Knowledge System
An internal knowledge assistant that answers employees' questions from a company's documents, cites its sources and says when it doesn't know.
Why Knowledge System?
A knowledge assistant that sounds sure while being wrong is worse than none. So the design starts from trust: every answer is cited, access is enforced inside the database search rather than filtered afterwards, the model never writes database queries (it fills in a small query form that the server checks and compiles), and a trace shows each search, source, delay and cost. Search combines meaning and keywords, because meaning alone can't tell one year's report from another's.
What is this?
A chat assistant for the question every large organisation has: where do I find this, and what is the rule on that? It answers from the organisation's own documents with numbered citations, and a cited PDF opens at the exact page. The demo runs over more than 2,100 public documents of a bank and its regulator, split across 17 departments.
When it can't answer, it says so. "I don't know" is a state of its own: what it searched, the closest sources, and which department would know. Each one is logged to a gaps view of what the company never wrote down.
Who is it for?
Organisations with a large pile of internal documents, policies or procedures and staff who spend time hunting for answers.