Your people already use AI assistants. Every one of those assistants starts the day knowing nothing about your company, so people paste documents in one at a time. It is slow, every answer depends on who pasted what, and nothing stops a confidential file from landing in the wrong conversation.
The standard fix is to connect the assistant to the company's documents: index the files, search them by meaning, hand the best passages to the model. It works, and it is where most companies are right now. It also skips a question the file server used to answer for you. Who may read this?
Four questions no index answers
A search index is built to find the passage that best matches the question. It is not built to ask whether the person asking, or the assistant asking on their behalf, is allowed to read it. The moment a shared index sits between your files and an assistant, four things that folders and habits used to handle stop being handled at all.
- Who may read this. The permissions lived on the file server. The index does not carry them. A salary band pasted into the wrong chat is a mistake a person makes once. An index that serves it to everyone makes it on every search.
- Which version is current. A policy gets replaced. The old draft is still in the folder, and it is often the better match for the question, because it says the same thing in more words. The model answers, with confidence, from last year.
- Who wrote it. Folders now hold documents written by machines: meeting summaries, generated drafts, notes an assistant took. An assistant that reads them is a model citing a model, and nobody reviewed the source.
- Whether it has expired. Price lists, procedures and contact sheets have a shelf life. Nothing in a search index expires them.
None of this is exotic. It is the ordinary state of a shared drive, and a shared drive was fine, because a person opened each file and knew what they were looking at. An assistant does not.
Where the files go is the other half
For many organizations the reason AI is not in the building yet is simpler than any of that: the documents cannot leave. A law firm's matter files. A clinic's clinical folders. A program's own records on its own servers. Every hosted retrieval service asks for those files to be indexed somewhere else, and the answer is no. So the assistants stay generic and the documents stay unread.
What we built
PremAgentic is a knowledge server that runs on your own servers and decides what the model is allowed to know, before anything is retrieved.
It indexes the Markdown, text, PDF, Word and Excel files your organization keeps and returns cited passages from them, with the file and the heading, page or sheet each one came from. It generates no prose; the assistant you already use does the talking. People reach it over HTTPS. Assistants reach it over MCP, the open standard for connecting AI tools to data. Agents read from it and never write to it.
The four questions above are four gates. Each one is checked on every search, inside the database, before the search runs.
- Access. Each folder carries an ordered list of allow and deny entries, set by an administrator. The first entry that names the caller decides, and no entry means no. If it cannot tell who may read a file, nobody gets that file. (What goes wrong when a permission filter meets a vector index.)
- Lifecycle. Superseded material is held back unless a caller asks for history, so the answer is the current version.
- Trust. A document that declares a machine author is held from assistants until a person has reviewed it.
- Freshness. Past its stale-after date, a document is flagged for people and held from assistants.
Because the gates are rules in the database and not instructions in a prompt, they hold the same way for a person at a keyboard and for an agent running unattended at three in the morning. Every question is logged with who asked, where the assistant's model ran and the passages it was given.
It all runs on hardware you own. The index, the embedding model that makes search by meaning possible, and the rules run on Windows or Linux, over a stock PostgreSQL database your own staff can back up and inspect. It needs no cloud service and no API key. Every assistant says where its model runs, and a folder marked never for hosted models is never served to one whose model runs outside the network.
Why us
We ran into every one of these problems ourselves. Two earlier posts here describe the knowledge layer we built for our own work: a wiki of our own notes with a local vector index, read by our own assistants. It worked on day one for one reader. The day more than one person and more than one agent read from it, the questions changed: who may read what, which version is current, and whether a passage was written by a person. The answers had to live in the database and be checked on every search, not in a prompt. PremAgentic is that layer, built to be installed inside someone else's organization.
Who it is for
Organizations whose files carry obligations, and whose people want the assistant anyway. A law firm where an attorney asks in plain words and gets the passage that answers, out of the files the firm keeps, and never a superseded draft. A clinic where the clinical folders are open to clinical staff and closed to the front desk, so one deployment serves the whole practice. A program whose files stay on its own servers, with a folder marked never for hosted models that is never served to one. Law, medicine and defense are examples. The shape is the same anywhere a document has a reader list.
Get it
PremAgentic 0.1.0 is out for Windows and Linux, and it is open source under the GNU Affero General Public License 3.0. The download, the source code and the manual are at premagentic.com. If your team wants AI assistants working over its own files, on its own servers, and wants it installed and running, that is the work we do: local AI, and private AI for Arizona businesses.