Your files stay on your computer
Nothing is uploaded, so there is no queue, no size limit nobody tells you about, and nobody waiting to approve your file. Add a document and it is searchable seconds later.
Check DevTools → Network yourself
Built for documents you are not allowed to uploadSecureRAG Cross-format file indexing.
SecureRAG Cross-format file indexing.
Runs locally. Never uploaded.
AI retrieval, fast and simple.
SecureRAG is a cross-format AI file search tool from Lilink. It accepts many file formats (PDF · DOCX · TXT · Markdown · CSV · HTML · JSON …) and uses AI to search their contents in one unified place. The whole process is fast and safe: your files always stay on your own machine, are never uploaded to the cloud, so your information stays completely private.
- Files never leave the device
- No account, no tracking pixels
- Not used for training
- Works offline once cached
Tool demo● LOCAL SESSION · 0 REQUESTS
Documents · 3 indexed
PDFExample file (1)42 pages · 84 chunks
DOCExample file (2)18 pages · 41 chunks
MDExample file (3)9 chunks
What is the termination notice period?
● Answer · generated locally
Either party may terminate on 60 days' written notice1. A material breach carries a 30-day cure period before termination rights activate2
Source 1 · Example file (1) · p.14 · §8.2
“…termination shall be effective sixty (60) days after written notice is delivered to the other party.”
0
documents leaked, ever
1request
model weights, cached after first load
25MB
default first download — nothing above 30 MB without your consent
100%
of parsing, embedding and retrieval happens in your tab
Capabilities
Search every file at once — no uploading, no sign-up
Search dozens of files at once
PDF, Word, Markdown, CSV, HTML, JSON and plain text, mixed together in one search. You do not convert anything first, and you do not open them one by one.
7 formats, one search box
Every answer links back to the source
Answers are built from sentences in your own files, each carrying a number that jumps to the paragraph it came from. You check the source instead of trusting a summary.
Click a number to read the paragraph
Works with the internet off
The first visit downloads about 25 MB of model weights. After that you can pull the network cable and keep searching and asking questions.
About 25 MB, once
Quotes only, or an explanation
Ask for the exact sentence and you get that exact sentence. Switch to the looser mode when you would rather have the wording explained than quoted.
Two search buttons: exact and fuzzy
Export the whole session
Every question, every answer and every matching sentence leaves as .txt, .md or .json, ready for your notes or a report.
txt / md / json
Trust center
Four claims we are willing to be tested on
Privacy posture
Test it yourself01No document transmission
Your files are read via the File API into memory. There is no upload code path in the application, and no endpoint that could accept one.
02No training on your data
Models run inference only. No gradients are computed, no feedback loop exists, and no content is retained outside your browser storage.
03One disclosed third party
Model weights are fetched from a public model host. That request carries no document content, no question and no identifier — only the model file name.
04Deletable by you, instantly
Clearing site data removes every chunk, vector and conversation. We hold no copy because we never had one.
Comparison
Local pipeline vs. uploading to a cloud assistant
| Criterion | SecureRAG | Cloud assistant file upload |
|---|---|---|
| Document custody | You; it never moves | Vendor infrastructure |
| Identity required | None | Account, often phone-verified |
| Network dependency | Only for the first model fetch | Continuous |
| Retention policy | Whatever your browser does | Plan-dependent, vendor-set |
| Answer citations | Paragraph + page, always | Inconsistent |
| Cost | Your electricity | Free tier cap, then paid |
| Natural-language fluency | Bounded by a 0.5–1.5B local model | Frontier-scale |
FAQ
Questions from people with sensitive documents
Can I verify the no-upload claim without trusting you?
Yes, and we would rather you did. Open DevTools → Network, clear the log, then add a document and ask several questions. The only request you should see on a first visit is the model weights. Repeat with the network disabled: it keeps working. The in-app shield panel reports the same records without DevTools.
What is the one third-party request, exactly?
A GET for the model file from a public model host (HuggingFace or a mirror). The request contains the file name and nothing else — no document, no question, no cookie identifying you. The models page lists every file by name and size so you can check what lands on your machine.
Is this suitable for regulated or confidential material?
Architecturally it removes the third-party transfer problem, because no transfer occurs. Whether that satisfies your specific obligations is a judgement only you and your compliance team can make — we provide the technical facts (data flow, storage location, deletion method) on the security page so that judgement can be made quickly.
How does the site fund itself?
Labeled display advertising, up to six units per page (three in-page banners, two side rails, one collapsible anchor) and never placed inside the chat workspace or above the primary action. There is no subscription and no data sale. Personalized ads load personalized only after you agree; outside the EEA, UK and Switzerland a refusal still shows non-personalized ads.