Legacy wikis and help centres were written for people who already knew where to look. We audit, restructure, tag and embed them so an assistant can find the right page and cite it.
Teams whose knowledge base has grown faster than anyone has maintained it
A Confluence space with ten years of pages, a help centre with three generations of product names, a SharePoint library where the real answer is in a comment on a draft. These are not unusual; they are the norm.
Retrieval-augmented generation does not fix content problems, it amplifies them. If three pages describe the refund process and two are out of date, an assistant will cite one of the wrong ones with full confidence. Migration is the work of making the corpus trustworthy before the model reads it. Our guide to data readiness for AI sets this in the wider context.
Outcomes we design for
One current page per topic, with duplicates merged or retired
An owner and a review date on every page that survives
Access labels that an assistant can enforce
Text that was trapped in screenshots and PDFs made searchable
Chunks and embeddings produced from a clean, versioned source
A question set that proves retrieval works before launch
What changes
From a sprawling wiki to a corpus an assistant can trust
The audit decides what survives; retired pages go to an archive the assistant never reads. Survivors are restructured, tagged and only then chunked and embedded, so the index is built from a clean source rather than from whatever happened to be in the wiki.
How we work
Five stages, each with a deliverable your content owners sign off
Content audit
Export, cluster near-duplicates, find pages with no views or no owner, and score each page as keep, merge, rewrite or retire. Output: a decision sheet per page and a readiness score.
Structure
Split mixed pages into single topics, agree a title convention, define the page template. AI drafts the splits; owners approve. Output: the new page map.
Metadata
Owner, review date, audience, access label, product and version on every page. The schema is designed so retrieval can filter on it. Output: tagged corpus and the glossary.
Embeddings
Chunking that follows the new structure, an embedding model chosen for your language mix, and an index with metadata attached. See chunking for why structure matters here.
Validation
A question set written with owners retrieves the right page at an agreed hit rate. Numbers reported before and after, then handover with an ongoing review cadence.
Audit decision
Typical signal
What happens
Keep
Single topic, viewed recently, owner confirms it is current
Tagged and migrated as is
Merge
Two or more pages on the same topic with overlapping text
AI drafts a merged page; owner approves; old URLs redirect
Rewrite
Right topic, wrong details, or text only in screenshots
Owner rewrites from an AI draft with extracted text
Retire
No views, no owner, product or process no longer exists
Moved to the archive; never indexed; redirect to the replacement if one exists
Content readiness checklist
What a page needs before an assistant should be allowed to quote it
Use this on your own knowledge base today. Pages that pass can be indexed now; the rest tell you the size of the migration.
Ready to index
A named owner who can be asked "is this still true?"
A last-reviewed date within the agreed cycle
One topic per page, stated in the title
Headings that follow the page template
No near-duplicate elsewhere in the corpus
Text is text, not a screenshot or an image of a table
Internal terms and acronyms defined in the shared glossary
An access label: public, all staff, team, restricted
Product and version tagged where content depends on them
Retired versions archived, not left alongside the current one
Anti-patterns we see in nearly every audit
"Notes" pages that mix meeting minutes, decisions and how-to steps
Draft pages published next to the final version
Procedures that exist only as attached PDFs or slide decks
Pages whose only content is a link to another tool
Comments and page history holding the real answer
Permissions set by accident of who created the space
Product names from three rebrands ago used interchangeably
Tables pasted as images because the editor was awkward
What we deliver
A migrated corpus and the tooling to keep it that way
Audit report and decision sheet
Every page scored, clustered and assigned a decision, with a readiness score for the corpus and a cost and time estimate for the migration.
Content model and templates
Page types, title conventions, metadata schema and access labels, configured in your wiki or help-centre tool so new pages start out ready.
AI-assisted rewriting workflow
Merge and split drafts, extracted text from images and PDFs, and suggested titles, queued to owners for approval with a clear diff.
Glossary of internal terms
Acronyms, product names and their old names, defined once and used by the assistant to understand questions phrased in any of them.
Embedding pipeline
Structure-aware chunking, a model chosen with our embeddings strategy, and a re-embedding plan for when content or models change.
Validation set and review cadence
A labelled question set, before-and-after retrieval scores, and a quarterly review process so the corpus does not drift back to where it started.
The migrated corpus feeds directly into an enterprise RAG build, a support copilot or an internal copilot. Many clients run the audit first and decide which of those to build from what it reveals.
Not always, but you have to audit. If the audit shows that most pages have an owner, a single topic and current content, we index in place and fix the rest as we go. If it shows years of duplicates and abandoned pages, indexing them first produces an assistant that confidently quotes the wrong one. The audit takes one to two weeks and tells you which case you are in.
Can we stay on our current wiki or help-centre tool?
Usually, yes. Migration here means restructuring and enriching the content, not necessarily changing the platform. We migrate platforms only when the current one cannot hold the metadata or permissions the AI layer needs, or when you wanted to move anyway.
How much of the rewriting can AI do?
More than you might expect, less than you might hope. Models are good at splitting mixed pages into single-topic pages, drafting summaries, suggesting titles and spotting near-duplicates. Deciding what is still true, who owns it and what to retire needs a person. We pair the two: AI proposes, the content owner approves.
What does 'AI-ready' mean in measurable terms?
Every page has an owner, a review date, an access label and one clear topic; near-duplicates are merged; text that lived in screenshots is text again; internal terms are defined in a glossary; and a labelled question set retrieves the right page at an agreed hit rate. We report these numbers before and after.
How long does a knowledge-base migration take?
Audit in one to two weeks; the migration itself typically runs 4–10 weeks depending on page count and how many owners need to sign off. Embedding and validation are a small part of it. Ask for an audit and we will size it from your export.
Talk to us about your AI project
Tell us what you are trying to do, which systems are involved and what is in the way. We reply within one business day with an opinion and a suggested next step.