Skip to content
Outrich Data

Data type

License your knowledge base and internal wiki for AI training

An internal wiki or help center your staff wrote can often be licensed for AI training. Outrich Data is a direct buyer and deals with the company that owns the content, after checking authorship and contributor terms. Names, credentials, customer details and pasted third-party text are removed first. You keep the knowledge base and approve each package in writing.

What counts as knowledge base

A knowledge base is the collection of articles your company uses to answer recurring questions. There are usually two. The external one is the help center customers search. The internal one is the wiki where staff record how systems are configured, how a known fault is fixed, which team owns what and why a decision was made.

Articles carry more than text. Page history, labels, parent and child structure, view counts, helpfulness votes and links to the tickets an article resolved are all part of the record. Community forum posts written by your customers, vendor documentation mirrored into the wiki and client workspaces you administer are not your content and are excluded.

Why AI labs want it

Public help centers are already in most training sets. Internal wikis are not. They hold troubleshooting notes written by engineers for other engineers, architecture decisions with the rejected options listed, and the unofficial fix that never reached the public docs. AI labs want that depth for assistants meant to answer questions inside a company.

A knowledge base is also a natural test bed for retrieval. Each article answers a known set of questions, and many are linked to the tickets or searches that led to them. Labs use those pairs to measure whether a model finds the right page and answers from it without inventing details.

Removed before licensing

What comes out first.

  • Author names, @mentions, avatars and page-watcher lists
  • Passwords, API keys, tokens and connection strings pasted into how-to pages
  • Internal hostnames, IP ranges and network diagrams
  • Customer names and account details used as worked examples
  • Vendor documentation and licensed articles copied into the wiki
  • Pages from HR, legal and security spaces, excluded in full
  • Embedded files, images and screenshots that show real records

Data title search

What gets checked.

  • Who wrote it. Staff-written articles are normally company property. Pages written by contractors, agencies or implementation partners depend on those agreements.
  • Community content. Forum answers and comments from customers are governed by your community terms, which may not grant a right to license them onward.
  • Third-party material. Vendor manuals, analyst reports and standards pasted into pages belong to their publishers.
  • Customer confidentiality. Internal articles written about one customer's setup can fall under that customer's NDA.
  • Employee data. Page histories tie edits to named staff. Check your employee notices, and for EU and UK staff the GDPR position, with counsel.

Who signsThe head of support or head of engineering usually sponsors it, depending on which knowledge base is involved. The general counsel approves the license and the wiki administrator handles space selection and export.

Value

What moves the price of knowledge base.

FactorEffect on value
Internal depthStaff-only troubleshooting and design pages are generally worth more than public help articles a model may already have seen.
Question and answer linksArticles tied to the tickets, searches or chats they resolved are useful for retrieval evaluation.
UpkeepA wiki with review dates, owners and few stale pages needs less filtering than one that was abandoned.
Page historyEdit history shows how an answer was corrected over time, which adds training signal.
StructureConsistent templates, labels and page hierarchy make the content easier to segment and describe.

Questions

Licensing knowledge base: what companies ask.

Our help center is public already. Would anyone pay for it?
Possibly, but less than for internal content. A public help center may already sit in training sets collected from the web, although a license gives a buyer clear permission that scraping does not. The internal wiki, page history and ticket links are usually the stronger part of an offer.
The wiki is a mess of stale and duplicate pages. Does that rule it out?
No. Stale pages are common and can be filtered by last-edit date, owner and view count. Duplicates are removed during preparation. You can gauge the size of the problem yourself with the free duplicate remover and dataset profiler on this site, which run in your browser.
Engineers paste credentials into wiki pages. How is that caught?
Secrets are a must-remove category. Pages are scanned for keys, tokens, passwords and connection strings, and matches are redacted before a dataset is built. Any credential found should also be rotated on your side.
Can we leave out whole spaces?
Yes. Selection works at the space, folder or label level. HR, legal, security and executive spaces are excluded by default, and you can add any others. The final inclusion list is approved by you in writing.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc