Prepare company documents for an AI knowledge base
Prepare company documents for reliable AI answers. Resolve conflicting versions, check PDFs and tables, preserve permissions, and test updates before rollout.
To prepare company documents for an AI knowledge base, start with the questions people need answered. Select the approved sources, resolve conflicting versions, inspect the extracted text and preserve who may read each part. Then test the answers and the effect of changing or removing a source.
The difficult case is a perfectly readable document that should not govern the answer. Imagine three hotel-expense policies: an approved $180 limit, a newer $220 draft and a current $240 regional exception. A successful upload could preserve every word and still produce the wrong recommendation.
Those figures are hypothetical. The preparation problem is real: someone must decide which rule applies, to whom and when. An AI knowledge base cannot settle an unresolved company policy merely by finding the newest file.
This guide focuses on preparing evidence that a system retrieves before producing an answer, often called retrieval-augmented generation, or RAG. It does not assume that uploading documents retrains the model. The worked policy and proposed tests are examples you can adapt, not results from a customer deployment.
Start with the answers people need
Choose a bounded task, such as helping employees find the applicable travel policy. Write down actual questions and have the policy owner identify the evidence needed to answer them. Include a question that requires an exception, one that needs clarification and one the available documents cannot answer.
For the hotel example, ask for the destination and travel date before selecting a limit. An answer about a particular trip needs the applicable policy, currency, effective period and any approval requirement. A general summary of the travel handbook is not enough.
Use that question set to decide what enters the first AI knowledge base. A folder may contain relevant records alongside obsolete drafts, personal notes and documents for a different business unit. Connecting the folder is not an editorial decision about every file inside it.
If a short, well-maintained reference page and ordinary search resolve the problem, start there. You do not need a language model to justify cleaning up a confusing policy library. The same preparation remains useful if a broader Business Brain implementation follows later.
Decide which source is allowed to settle a conflict
Create a source register before conversion. A spreadsheet is sufficient for a small pilot. Record each original location and stable ID, its business owner, approval status, effective period, scope and allowed audience. Add the ingestion decision and its reason so the next person does not have to reconstruct your judgment.
The hypothetical travel register would distinguish these files:
| Document | Authority and scope | Preparation decision |
|---|---|---|
| Approved domestic hotel policy, $180 | Current for the stated domestic trips. | Include with its effective date, owner and source link. |
| Newer draft, $220 | Proposed change, not approved. | Keep outside the current-answer collection; retain as a draft where appropriate. |
| Approved regional exception, $240 | Current only for the named region and conditions. | Include with explicit scope and a link to the general policy. |
| Previous approved policy | Relevant to historical trips, not necessarily new ones. | Separate historical use from current-policy answers. |
Do not collapse the regional exception into the general policy because the wording looks similar. Similarity tools can propose duplicate candidates; an owner must decide whether they are duplicates, amendments or legitimately different rules.
Atlan’s preparation guide includes source ownership, version selection and approval as governance work. Apply those ideas with a business rule for authority, not a blanket newest-file-wins rule. A timestamp tells you that something changed, not that the change was approved.
Removing superseded material from an AI knowledge base is also different from deleting business records. Preserve the originals according to your organization’s retention decisions. If historical questions matter, retain effective periods and require the assistant to distinguish past policy from current guidance.
Inspect what the system actually extracts
Keep an untouched original and a prepared derivative linked to it. Run a small, representative sample through the intended extraction path before converting the entire library. Read the extracted output beside the original, including a difficult table and a document with important footnotes.
These checks target different ways meaning can be lost:
| Input | What to inspect in the extracted result |
|---|---|
| Scanned PDF | OCR accuracy, reading order, minus signs, decimal points and words such as “not.” A searchable file can still contain recognition errors. |
| Word, HTML or Markdown | Heading hierarchy, procedure sequence, links and exceptions separated by page breaks or layout. |
| Spreadsheet or PDF table | The relationship between each value and its row, column, units, date and footnote. Preserve formulas or their business meaning where the answer depends on them. |
| Diagram or screenshot | Whether the instructions depend on visual information missing from extracted text. Add a reviewed description or use a supported visual-processing route. |
| Recorded explanation | Transcript accuracy, speaker context and references to material shown on screen. Captions alone may omit the decisive evidence. |
AWS’s RAG writing guidance recommends clearer structure and explicit treatment of tables, images and self-contained topics. The useful principle is to make necessary context available with the answer-bearing text. Do not simplify away a qualification just to make the document shorter.
