What is an AI knowledge assistant?

What it actually does

Strip away the branding and an AI knowledge assistant does three things. It keeps an index of your content up to date. When someone asks a question, it finds the passages most likely to answer it. Then it has a language model write an answer from those passages, with citations, or decline if they do not answer it.

The technique underneath is retrieval-augmented generation. What separates one assistant from another is less the model it uses than everything around the model: which sources it can read and keep in step, how good its search is, how it decides it does not know, and where people can reach it.

How it differs from a chatbot

The chatbots of the last decade were decision trees with a text box. Someone wrote intents ("track order", "reset password"), example phrasings and scripted replies. Anything outside the script got "Sorry, I didn't understand" or a menu.

A knowledge assistant needs no script for questions about information. If the answer is written down somewhere it can read, it can answer - in the user's words, combining two documents if needed. Scripts still have a place for actions such as raising a request or changing an account, which is why better assistants combine both: free answers from knowledge, controlled steps for actions, with a confirmation before anything changes.

How it differs from search and a knowledge base

A search box returns a list of documents and leaves the reading to you. A knowledge base is the collection of articles itself. An assistant sits on top of both: it does the reading, and returns the two sentences that answer the question with a link to the place they came from.

That matters most when the answer is buried - the fourth paragraph of a 20-page policy, or a table in a product guide. It matters least when your help centre already has one short article per question; there, an assistant mainly saves people from guessing the right search words.

• Search: finds documents. You read them.
• Knowledge base: holds the documents. You search them.
• Knowledge assistant: reads the documents for you and answers, with the source.

How it differs from ChatGPT and other general assistants

General assistants are superb at writing, summarising and reasoning. Given a pasted document, they will answer questions about it. But they are designed to answer everything, which is the opposite of what you want from an assistant speaking for your company.

A company knowledge assistant should refuse questions your content does not cover, never fill gaps from general knowledge, keep each audience to its own content, and be reachable by customers who will never have an account with an AI vendor. It also needs to be kept in step with your sources without anyone pasting documents into a chat.

Who uses one, and where

The same knowledge tends to be needed by several audiences, which is the strongest argument for one assistant rather than several disconnected bots.

• Customers - a chat on the website, help centre or inside the product, answering how-to questions at any hour.
• Support agents - a reply draft built from the same knowledge, inside the help desk, for a person to check and send.
• Employees - policy, process and IT questions asked in one place instead of an HR or IT inbox.
• Your own tools - the same search and answers through an API, for a portal, an app or an internal system.

What a good one needs

Most of the work in a knowledge assistant is unglamorous, and it is where cheap tools cut corners.

• Sources that stay in step: website and help-centre crawling, cloud drives and wikis, uploads, and an API - re-read on a schedule, fetching only what changed.
• Hybrid search: meaning and exact terms together, so error codes and product names are found.
• A notion of authority: official documents outrank old tickets and informal notes.
• A confidence gate that refuses rather than guesses, and citations that are checked, not just displayed.
• Separation: each audience or customer sees only the content meant for it.
• Feedback: ratings on answers, and a list of questions it could not answer, so someone can fix the documents.

Twelve questions to ask a vendor

Ask these with your own documents loaded, not in the vendor's demo workspace. The answers to 3, 4 and 9 will tell you most of what you need to know.

• 1. Which of our sources can you read today, and how do you keep them in step when they change?
• 2. Is search vector-only, or hybrid with keyword search?
• 3. What happens when the answer is not in our documents? Show me.
• 4. Are citations checked against the source, or only displayed?
• 5. Can we mark some sources as more authoritative than others?
• 6. How do we keep internal content out of customer-facing answers?
• 7. Can answers about one customer's data be restricted to that customer?
• 8. Where does it answer - website, app, help desk, internal chat, API?
• 9. Which AI models process our documents, and where do they run? Are our documents used for training?
• 10. If it can act in our systems, what stops it acting without confirmation, or for the wrong person?
• 11. How is it priced - per seat, per answer, per resolution? What happens when we run out?
• 12. What does it show us about questions it could not answer?

When you do not need one

If your volume of repeat questions is small, your help centre is short and well organised, and the people asking are happy to search, a good knowledge base may be enough. If your documents are mostly out of date, fix that first - an assistant will faithfully repeat whatever they say.

And if the questions you get are really requests for action - refunds, changes, approvals - you need a help desk with sensible workflows before you need an assistant.

Where SupaCortex fits

SupaCortex is an AI knowledge assistant built around one shared brain. It learns from uploads, website and help-centre crawls, Google Drive, SharePoint and Confluence, and reviewed answers mined from past tickets. It answers - with checked citations, or an honest refusal - on a website and in-app widget, as reply drafts in UnifyDesk or any help desk on its Draft API, in a team chat across all your knowledge, and through an API.

Knowledge is organised into projects, so you decide what each audience can ask about. It is priced by plan with monthly AI credits; website visitors and app users are never seats.