Our approach to knowledge management

Most of our articles describe what Desk does for a team without saying how. This one is the how. If you have wondered whether Desk is "RAG" or "a knowledge graph", or what happens to a file between upload and answer, this is the place.

The problem in one sentence

Documents are written for people to read one at a time. Business questions are asked across many of them at once. Something has to close that gap, and there are two ways to do it.

Layer one: search over the text

The first thing Desk does with a file is keep its text and make it searchable, by exact words and by meaning. "By meaning" uses embeddings: each passage is turned into a numeric fingerprint so that a question phrased differently from the document still finds the right paragraph. This is the retrieval that document chat tools are built on, and it is good at what it is good at: quoting what a specific document says. Desk uses it for exactly that.

Layer two: records and the links between them

The second thing Desk does is read the file. A model goes through the text and writes down what it mentions: the people, organizations, places, pieces of work, things, agreements, events, issues, commitments, and topics. Those ten kinds are the same in every Desk knowledge base, in every industry, and that is deliberate. A clinical study, a consulting engagement, and an acquisition are all pieces of work with a start, an end, parties, and outcomes. The industry's own word for the thing, "study" or "engagement" or "deal", is kept on the record so nothing is lost, but the shape underneath is shared.

Each record gets its details, its links to other records, with the document's own wording kept on the link, and its facts. A fact is a specific statement, such as an enrollment target or a payment term, with the exact quote and file it came from. When a second file mentions the same person or study, it adds to the existing record rather than creating another, so a record accumulates sources over time.

In technical terms this is a knowledge graph: typed nodes, typed edges, attributes, provenance on everything. In practice it is the reason Desk can count, compare, and follow connections across hundreds of files, which search alone cannot.

Why the structure is fixed

Knowledge graph projects usually begin with months of schema design and end with a schema nobody maintains. We removed that step. Because the ten kinds cover any business, there is nothing to set up before the first upload and nothing to change when a new kind of document arrives. Files are read once and never re-read because of a structure change, since the structure does not change. Specifics live in facts, which need no schema at all.

What happens when a question is asked

The model answering the question does not receive your files. It receives tools: search the text, look up a record, list records of a kind with a filter, follow a record's links, read a stretch of a document. It uses them, usually several times, and every record and file it touches becomes a citation on the answer. It is instructed to state only what the files state, to keep their wording for reasons and decisions, to label anything beyond that as inference, and to say plainly when the files are silent. Two people asking the same question get the same answer because it comes from the same records.

Two small things that matter a great deal

Versions of a document, such as an SOP at v2.1 and v3.0, are recognised as versions rather than duplicates. They are linked oldest to newest, the newest is marked current, and an unversioned mention resolves to it, so an answer about "SOP 014" quotes the current version and says so.

Records that are the same thing under two spellings are found and offered for a merge, with the reason shown. Names that differ only by a number, such as two protocol codes or two audits in different years, are never confused for one another.

Where this does not matter

For a single document, or one person's small working set, plain search gives the same answer Desk would. The structure earns its keep as the library grows and questions begin with "which of our", "how many", "compare", or "who".