For the hotel table, an extracted “240” is unusable without the region, currency, per-night unit and conditions. Check an actual question that needs those fields. A visual preview of the original PDF cannot show whether the AI knowledge base received them correctly.
Markdown is an option, not a guarantee. Microsoft’s Markdown indexing documentation says its parser splits on headings and passes tables and other non-heading elements as text. That behavior is specific to the documented parser, but it illustrates why you must inspect the output of your chosen system.
Keep original-language text where translation could change a policy term. If you add a translation, label it, link it to the source and have an appropriate reviewer check important terminology. Do not silently replace an approved policy with an unchecked AI translation.
Give each retrieved passage enough context
Many retrieval systems divide documents into smaller passages. This lets an AI knowledge base retrieve relevant material without loading every source into a single request. It also creates a new failure point: the answer-bearing passage may become separated from the condition that makes it true.
An illustrative prepared record for the regional exception could look like this:
Source ID: travel-policy-regional-04
Status: approved
Applies to: Region R, eligible business trips
Effective period: 2026-01-01 through 2026-12-31
Owner: travel operations
Rule: hotel room reimbursement cap is USD 240 per night before taxes
Conditions: Region R trips approved by travel operations; meals excluded
Source location: travel-policy-regional-04, section 2
Access reference: source-controlled travel-policy-readers groupThese dates, conditions and identifiers are invented to make the example concrete. This is not a schema accepted by every product. In a real record, retain a working link to the governing source revision. The access reference must connect to enforced authorization; writing it into a text block does not secure the content.
Prefer meaningful sections with descriptive headings over arbitrary cuts that separate a rule from its exception. Where a procedure spans several passages, retain ordering and the relationship between them. Test short factual questions as well as questions that need several sections.
There is no universal chunk size to prescribe here. Your documents, extractor, retrieval method and model limits affect the choice. Compare candidate settings against the same reviewed questions, including missing and conflicting evidence. Do not use a larger overlap simply to conceal a document whose rules remain ambiguous.
AI can propose headings, a glossary or a short summary, but the result needs review. A generated FAQ that invents a refund condition creates a new unsupported source. Keep such additions out of the approved AI knowledge base until the owner has checked them against the governing material.
Preserve permissions through every derived copy
Classify sensitive material before ingestion and confirm whether the chosen service is approved to process it. Exclude secrets and unrelated personal information from the pilot. The business owner decides whether a document is useful; the appropriate data or security owner decides whether this use and audience are allowed.
An AI knowledge base needs more than a “confidential” label. The identity requesting an answer must be checked against the content it is allowed to retrieve. Derived passages, summaries and cached answers must not quietly widen the source’s audience.
Microsoft’s document-level access guidance distinguishes metadata from query-time enforcement and explains that indexed permissions need synchronization. Its chunked-index guidance also requires permission information on derived records. Several native permission features described there are in preview; this is not a claim that every connector provides the same protection.
Ask the system owner to show where authorization is enforced and what happens when permission metadata is absent or stale. Agree a safe failure behavior before importing restricted material. An instruction telling the model to keep secrets cannot replace that control. If the service cannot maintain the required separation, reduce the pilot’s scope or choose another approach.
For the wider connection decision, see the Slack, Google Drive and CRM guide. Document preparation cannot compensate for a connector that reads as the wrong person.
Test the evidence, the answer and the audience
Run the approved question set as an ordinary intended user, not only as an administrator. Record the source passages retrieved and compare the resulting answer with the policy owner’s expected result. Where logging itself contains sensitive information, limit its access and retention.
