Company knowledge with sources: how the workflow works

An employee needs the current procedure. Two versions are in the catalogue, and some documents belong to HR alone. Access must be resolved before an answer is prepared.

Same company. Two people. Different access.

Ask a question. The reading room searches accessible documents and shows a source. The user can open the passage supporting the answer.

Check the scope. Group filtering applies during retrieval, and the database enforces access again at row level. A restricted document should never enter model context.

Find the current version. An outdated document and conflicting sources are separate states. Without an accessible basis, the system declines to answer.

Rebuilt local application, 25 September 2026. Data provenance and execution boundaries are described below.

What the rebuilt application lets you do

The new interface combines catalogue search with a full source reader. Switching the sample persona changes the visible subset of 600 documents and replays dated answers. The reader exposes the text, identifier and version. This is a local demonstration of access scopes; it is not an authentication test or a fresh model query.

Load more documents through the end of the selected persona’s available catalogue. Changing the persona or search resets the displayed list.

What the code does

Access control runs before answer generation. One layer limits retrieval to the user’s groups; another enforces a PostgreSQL row policy. Removing a citation from the screen later cannot undo the model having seen a restricted document.

Permission-matrix accuracy must stay separate from retrieval accuracy. The archived matrix questions contained document identifiers. A separate paraphrase trial produced weaker retrieval, which matters directly to employees asking in their own words. These evaluations answer different questions.

A useful first pilot begins with an organised collection, named owners and current versions. That material can then support comparisons of search, embeddings and answers. The pilot must also check whether revoking access actually changes the next query’s result.

What the evidence covers

The archived matrix covers 750 question–identity pairs. A separate paraphrase test recorded 0.511 top-result accuracy; the two measures should stay separate.

The corpus and identities are synthetic. Archived paraphrase tests exposed retrieval limitations; permission accuracy does not measure answer quality.

A first trial in the business

Start with one team and a controlled document collection. Test questions in users’ own words and access revocation after a role change.

Prepare:

• Documents with owners and versions
• Employee groups and permissions
• Questions with verified answer sources

Open the example and walk through its screens.

Source of this description

Demo source note. This description is grounded in recovered code and documentation. Archived measurements retain their original scope; they are not client-deployment results.