Blog · 3 November 2026 · 7 min read
Local AI and documents that change: versions, duplicates, sources
A price list changes every season, a contract has three revisions, a procedure gets rewritten. An AI that answers with the document from two years ago is worse than no AI. Here is how the archive is kept up to date, and the part that no tool can do for you.
In short: yes, the local AI notices changes. Syncing compares the files in the connected folders with those already in the archive: it uploads only the new and modified ones and removes those that have been deleted. On the computer's folders this happens as soon as the file changes; on cloud sources, at scheduled intervals. Under every answer you find the document and the passages used. What the system does not do is decide which version is the right one: if the old price list and the new one are both left in the folder, it reads both.
If I edit or replace a file, does the AI know?
Yes, without you doing anything. Open WebUI's synchronisation tool calculates a fingerprint of the content for each file and compares it with that of the file already loaded. This gives three cases:
- New file: it is uploaded and indexed.
- File modified: the fingerprint has changed, so the new version replaces the old one in the archive.
- File deleted from the folder: it is removed from the archive too, and the AI stops using it.
Files that have not changed are left untouched: that is why the update is quick even with thousands of documents.
How quickly? It depends on where the file is. For the computer's folders there is a mode that reacts to disk events: the file comes in a few moments after it is saved. For cloud sources, such as Google Drive, synchronisation runs at scheduled intervals, for example every hour. In both cases some time is then needed for the document to be read and indexed: a long PDF is not available the instant you save it.
What happens to old versions?
It depends on how you save them. If you overwrite the file, the old version disappears from the archive. If instead you save listino_v2.pdf next to listino_v1.pdf, the system treats them as two different documents: it reads both, and the AI may answer with the old price.
It is not a flaw in the tool: no program can know which of two revisions of a contract is the signed one. A simple rule is needed, and in practice three are enough:
- An "archive" folder excluded from search. Superseded versions are moved there: they are kept, but the AI doesn't consult them. Syncing lets you exclude folders and file types.
- Just one file per "live" document. The current price list always has the same name and is overwritten.
- The date in the document. If a file states "valid from 1 September 2026", the AI can tell you so in its answer and you see straight away whether it is the right one.
Tip: if you also need to consult the history ("what price were we charging in 2024?"), the archive can become a second collection, queried only when you ask for it.
And duplicates?
A file identical to the one already loaded is recognised and is not reloaded. Cleaning up scattered copies, however, remains your job: the same PDF saved in three connected folders, or the copy "definitivo_ok_2.docx", are separate documents as far as the system is concerned.
The effect is not serious but it is noticeable: when the AI looks for the most relevant passages, the copies take the place of other useful information, and the answer becomes poorer. Before connecting a folder, an hour of tidying up is worthwhile: it is the job that improves the quality of the answers most, more than the choice of model.
How do I know where an answer comes from?
The sources appear below the answer: the document and the passages the AI used. You can open them and compare them with the original. It is the most important check, because it tells you two things at once: whether the information really exists and whether it comes from the right document.
Getting people used to looking at the source is half the training. If the answer cites the archived price list, the problem is in the folder, not in the AI, and it is solved by moving a file. How document search works is explained in the guide on how to connect a local AI to company documents.
Which formats does it read?
Those in common office use. With the document reading engines that work alongside Open WebUI, for example Docling, the archive accepts:
- PDFs, including scanned ones;
- Word, Excel and PowerPoint, and the equivalent open formats;
- text, CSV and web pages;
- emails saved as files;
- images containing text, read using character recognition.
New documents come in by themselves: everything you save in the connected folders is loaded at the next synchronisation, with no manual upload involved.
What it does not do well
- Low-quality scans. Crooked photocopies, stamps over the text and handwriting produce reading errors, and a misread number becomes a wrong answer.
- Complex tables in PDFs. Merged cells and tables split across several pages lose their structure. It is better to give the AI the source Excel file.
- Very large or protected files. There are size limits, and password-protected documents cannot be read.
- Contradictions between documents. If two files currently in force say different things, the AI may cite one without noticing the other.
- What isn't written down. If the real rule is "we always give our long-standing customer an extra 5%", and it isn't in any document, the AI doesn't know it.
How we set it up
Before installation we look at the folders together, on site at the client's premises and throughout Italy: which to connect, which to exclude, where to put the archive of superseded versions. We then configure the synchronisation and test it in front of you, replacing a document and checking that the answer changes. In the training we insist on one habit only: looking at the source.
From that point on the archive updates itself and the company manages it independently. The full project is described on the Local AI for businesses page; if you would like an opinion on how your documents are organised, write to us.
Sources
- Open WebUI – Knowledge: syncing a folder and comparing files
- Open WebUI – Knowledge Base Sync (oikb): incremental sync and watch mode
- Open WebUI – oikb daemon: interval and scheduled syncs
- Open WebUI – RAG: document search and citations
- Open WebUI – Extracting text from documents (PDFs, scans, images)
- Docling – Supported file formats
About the author: Dabryx
Dabryx is an Italian software house that installs local AI in companies, on site across Italy, on servers or Macs. The basic installation starts at €2,900, hardware not included, and consulting at €100 per hour. The office is in Rocca Priora, near Rome.
We also install local AI in Rome, Milan, Parma and Reggio Emilia, and we develop custom management systems and custom software.
Frequently asked questions
If I replace a document, does the local AI use the new one straight away?
Yes. Synchronisation recognises the modified file and replaces the old version in the archive. On the computer's folders this happens shortly after saving, on cloud sources at the next scheduled synchronisation.
Can the AI answer using an old document?
Yes, if the old version is still in a connected folder. The system does not know which revision is in force: superseded versions should be moved to an archive folder excluded from the search.
Does the local AI cite the source of its answers?
Yes. Below every answer based on the documents the file and the passages used appear, and you can open them and compare them with the original.
Do I have to upload new documents by hand?
No. Everything you save in the connected folders enters the archive by itself at the next synchronisation.
Want an estimate for your company?
Tell us how many people will use the AI and what for: we'll tell you which configuration you need.
Response times
Within 2 hours on business days (9:00–18:00) for clients with a support contract.
AI consulting
Basic installation from €2,900, consultancy from €100 per hour, on your company's server or Mac.
Where we work
Throughout Italy: we come on site to the client's premises and also support you remotely.