A citation
The answer should lead to a document and an excerpt, not to the model’s “memory”.
Documents
An answer based on your files, with the excerpt in view.
We search the company’s materials first, then the model drafts an answer only from what was found. Embeddings and copies of excerpts stay on the organisation’s server.
A local index
Vectors are not sent to an embeddings service in someone else’s cloud.
Permissions
A person should not receive a passage they cannot open in the source system.
How it works, without magic
The collection is split into excerpts, their representation is computed, and the index sits next to the model. A user’s question hits the index first. Only the retrieved passages and the question itself go to the model.
The model does not need to “remember” the company. It reads what it was just given, and it should say when those passages do not contain the answer.
Which materials
The format list is closed against your collection, not against a slide.
- Office documents, text PDFs, intranet pages, helpdesk exports.
- Scans without a text layer need a separate recognition step. We do not promise that every scan will be readable.
- Out-of-date materials have to be switchable. Otherwise the assistant will quote an old procedure with a straight face.
- If files hold personal data, the index scope is a decision for the controller. General description: security and compliance.
Permissions
The index itself does not know your departments. If sales staff cannot read HR files, the deployment has to mirror that: separate collections, or a filter tied to the account.
The method depends on the system where the files already live. We do not sell a universal plugin “for every drive”. When the source is Odoo or tickets, we look at integrations.
Questions
Can you see where an answer came from?
That is the point of this layout: the answer keeps the document name and the excerpt it used. The user should be able to open the file and check, rather than take the model’s word.
Will every employee see every document?
Not if we design it that way. The split has to follow permissions in the source system. The model will not invent that, so we settle it while reviewing the collection.