Context Engineering for a Knowledge Base: What to Retrieve

Context engineering for a knowledge base means selecting the source passages an AI needs for a task, preserving their meaning and checking that the right material was retrieved. Start with the question, identify authoritative documents, attach current decisions and verify the answer against its evidence.

A reference document rarely says where active work stands. fenbs can supply the project’s decisions, rules and latest progress alongside those documents, giving the assistant both the background and the current situation.

How do you choose material for the next task?

1. Write the question the assistant must answer and the output you expect.
2. Find the authoritative source for each important claim or instruction.
3. Select passages that include the relevant rule and its exceptions.
4. Add the current project decision and work state separately from reference material.
5. Test whether the assistant can find, cite and correctly apply those passages.

Say a hypothetical home-services company wants an assistant to draft a response about rescheduling. The useful material is the applicable appointment policy, the service area and whether this appointment has already been changed. A company history page may be accurate but contributes little to that response.

Write a small source map before automating anything. One row per question is enough: question, controlling source, relevant section and person who maintains it. This reveals gaps that adding more documents would hide. Start with a few recurring questions that someone on the team can answer correctly without AI.

Choosing the application itself belongs in the knowledge-base software guide. This workflow applies to the sources you already have; it does not require replacing their home.

What labels make a retrieved passage usable?

A paragraph needs enough surrounding information to tell the assistant when it applies. Keep the document title, section, version or effective date, owner and status attached to its content. “Draft” and “approved” are meaningful facts, not decorative labels.

Preserve qualifications inside the passage too. If a rule applies only before a technician leaves, retrieve that condition with the rule. Splitting documents into smaller pieces can make them easier to find, but a fragment that loses its exception can support the wrong answer.

Microsoft’s RAG overview↗ (opens in a new tab) describes retrieval-augmented generation, which supplies source material to a model, and stresses relevance and authorized access. Use those principles even for a manual workflow: choose useful passages and confirm that the reader is allowed to receive them.

What should accompany the source documents?

Add a small work record: the goal, agreed exceptions for this project, unresolved questions and next step. Label it separately so a temporary project choice does not become a company-wide policy. Keep a link back to the document instead of reproducing a second full copy.

For the scheduling example, the policy describes the general rule. A person’s recorded decision may permit an exception for the current task. The assistant needs the decision’s scope and reason, plus any remaining approval, before drafting the message. If the record merely says “exception discussed,” it should ask whether anyone approved it.

This is where fenbs AI context helps: Rules and Where we left off can explain how the source applies to current work. The source documents still hold the detailed reference material, and the work record explains what people chose and what happens next.

How do you keep old material out of current answers?

Assign someone to maintain each controlling source and make replacement explicit. When a policy changes, update its status, identify the new version and check the places that still retrieve the old one. An archive can remain useful for history while being excluded from ordinary current-policy answers.

If you use a search index, which is a searchable copy of source content, verify its update and removal behavior. Microsoft’s change and deletion documentation↗ (opens in a new tab) illustrates the distinction: detecting changes and removing deleted source material require different handling in that system.

Test the transition with a known old statement. Ask the assistant a question that used to produce that statement and inspect the evidence it now returns. A fresh date on the source page does not prove the retrieval system has refreshed its copy. Check the actual answer path.

For work involving independent assistants, context engineering for AI agents covers the broader session workflow. Keep this retrieval check focused on whether the active source reached the current task.

How can you tell whether retrieval is working?

Keep a small set of questions with known answers and supporting passages. Include an ordinary case, an exception, conflicting documents and a question your sources cannot answer. Write the expected behavior before testing so an eloquent answer does not change your standard.

• Source selection: did the relevant passage appear, with its conditions?
• Authority: did the assistant use the approved version and flag conflicts?
• Application: did it apply the rule to this case without widening its scope?
• Missing evidence: did it identify an unanswered question instead of inventing a policy?
• Access: could each intended reader receive only appropriate material?

Change one part at a time when a test fails. A missing passage suggests a search or selection problem; a correct passage used incorrectly suggests an instruction or interpretation problem. The difference between those layers is explained in context engineering vs prompt engineering.

How to do this with fenbs

1. Create a free account and add the project whose work depends on your knowledge base.
2. Add tasks for the actual outputs. In their plans, identify the sources, sections and verification each output needs.
3. Record the people-approved decisions and rules. Use Links and listings to keep useful source references with the work, and context notes for facts the next assistant needs.
4. Connect a supported AI app from Integrations. Ask it to read project context, open the relevant sources through an available supported method and identify gaps before working.
5. Review its answer against the source, then save progress and any lesson about a failed retrieval. Update the work record when people change the decision.

fenbs is useful for current work context; it does not index your knowledge base, provide vector search or automatically import its documents. If your main need is a searchable document portal, keep that capability in your knowledge-base system. Use the context engineering hub to explore the other parts of an assistant’s context.