Knowledge sources

What is learned — and what is looked up.

replytor does not have one knowledge base but several sources. The distinction that matters is not between systems; it is between knowledge that holds and facts that change by the hour.

The one distinction everything rests on

	Indexed	Looked up live
Content	Rarely changes, holds generally	Changes constantly, holds for one case
Example	Manual, return policy, resolved case	Stock, order, ticket status
Route	Made searchable in advance	Fetched at the moment of answering
Citation	Reference to the passage	The source's reply with a timestamp

Four kinds of source

Documents

PDF, Word, Excel, Markdown, HTML and plain text. Manuals, data sheets, price lists, policies. Scanned PDFs go through text recognition — anything not reliably readable is not indexed but reported.

Ticket and project systems

Jira, Jira Service Management, Confluence, Odoo, Zendesk, Freshdesk and their kind. Resolved cases and knowledge pages become searchable; the status of a single case is fetched live.

eBay inventory

Listing, stock, shipping and return terms, order status — through the eBay interfaces, live and read-only. The agent cannot change a listing or correct a stock level.

Your own databases

Read access to views you release, through queries defined in advance. The language model writes no database queries — it only picks which vetted one to use.

When two sources disagree

With several sources, disagreement becomes routine: the manual says fourteen days for returns, the marketplace terms say thirty. There is a fixed order of precedence — the live lookup for the specific case counts most, then your binding documents, then remaining documents, and last resolved cases.

If two sources on the same level disagree, the case is handed over and the contradiction reported to you. One of the two is wrong, and that is not a question for a machine.

Which sources do you have?

Usually more than you think — and most of them sit in systems, not in documents.

Let's talk it through