Most companies already own the raw material for a private AI assistant: ten years of proposals, contracts, manuals, reports and correspondence sitting on a file server or NAS, largely unread. Turning that archive into something staff can ask questions of is realistic, but the NAS alone does not do it. The NAS stays what it is, the organised and protected home of the documents. A separate AI computer runs the language model, and retrieval software sits between them, indexing the files and fetching the relevant passages for each question. Whether the result is useful depends less on the AI than on the state of the files: duplicates, outdated versions, scanned pages with no text and loose permissions all show up directly in the answers.
What are the three parts?
| Part | Job | What it is |
|---|---|---|
| Document store | Hold the files, with permissions, versions and backup | NAS or file server |
| Retrieval software | Read the files, build a searchable index, find passages for each question | RAG application with a vector database |
| AI compute | Run the embedding model and the language model that writes the answer | GPU workstation or desktop AI system |
How retrieval compares with the alternatives is covered in RAG vs fine-tuning vs long context, and ready-made retrieval front ends in Open WebUI vs AnythingLLM.
Is your NAS ready for the AI era, or just storing files?
A quick test. The more of these you can answer yes to, the less work lies ahead:
- Is there one agreed location for each kind of document, or copies scattered across personal folders?
- Can you tell the current version of a policy or price list from the superseded ones?
- Do folder permissions reflect who should see what today?
- Are scanned documents searchable text, or just images of pages?
- Is confidential material such as HR and payroll kept apart from general files?
- Is the NAS backed up, with snapshots?
- Can another machine on the network read the shares at a reasonable speed?
What housekeeping matters most?
- Decide the scope. Start with one well-kept area, such as product documentation or standard operating procedures, not the whole server.
- Remove or separate the outdated. An assistant that finds three versions of a price list will quote the wrong one with confidence. Archive superseded material outside the indexed area.
- Deal with duplicates. Copies of the same file crowd out other results.
- Make scans readable. Run optical character recognition on scanned PDFs. Text that cannot be extracted cannot be found.
- Fix permissions first. Retrieval should return only what the person asking is allowed to see. If the folder permissions are wrong, the assistant will leak, however good the model is.
- Fence off sensitive data. Keep personal and confidential records out of the first phase entirely.
- Use meaningful names and folders. File name, folder and date become clues the retrieval system uses and shows as sources.
- Appoint owners. Each indexed area needs someone responsible for keeping it current.
This is unglamorous work, and it is where most of the value comes from. A side benefit: a tidied, permission-checked file share is better for people too.
Can 10TB of data be turned into business intelligence?
Some of it. Volume is a poor guide to value. CALC. Suppose a 10TB share is 70 percent video, images and software installers, 20 percent duplicates and old versions, and 10 percent current documents. The useful text is then about 10TB x 0.10 = 1TB of files, and the text inside them is far smaller again: a million pages of plain text is only a few gigabytes. The split here is illustrative; an inventory by file type and age on your own NAS will give the real figure, and it is usually a surprise.
What a retrieval assistant does well with that text is answer questions that have an answer in a document: what did we agree with this customer, what is the procedure for this fault, which projects used this component. What it does not do is analyse numbers across spreadsheets or databases; that is a reporting and analytics task for other tools.
An AI assistant does not make an untidy archive wise. It makes it fast. Whatever is in there, right or wrong, current or stale, now reaches staff in seconds.
What does the hardware look like?
- NAS: the existing one may do, if it has capacity, snapshots and a wired link of 2.5GbE or better. Indexing reads every file once, then only changes.
- AI compute: a system able to run a capable language model and an embedding model for your number of users. See analysing company documents locally on GB10 and, for larger deployments, from 10,000 documents to one private AI brain.
- Index storage: fast local SSD on the AI system, backed up to the NAS. See vector databases explained.
Sizing the indexing job itself is covered in can AI search 10 years of company documents in seconds.
What should you watch for?
- Wrong answers from stale documents. Show the source and its date with every answer. See AI hallucinations explained.
- Instructions hidden in documents. Files from outside the company can carry text aimed at the AI. See prompt injection explained.
- Personal data. Indexing HR or customer records makes them easier to find. Apply the same care, and the same legal obligations, as for the original files.
- Drift. New files need indexing and deleted ones removing, on a schedule.
Buying in Singapore
Document store: the Synology DS925+ (S$1,260) or Synology DS1825+ (S$2,148), with Seagate IronWolf Pro drives from S$390. AI compute: the NVIDIA DGX Spark 4TB (S$9,450), ASUS Ascent GX10 1TB (S$10,350) or HP ZGX Nano G1n 4TB (S$11,499) (listed 12 Nov 2026; prices change; ask for a current quotation). Retrieval software is chosen separately and is not included with this hardware. To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can a NAS be used as an AI knowledge base? A NAS can be the document store for one. The AI model and the retrieval software run on a separate AI computer, which indexes the files on the NAS and answers questions from them.
Does a Synology NAS run the AI model? No, not a large language model at useful speed. The NAS stores and protects the documents; a GPU system runs the model.
What needs to be done to files before using them with AI? Remove or separate outdated versions and duplicates, make scanned documents searchable, correct folder permissions, keep sensitive records apart and assign owners to keep each area current.
Will the AI respect who is allowed to see which documents? Only if the retrieval software is set up to apply each user's permissions and the folder permissions themselves are correct. This should be designed in from the start.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
