Searching your enterprise documents, and finding the answer
Finding them is the hard part, and so is trusting what you read. An answer that cites no source is worth no more than a rumour.
In short
You ask a question in plain language, as you would ask a colleague. The platform searches your document base and the knowledge graph, then writes an answer in which every claim points back to the document that justifies it. Permission filtering applies during the search, so nobody reads through it what they are not allowed to read.
What does looking things up really cost?
In most companies, finding information you know exists follows the same path. You search the document library. Nothing. You ask a colleague. They no longer remember where it is. You end up reconstructing the answer, or deciding without it.
The problem is not that the information is missing. It is that it sits across a shared drive, a mailbox, a business tool and somebody's laptop, and none of those places knows what the others hold. A document management system (DMS) files things better, but it still does not answer.
Ask the question, not the keywords
Omniscient takes the question as you would put it to a colleague. In plain language, in one sentence, with no special syntax.
The assistant reads the question, decides where to look, queries the graph and the documents, then writes the answer. Five search strategies exist, from targeting a single entity to surveying a whole domain. They are picked automatically according to what the question needs, without the reader having to know which one fits.
The answer is written as it goes, and the search steps are shown live. You see what was consulted before the result arrives.
How does semantic search work here?
Searching by keyword assumes you use the document's own words. Nobody knows them in advance, and in-house vocabulary shifts from one department to the next.
Semantic search lifts that constraint. Every passage of a document is turned into a numeric representation of its meaning, called a vector, and the question goes through the same treatment. Retrieval then becomes a matter of closeness in meaning rather than coincidence of words: a question about "late payment penalties" brings back a paragraph about "compensation where the deadline is exceeded", with not a single word in common.
This vector search has a clear limit, and it is worth knowing: it retrieves what resembles, not what follows. A question requiring a count, a crossing of two lists or a chain of ownership has no answer inside any passage. That is where the knowledge graph takes over.
An answer is checked, not believed
This is where everything is decided.
Every claim in the answer points back to the document that justifies it. The citation is clickable: it opens the relevant entity's record or the source document, at the useful spot.
The practical consequence is simple. An answer you doubt can be verified in seconds instead of being set aside out of caution. That decides whether a tool like this is genuinely adopted, far more than the quality of its prose.
What that mechanism actually prevents, and what is left to watch for in spite of it, takes a page of its own: avoiding hallucinations.
What a document engine cannot do
Some questions have no answer in any single document.
Which suppliers work on more than one site, and which of them are covered by a contract that has expired?
Answering that means crossing a supplier list, site work and contract dates. Three sources, three formats.
Omniscient can answer because it does not only search the documents, but the knowledge graph drawn from them. Objects there are linked, so the crossing is a walk rather than a search.
Permissions first, search second
A platform that reads a company's entire document estate is worth nothing if it ignores existing access rules. That is the first objection people raise, and it is a fair one.
Filtering applies during the search, not after it. An answer cannot rest on a document its reader may not see, because that document never enters the scope being consulted. Two people asking the same question may therefore get two different answers, each grounded in what they are allowed to read.
Rules are set per role and per object, on entities, relations and documents. What that timing changes, and what stays on record, is covered on the security and access control page. Five roles exist by default, and custom roles are defined per resource and per action.
What can the platform read?
Sources connect without development, and office documents are handled in depth.
| Family | What is covered | What is drawn from it |
|---|---|---|
| Office documents | PDF, Word, PowerPoint, Excel | Text, structure, tables |
| Images | Common formats, standalone or inside a document | The text they contain |
| Structured files | CSV, TSV, Excel | One entity per row, typed columns |
| Databases | SQL | Existing tables and relations |
| Online workspaces | Google Drive, SharePoint, Notion | Files and pages, with their tree |
| Business applications | HubSpot, REST APIs | Business objects and their fields |
Large files are read as a stream, which avoids saturating the machine on a spreadsheet of several hundred megabytes.
What does it cost in tokens, and how do you know?
A question put to a language model has a cost. Across daily use by several dozen people, that cost stops being negligible.
The platform counts it. Every run leaves a trace: what ran, how many tokens were consumed, for what amount. Generation tokens and vectorisations are counted separately, per run and per period.
The assistant's audit console adds the rest: question volume, durations, routing decisions, ratings given to answers. The AI bill stops being an end-of-month surprise.
Where to start
A useful demonstration is built on your own documents and on a question you actually have. That is the only way to judge, because a demonstration on catalog data says nothing about what yours will give.
If your documents cannot leave your network, the sovereign AI page covers the three installation modes, including the one that runs entirely on your own machines.
Tell us what you are trying to get, and we will show you the path to the answer.