, chunking, metadata filters, retrieval, reranking, citation, source access, freshness, correction, and deletion. An answer should say when the available evidence is weak or contradictory.

  • Build questions whose answers live in different source types and sections.
  • Measure whether the right evidence was retrieved before grading the prose.
  • Assign an owner and review date to every approved source.

Sources used for this check

Practice

Before you move on

  • Choose ten high-value, approved sources for one pilot.
  • Identify two documents that should not be retrieved by the agent.
  • Test a question with no approved answer and confirm the agent stops.
  • Sources have owners and review dates.
  • The agent can show what supported an answer.
  • Stale or irrelevant material is excluded.
  • Unknown answers have a human route.

Lesson progress

Finished this lesson?

Save your place on this device so it is easy to pick up where you left off.

Not marked complete yet.