Sources and lessons
Knowledge is broken down into two layers. The source captures where the information came from: a closed ticket, a training session, a manually entered text or a downloaded document (text formats: TXT, MD, CSV, JSON, YAML, XML, HTML, SQL, LOG). The lesson formulates the knowledge itself and is obliged to refer to the source: material without origin does not enter the database. You cannot learn on an open ticket, only on a closed one. Lessons are versified. Editing creates a new version, not editing an existing one: the active version continues to respond until a new one is approved. The lesson may have an expiration date, the expired lessons are turned off themselves.Adoption
Approving a version of a lesson is the only action in the entire module that creates active knowledge, and the owner and administrator of the AI have the right to do it; operators and the knowledge inspector cannot approve it. Before approval, conflicts with existing lessons on applicability are tested, the conflicting version cannot be silently confirmed: the conflict must be explicitly acknowledged, so the base does not gain contradictions unnoticed.Review and quarantine
There are two ways to respond to erroneous knowledge: shutdown or quarantine. Quarantine eliminates the lesson from the answers immediately and triggers an impact analysis: what else is built on the same material. Influence analysis returns two lists and never mixes them: direct dependencies: materials that are provably used by a source, with a suggestion of a specific action; similar materials: just textual proximity, for which only human verification is offered, because similarity does not prove infection; for already sent answers, actions are not proposed at all: they cannot be returned. Removing the source content (on request from the client or data policy) empties the text but leaves a record: lessons continue to show their origin.Proposals by operators
The operator from the ticket can click “Proposal to study.” It creates nothing in the database: the sentence enters the “Proposal by operators” list in the “Training” section, where the AI administrator accepts it or rejects it. The accepted sentence becomes material for the lesson in the usual way, through a draft and approval.Playground training
The Learning section is designed as a messenger with an AI support, and you create a named session and you talk in two modes that you switch right in the composer. Customer mode: you write on behalf of the client, a real conveyor responds to validated knowledge, like a real ticket, you can put screenshots (up to four, compressed in place) and see how the AI recognizes them. The answers come with type (answer, refinement, transfer to the operator), confidence and sources. Learning mode: you write like a curator: rules, instructions, debriefings. You can mark the line with the right example or the wrong example. AI confirms understanding, but nothing from the session enters the knowledge base automatically. The “Analysis” button runs through the session summarizing and evaluating the mood with the risk of leaving: it is convenient to check how these models see the dialogue. The “Leave in a lesson” button distills the session into a draft lesson; then it goes through the usual way of checking and approval. At the end of the session, the panel will separately ask whether to prepare a lesson.Knowledge gaps
The questions that the database didn’t answer are piled up in analytics as knowledge gaps grouped by topic. This is an AI administrator worklist: each topic is a candidate for a new lesson. The closed gap is marked manually. Clustering is rough and points to a topic rather than pretending to be accurate. Roles and rights to act with knowledge are discussed on the page Operators and roles: operators have only a view and offer, the knowledge inspector can quarantine, says AI administrator
