Why the documents decide the outcome
When a knowledge assistant gives a poor answer, the model is rarely to blame. Far more often it found an outdated policy, two procedures that disagree, a scanned page it could not read or nothing at all because the answer lives in someone's head. The assistant is a mirror held up to your documentation, and many organizations are surprised by what they see.
That is useful news, because documents are something a business team can fix without any technology. The steps below can begin before a supplier is chosen, and they will improve daily work even if the assistant is never built.
Take an inventory of where knowledge lives
| Source | Typical content | Usual condition |
|---|---|---|
| Policy and procedure library | Approved policies, standard procedures, forms | Good structure, some outdated items |
| Shared drives | Everything, in every version | Duplicates, drafts and personal copies mixed with finals |
| Intranet or wiki | How-to pages, announcements, team spaces | Uneven. Some pages current, many abandoned |
| Help desk or ticket system | Resolved cases and their answers | Rich in real questions, but with personal data and old workarounds |
| Email and chat | Decisions and explanations | Unstructured and private. Rarely suitable as a source |
| Product and technical documentation | Manuals, specifications, release notes | Usually reliable and well versioned |
| People | What was never written down | Not available to any assistant until someone writes it |
For each source, note roughly how much there is, who looks after it, what formats it uses and how access is controlled. A one-page table is enough. It will show at once which sources are ready and which need work.
Choose a first set by the questions people ask
Do not start by loading everything. Start from one group of users and the questions they really ask, then find the documents that answer those questions. If the experts in that area were asked where to look, which documents would they name? That is the first set.
Good first sources
- Documents experts already send to colleagues who ask
- Approved, dated and owned
- Written as text, with headings
- Changed through a known process
- Open to the whole user group
Sources to leave for later
- Shared drives with years of drafts
- Chat and email archives
- Scanned paper with handwriting
- Material with mixed confidentiality
- Anything nobody can vouch for
Give each document an owner and one current version
- 01
Name an owner
A person or role responsible for keeping the document correct. A document without an owner should not be a source.
- 02
Keep one current version in scope
Archive superseded versions somewhere the assistant does not read. Old versions are the leading cause of confidently wrong answers.
- 03
Settle contradictions
Where two documents disagree, decide which is authoritative for that topic and correct or retire the other.
- 04
Date everything
An effective date and a review date on each document let the assistant, and its users, judge how fresh an answer is.
- 05
Write down what is missing
Questions with no written answer need one. Have the expert write a short page, starting with the most frequent questions.
Make the content readable by a machine
An assistant works by splitting documents into passages and finding the ones that match a question. Structure helps it split in sensible places and understand what each passage is about.
- Real text, not pictures of text. Scanned pages must be converted, and the result checked. Text inside images and diagrams is usually invisible unless described in words.
- Headings that mean something. A clear hierarchy of headings tells the system where a topic begins and ends. A section titled Annual leave is found. One titled Section 4 is not.
- Self-contained passages. A paragraph that says as above or see the table opposite loses its meaning when read alone. Repeat the subject.
- Tables with care. Simple tables with a header row survive. Merged cells and tables used for layout do not. Consider restating key tables as sentences.
- Defined terms and abbreviations. Spell them out at least once per document, and keep a glossary the assistant can also read.
- Useful titles and file names. A document called final v3 new tells neither people nor machines what it contains.
Record who may see what
If the assistant can read a document, it can quote it to whoever asks, unless access is enforced. Before any confidential source is included, its access rules must be clear in the system where it lives, because the assistant should take them from there and apply them to each user at the moment of the question.
Check three things for each source. Are permissions set on purpose, or has everything simply been left open? Do any documents mix content for different audiences, such as a staff handbook with a management-only annex? Does the source contain personal data that should not be searchable at all? Fixing loose permissions is worthwhile regardless, since an assistant only makes existing oversharing easier to discover.
A worked example
Preparing HR documents at a logistics company
This example is invented to show the amount and kind of work involved. A logistics company wants an assistant that answers staff questions about HR policies.
- InventoryPolicies sit in three places: an HR folder with 212 files, the intranet and attachments in old emails. The HR team identifies 31 documents it would actually point a member of staff to.
- VersionsOf those 31, 9 exist in more than one version. The leave policy has four. One current version of each is placed in a new folder, with an effective date and an owner.
- GapsA month of questions to HR is reviewed. A fifth of them, mostly about shift swaps and uniform allowances, have no written answer. Two short pages are written.
- FormatSix policies are scans of signed paper. They are replaced by the original text files. The pay scale table, full of merged cells, is rebuilt as a plain table.
- AccessThe disciplinary procedure has a manager-only annex in the same file. It is split into two documents with different permissions.
The assistant starts with 33 documents instead of 212, and every one has an owner. HR staff report that the tidy folder alone has reduced the questions they receive from managers.
Set up the routine that keeps it current
- 01
How does a new or changed document reach the assistant?
Why it mattersIdeally by being saved in the right place, with the index updated automatically. A manual upload step will be forgotten.
- 02
How is a withdrawn document removed?
Why it mattersDeleting or archiving at the source should remove it from answers promptly.
- 03
Who reviews documents that reach their review date?
Why it mattersThe owner, prompted by a reminder. Stale content should be flagged or excluded.
- 04
Who sees the questions that went unanswered?
Why it mattersA regular look at failed or low-rated answers shows what to write or fix next.
- 05
Who decides when a new source is added?
Why it mattersOne person, applying the same tests as for the first set.
Document preparation checklist
Document preparation checklist
0 of 11 checkedWhat to do next
Ask the experts in one department to list the twenty documents they send people most often, and check each for an owner, a date and a single version. That afternoon's work is the foundation of any assistant, and our RAG implementation checklist covers the stages that follow.
Our enterprise AI assistant development page describes how we build on prepared content. Tell us which documents you would start with and we will say what we would check.
Common questions
How many documents does a knowledge assistant need?
Fewer than most teams expect. A few dozen well-kept documents that answer the common questions will outperform thousands of mixed quality, because every weak document is a chance for a wrong answer. Start with the smallest set that covers one user group's real questions, measure, and add sources where the unanswered questions point.
Do we have to rewrite our documents?
Mostly no. The main work is selecting, removing old versions, assigning owners and fixing a small number of documents with format problems. Rewriting is needed where knowledge was never written down, where passages only make sense in context or where key information is locked in complex tables or images.
Can the assistant read scanned documents and PDFs?
Text-based PDFs are generally read well. Scans are images and must be converted to text first, with accuracy that depends on print quality, layout and language, and handwriting is unreliable. Where the original text file exists, use it instead. Always have a person check converted documents that matter.
What should we do about confidential documents?
Decide per source whether it belongs in the assistant at all. If it does, make sure permissions are correct where the document lives and that the assistant enforces them for each user at the time of asking. Split documents that mix audiences. Have someone without access test that restricted content cannot be reached.
Who should be responsible for the content afterwards?
Each document keeps its own owner, and one coordinator looks after the whole: approving new sources, watching review dates and reading the questions the assistant could not answer. In a small organization that is a part-time role. Without it, answer quality declines as documents age, and people stop trusting the assistant.