The following is a proposed test set for the hypothetical policy, not a report of tests run against a customer system:
| Test | Expected behavior |
|---|---|
| Current domestic trip | Uses the approved $180 policy with its relevant conditions and citation, not the newer draft. |
| Qualifying Region R trip | Uses the $240 exception and states the scope that makes it applicable. |
| Destination omitted | Requests the missing detail instead of choosing a policy silently. |
| Question outside approved evidence | Says the evidence is insufficient and points to the agreed owner or process. |
| User without source access | Does not disclose restricted contents or revealing source metadata. |
| Access revoked after an earlier answer | New retrieval and reused context follow the agreed revocation behavior; the observed delay is recorded. |
| Policy corrected or removed | Old answer-bearing fragments stop supplying current guidance after the documented update process. |
A citation is necessary evidence to inspect, not an automatic pass. Follow it and confirm that the cited version actually supports the answer. A correct number from the wrong regional policy is still a failed answer.
For revocation, measure from the source-permission change to the point when the system stops retrieving or reusing the restricted material. Test both a fresh question and an existing conversation. Revoking access cannot erase information someone already viewed, so distinguish future access prevention from deletion of retained system copies.
When the AI knowledge base fails a test, keep that question and locate the first incorrect step. If extraction lost a footnote, repair extraction. If retrieval selected a draft, fix source selection or version handling. If the correct evidence was retrieved but the response changed its meaning, investigate answer generation and its instructions. Rewriting the prompt for every failure can hide the underlying defect.
Have the business owner approve the corrected examples, then rerun the previously passing cases. A small sample cannot prove that all future questions are safe, but it gives you a repeatable release check and a way to diagnose regressions.
Make updates and removals part of preparation
Assign an owner for each source and someone responsible for the ingestion process. Define which changes require review, how quickly important changes must reach answers and what to do when a refresh fails. Match the review interval to the material: a stable glossary and a frequently changing operational policy do not need the same schedule.
Keep a mapping from the original document and revision to the derived records in the AI knowledge base. Track the last successful processing time separately from the source’s modification date. A changed file is not evidence that its indexed representation changed.
Test removal, not just replacement. Microsoft’s Markdown documentation warns that its one-to-many indexing mode can retain obsolete sections when a file is shortened. Other systems have their own behavior. Verify the deletion path in yours, including caches and retained conversation context, without confusing removal from an active index with disposal of the original record.
Agree a safe response when a time-sensitive source is unavailable or overdue for review. For some questions, the correct result is a handoff rather than an answer based on material nobody can confirm. Include that operational work in the implementation budget.
Before importing the next folder, complete one source register entry and one end-to-end question test. The evidence should show which approved document supplied the answer, which conditions survived extraction, who could see it and how a subsequent correction reached the result. Use the gaps in that record to choose the next preparation task.
Questions people ask
What is the best file format for an AI knowledge base?
Use a format your chosen system can extract accurately. Clean Markdown or HTML can preserve useful headings, but PDF, Word and spreadsheet inputs may also work. Inspect the extracted content, including tables, footnotes and image-dependent instructions. A supported file extension does not prove that the meaning survived. Keep the original alongside a traceable prepared copy.
Should I merge all company documents into one file?
Usually, preserve separate documents with their owners, source links, versions and permissions. Merging unrelated material can obscure those boundaries. If a system requires a combined input, retain explicit section-level provenance and access controls, and test whether it can still return the right source and exception. Do not use historical forum upload limits as current product requirements.
Can AI clean and summarize documents for me?
It can help propose headings, summaries and duplicate candidates. Treat those outputs as drafts. A knowledgeable owner must check numbers, conditions, exclusions and source links before the derived content becomes approved evidence. A fluent summary is not a substitute for a missing company decision.
How much content does an AI knowledge base need?
Enough approved evidence to answer a defined set of useful questions, including their important exceptions. There is no universal document count or word target. Start with a bounded topic, identify unanswered questions, and add reviewed sources where the tests reveal gaps. Stored corpus size and the context used for an individual answer are different measurements.
Does updating a source automatically update the AI answer?
Not necessarily. The change may need to reach extracted text, indexed fragments, caches and conversation context. Test replacements and removals separately, then measure when answers stop using the old material. Source-permission changes need their own propagation test. The behavior depends on the product and configuration.