- In short
- Curating uploaded knowledge means keeping it current, relevant, and free of duplicate or superseded versions of the same document. A knowledge base holding several versions of the same policy invites Claude to cite the wrong one, so curation involves removing deprecated versions as new ones are added, not just adding new files. Uploaded knowledge needs the same ongoing care as a connected external source, because duplicate or conflicting sources degrade output even when every individual document is accurate.
Adding a document is only half the job
The knowledge base is where uploaded reference documents live, and it is easy to treat as a folder you only ever add to. The CCAO-F exam treats keeping it curated as an apply-level skill precisely because the add-only habit is what breaks it. Uploading the new version of a policy without removing the old one leaves both in the base, and now Claude has two sources of truth for the same fact.
The failure this produces is subtle because it is not about any single document being wrong. Each version of the policy may be perfectly accurate for its own time. The problem is the conflict: with two versions present, Claude can cite either, including the superseded one, and the answer that comes back may be correct-looking but based on the wrong source.
- Curating uploaded knowledge
- The ongoing practice of keeping a Project's uploaded knowledge current, relevant, and free of duplicate or superseded versions of the same document. Curation includes removing deprecated documents as new ones are added, not merely adding files. Uploaded knowledge needs the same continuous care as a connected external source, because conflicting sources degrade output even when each document is individually accurate.
Why duplicates degrade output
When a knowledge base holds three versions of the same policy, nothing announces which one is authoritative. Claude may draw on whichever it retrieves, and if that is a deprecated version, the output is grounded in outdated facts even though the base "contains the right answer" somewhere in it. The presence of the correct document does not neutralise the presence of the wrong one.
This is why the failure survives a naive quality check. Someone auditing the base might open each document, confirm it is a real and once-valid policy, and conclude everything is fine, because every file is individually accurate. The defect is not in any file; it is in the collection holding conflicting versions side by side. Curation targets exactly that: the health of the set, not just the correctness of each member.
Curate it like a shared drive
The practical model is to treat uploaded knowledge the way you would a well-run shared drive: keep it current, keep it relevant, and remove deprecated versions as you add new ones. When a new policy version arrives, the job is not "upload the new one" but "upload the new one and remove the old one," so the base always holds a single authoritative version.
The other half of the lesson is that uploaded knowledge is not set-and-forget. It needs the same ongoing care a connected external source gets. A connector reaching a live system stays current on its own; a static upload only stays current if someone maintains it. Assuming an uploaded document is permanently fine because it was accurate when added is the mindset that lets duplicates accumulate.
What the CCAO-F exam trips candidates on
The exam sets two traps. The first is uploading a new policy version without removing the superseded one, leaving Claude able to cite either. The scenario looks like a diligent update, but adding without removing is exactly the miss; the credited answer removes the old version so only the current one remains.
The second is assuming uploaded knowledge is set-and-forget once added, unlike a live connector. This is the mindset that produces the first trap over time. Uploaded knowledge needs the same continuous curation as a connected source, and a question that hinges on why a well-populated base still produces wrong citations is pointing at neglected curation, not a bad individual document.
Worked example
A team's Project keeps producing answers that cite last year's expense policy, even though this year's policy was uploaded weeks ago. An audit finds the knowledge base contains Expense_Policy_2025.pdf and Expense_Policy_2026.pdf, both accurate documents. What is wrong, and how is it fixed?
The defect is in the collection, not in either file.
Both documents are individually accurate, which is why the audit finds nothing wrong with them one by one. But holding two versions of the same policy means Claude can cite either, and here it is sometimes citing the superseded 2025 version. The team did the add half of the update, uploading the 2026 policy, but not the remove half, so the old version still sits in the base inviting the wrong citation.
The fix is curation: remove Expense_Policy_2025.pdf so the base holds a single authoritative version. Notice the underlying habit that caused this. Uploaded knowledge was treated as set-and-forget, as if adding the new policy was the whole job. Keeping the base free of superseded duplicates, the way a shared drive is kept clean, is the ongoing care uploaded knowledge requires, and it is what returns the citations to the current policy.
Common misreadings to avoid
Misconception
Updating a policy in the knowledge base just means uploading the new version.
What's actually true
Misconception
If every document in the knowledge base is individually accurate, the base is fine.
What's actually true
How this shows up on the exam
Domain 5 questions describe wrong or outdated citations from a base that clearly contains the right document too. Read for a superseded duplicate left in place, and choose the answer that removes the old version rather than one that blames an individual file or the model. Remember that uploaded knowledge needs the same continuous care as a live source.
This apply-level skill feeds the decision in choosing connector access versus manual upload, where the maintenance burden of manual uploads is part of the trade-off, and it is a recurring item in a review cadence. It builds on the instructions-versus-knowledge distinction and pairs with connectors as authorized external reach.
A Project keeps citing last year's policy even though this year's version was uploaded. The knowledge base holds both versions, and both are accurate documents. What is the best fix?
People also ask
Why does a duplicate policy version cause wrong answers?
Do I need to remove old documents from a knowledge base?
Is uploaded knowledge set-and-forget?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.