6,869 Truth and Reconciliation Commission records, catalogued and linked back to their source
Evidence Systems, our own project
Data System Build
AI retrieval is only as good as the documents behind it. An assistant that searches your own material gives weak or unsupported answers when files are duplicated, out of date, badly named or missing basic metadata.
Getting ready means deciding which questions the system must answer, cleaning and structuring the source set, adding the metadata retrieval depends on, and testing answers against known sources before anyone relies on them. These guides cover each step. We build data systems where every claim traces back to its source, and we can run them for you.
Start here
A step-by-step method for making a document library readable, structured and traceable before any AI tool searches it.
The metadata fields that let an AI knowledge base filter the right sources and trace every answer to a page or row.
A test plan for AI knowledge bases that checks sources, retrieval, citations and refusals before people rely on the answers.
How to tell whether a weak AI answer comes from the model or from old, duplicated, vague or conflicting source documents.
The dated record of South Africa's withdrawn draft AI policy, and the citation checks that stop invented references reaching a public draft.
Evidence Systems, our own project
Data System Build
Specialist drafting team, for COGTA
Data System Build
The working system, built and handed over, with the trail back to source intact.
New guides, tools and case studies from our team, about once a month. You can unsubscribe at any time.
Tell us what you are working with. We will tell you plainly whether we are the right fit.
Tell us what you are working with and what it has to answer. We will tell you plainly whether we are the right fit.
Choose a time (opens in a new tab) The call is on Google Meet. Bookings are handled by Google Calendar, which may set its own cookies.