Feature

A knowledge base with an editorial cycle, not a folder of files

Dropping a folder of documents into a model gives you an agent that quotes last year's price list, because nobody marked it void. Here content has a state, a version and a person who approved it.

Four states, one direction

The agent sees only published versions. A draft in the editor has no path into an answer.

StateMeaningVisible to the agent
DraftWork in progress, visible to editorsNo
ReviewedRead by a second person, awaiting publicationNo
PublishedApproved for use in conversationsYes
ArchivedWithdrawn, kept for the recordNo

Versioning, or the answer to a hard question

What the agent knew that day
Every publication creates a version. When a customer returns a month later disputing an answer, you see exactly the text in force then, not today's.
Who approved it
Each transition between states is recorded together with the person who made it.
Rollback
Undoing a change does not require reconstructing the text from memory.

Where content comes from

The panel editor

Write or paste content directly. The simplest route for procedures that so far lived in people's heads.

Import from a URL

You give a page address and the system pulls the content. The import is guarded against reaching internal network addresses.

The gap register

Questions the agent could not answer come back as a list of things to write. That closes the loop between conversation and editing.

File upload

PDF, DOCX, TXT and Markdown straight from disk. They enter the same editorial cycle as articles written in the panel, so a draft pulled from a file does not answer a customer either.

How the agent finds the right passage

Meaning based matching alone loses catalogue numbers and proper names. Word based matching alone loses questions asked in different words than the document uses. We run both.

Two paths, one result
Semantic search and keyword search run in parallel, and the results are merged by a single rank fusion method.
Scope inside the query
The limit to a single organisation is a condition of the query itself, not a filter over results. Another tenant's passage has no way into the context.
Citations attached after generation
Naming the sources is not a request made to the model but a separate step on its output side.

Your knowledge already exists, it is just scattered

We will show how much of it the agent handles without writing anything from scratch.