How do we load our documents into a private AI? Decide scope and access first; the loading itself is the easy part.
Short answer. Decide what goes in, clean it, set who can see it, load a small batch, then test it with questions whose answers you know. Behind the scenes the system extracts text from each file, splits it into chunks, turns each chunk into a numerical vector called an embedding and stores them in a vector database. At question time it retrieves the closest chunks and gives them to the model. The risks sit before and after that: outdated or confidential files that should not be included, scans that need character recognition, and an index that ignores who is allowed to read what.
01
Before loading: decide, clean, label
Decide scope by purpose, for example signed policies, standard templates, closed matters or past proposals. Leave out drafts, duplicates and anything superseded.
Check confidentiality: client files, personal data, privileged material. Decide which team may query which collection.
Convert scans. Image-only PDFs need optical character recognition so the text can be read; some systems use a vision model for this.
Name and tag: dates, owner, client or matter, and a status such as current or superseded.
Keep originals in your normal file store. The AI’s copy is a working copy.
02
What happens when you load a file
Step
What happens
What can go wrong
Extract
Text is pulled out of the file; scans go through character recognition
Tables and columns come out jumbled; hidden text is extracted too
Chunk
The text is split into pieces by size, with some overlap, sometimes along headings
Chunks that are too small lose context; chunks that are too large blur the match
Embed
Each chunk becomes a vector, produced by an embedding model
Changing the embedding model later means embedding everything again
Store
Vectors, text and metadata go into a vector database
Metadata such as owner and team is what access rules use; without it everyone sees everything
Retrieve
At question time the closest chunks are fetched and passed to the model
Old versions rank alongside new ones unless marked superseded
03
Access and sensitive files
OWASP lists unauthorised access and cross-context leaks as weaknesses of retrieval-based systems, and advises permission-aware vector stores with strict partitioning between groups. It also advises validating input documents before they enter the knowledge base, and using text extraction that detects hidden content, because hidden instructions in a document can steer the model.
Open WebUI’s documentation suggests keeping file upload and knowledge permissions off by default and granting them only by group. The UK government’s AI Playbook states the principle plainly: the model should only have access to the data the user of the model should be able to access.
04
Test before you trust
Write 20 to 30 questions whose answers you know, including some whose honest answer is “not in the documents”.
Check that every answer shows its source passage and that the passage is the right one.
Ask a restricted question from an account that should not see the file.
Ask about a document that has two versions and see which one is used.
Repeat after any change of model or settings.
05
What ANVE says
ANVE’s documents maker answers questions over the office’s own documents and shows the source passage. Its vision models read scanned documents (OCR), photos and diagrams. Full document drafting with design templates is not on ANVE yet.
Connectors, for example to Google Workspace or Zoho, work only if an administrator approves one. A connector sends that service only what it is asked to fetch or send, and every call is logged on the box. Ready-made Google Workspace and Zoho connectors are not on ANVE yet, and so are teams and finer roles (see /security).
How files otherwise get into ANVE (upload or folder), which file types it reads, how document access is set and how long indexing takes are not stated here. ANVEAI confirms them in writing before a paid pilot; ask to try a sample of your own files, including your worst scans.
06
Questions people ask
How many documents can we load?
Storage and memory limit it more than the model does. The index is small next to the originals but grows with every chunk. See the storage question.
Can it read scanned PDFs?
Yes if the system has character recognition or a vision model. Check it on your worst scans first.
Do we need to reload everything when we change the chat model?
No. Only a change of embedding model means embedding again; Open WebUI’s documentation says that change requires a re-index of all knowledge-base documents.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
