A chat-with-your-documents tool has read nothing until you ask. Desk has already read everything. That one sentence is most of the difference, and the rest of this article explains what it buys you and when it does not matter.
The everyday version of it is the difference between searching a chat history and opening a filed archive. Chat search finds messages containing words like the ones you typed, which is exactly what you want when you half remember what was said. Ask it how many clients agreed to a particular term last year and it hands back a few plausible messages and no answer, because nothing in a chat log knows that four threads are about one client. An archive was organized on the way in, so the same question is a lookup.
How a document chat tool answers
Upload files, and the tool cuts them into passages and indexes them. Ask a question, and it finds the passages that most resemble the question, hands the top handful to a language model, and the model writes an answer from them. This is called retrieval-augmented generation, and it works well for a certain kind of question: "what does this document say about termination notice?" The answer is in one place, the passage that contains it ranks first, and the model summarizes it.
Where it runs out
Business questions are rarely about one passage.
- "Which of our sites has run a phase 2 study with more than 20 enrolled?" needs every study, not the ten passages most similar to that sentence. The tool cannot know whether there were forty relevant passages and it saw eight, and it cannot compare a number in one file with a number in another.
- "What has this sponsor done with us?" needs the sponsor's mentions across proposals, contracts, reports, and emails, and it needs to know that "Helix", "Helix Therapeutics", and "HTX" are one company. A passage search hopes the answer is in a single paragraph.
- "Who here has worked with Dr. Chen?" needs to know that "Dr. Chen" in a CV and "Marcus Chen" in a monitoring report are the same person. Every question starts from zero, so nothing carries over.
The tell is in the answers. They read fluently, they cite something, and they quietly answer a smaller question than the one you asked.
What Desk does instead
Desk reads each file once, when it arrives, and turns it into records: the people, companies, projects, agreements, events, and decisions it mentions. Each record has fields, links to other records, and facts with the exact quote and file they came from. Marcus Chen is one record with four source files. A study is one record with its sponsor, site, investigator, enrollment, and phase, each traceable.
When you ask a question, Desk uses both layers. It searches passages by keyword and by meaning, exactly as a document chat tool does, so quoting what a file says stays fast and precise. And it filters, counts, and follows links across the records, so "which of our…" is answered over the whole knowledge base rather than a sample. Every claim in the answer links to the record or file behind it, and when the files are silent, Desk says so.
Three things this buys you
- Complete answers, not a sample. Counts, comparisons, and "which of our" questions run over everything you have uploaded.
- Connected knowledge. The same person, company, or project across a dozen files is one record. Following the links is how "what have we done with them" gets answered.
- Verifiable answers. Facts carry their quote and file. Records show their sources. You can check, and so can the person you forward it to.
When the difference does not matter
If the whole library is one contract, or one person's working notes, a document chat tool gives the same answer Desk would, and faster to set up, because Desk spends a minute or two reading each file on upload. The gap opens as the library grows and questions start spanning files. If your question has "which", "how many", "compare", or "who" in it, you are asking a question the records were built for.
The other difference is what survives. A document chat tool assembles an answer and discards the work that produced it. Desk keeps what it read, so every later question starts from everything the earlier ones established.
How Desk helps
Desk searches like a document chat tool and also knows what it has read.
- Passages are searchable by keyword and by meaning, so quoting a file stays fast and exact.
- Every file is also read into records and relationships once, so counting, filtering, and comparing across the whole library is a query rather than a sample.
- Every claim links to the record or file behind it, and Desk says plainly when the files are silent.
Thinking about how this would work for your team?
Let’s chatKeep reading
"Have we done this before?" is the most expensive question in your company
It gets asked a dozen times a week, it is almost always answerable from files you already have, and the answer usually arrives too late to matter.
Read more →Tribal knowledge: what it costs and how to capture it without a reorg
The knowledge that walks out the door is rarely written nowhere. It is written somewhere unfindable. The fix is not asking people to document more.
Read more →