Smarty Your knowledge, kept in-house

All your documents can finally answer questions. Without a single byte leaving the building.

The AI assistant that needs no internet — built for organizations whose documents were never meant for the cloud.

Every school, public office, hospital, and company sits on years of accumulated knowledge: regulations, case files, medical protocols, contracts, meeting minutes, manuals, scanned archives. Finding anything means knowing who to ask — or which of a thousand folders to dig through. Modern AI assistants solve exactly this problem, but they come with a condition most of these organizations can never accept: hand your documents to someone else's cloud.

Smarty removes that condition. The AI, the search index, and the documents all live on your hardware — a single server, or even one capable office PC. Staff ask questions in plain language, in any language, and get answers with clickable citations to the exact source page. No cloud accounts. No data transfer agreements with AI vendors. No internet connection — not even an optional one.

Offline by design

The most secure network connection is the one that doesn't exist

Most "private" AI products are cloud services with privacy settings. Smarty is the opposite: a system that physically cannot send your data anywhere, because everything it needs — the language model, the search index, the documents — is already inside your walls. It runs on an isolated network, or with the network cable unplugged altogether. Your firewall doesn't have to trust Smarty; it can simply never see it.

Your premises server room, office, or a single PC Staff browsers local network only Smarty + AI engine models on local disk Documents database + files Everything the system needs lives inside this box — questions, answers, files, AI. no connection exists The internet cloud AI vendors · trackers anyone else
Not "encrypted in transit". Not "we don't train on your data, per our terms". Simply: no wire. The system is fully functional with the internet physically disconnected — questions, answers, imports, everything.
Who it serves

Built for the organizations that can't use the cloud

Schools & universities

Student records, individualized support plans, staff policies, years of internal rules — searchable by the people authorized to see them, and only them.

"What does our policy say about adapting exams for a student with dyslexia?"

Public offices

Ordinances, administrative decisions, archived correspondence, procedure handbooks. Citizens' personal data stays inside the institution, as the law expects.

"Which permits does a food stall at a public event need, per our current ordinances?"

Hospitals & clinics

Clinical protocols, internal guidelines, department procedures, scanned legacy documentation — instantly answerable at the point of need, with the source one click away.

"What is our current pre-operative fasting protocol for pediatric patients?"

Companies with sensitive IP

Contracts, R&D documentation, HR files, board minutes. Everything an NDA protects from the outside world stays protected from AI vendors too.

"What warranty terms did we agree to in last year's supplier contracts?"
Zero external connections Works fully offline Any language in, same language out Every answer cites its source Reads scans & images too Deletion is immediate and real
Security

Defense in depth — starting with the moat

Air-gapping is the foundation, not the whole story. Inside the walls, Smarty is built like a system that expects to hold pupil files, patient protocols, and boardroom minutes:

Isolation as architecture

No external services, no telemetry, no auto-updates phoning home, no third-party scripts in the interface. The attack surface of a cloud pipeline simply isn't there.

Access control

Every route requires an authenticated session; sign-ins use constant-time password verification and cryptographically signed, expiring session cookies. Accounts and role-based permissions govern who sees which collections.

Your data is never training data

The AI models arrive pre-trained and run locally. Nothing you store, ask, or receive ever flows back to any model provider — there is no channel it could take.

Answers you can audit

Every claim carries a numbered citation; one click shows the exact passage, file, and page. And when the documents don't cover a question, Smarty says exactly that — the AI is never consulted, so it cannot invent an answer.

Hardened at rest and in transit

Documents and the search index live in an encrypted store; browser traffic on your internal network runs over TLS. Defense holds even if a disk walks away.

Deletion means deletion

Removing a document erases the file, its text, and its search index in one stroke. The very next question already excludes it — there is no lingering copy in someone else's data center, because there is no someone else.

Architecture

Five pieces, all under your roof

Staff use a simple web page. Behind it, an API server handles logins, uploads, and chat, while a separate worker does the slow work — reading documents, OCR, indexing — in the background. They coordinate through the database, which doubles as a job queue.

Browser staff web page API server login · upload · chat Worker parse · OCR · index PostgreSQL + pgvector documents · search index job queue Local AI engine language + vision models File storage original files on disk asks · uploads queues a job per document worker claims jobs embed questions · generate answers embed passages · describe images reads files for parsing
Every box runs on your hardware. The teal path is the heart of the design: the API records a job in the database, and the worker picks jobs up safely one at a time. Original files are stored named by content fingerprint, so the same file can never be stored twice.
Getting documents in

Two doors, one pipeline

Documents arrive through the upload button, or through watched folders: point Smarty at a directory on your server or network share — say, the one your scanner already drops files into — and it keeps importing whatever appears there, subfolders included. Both doors feed the same pipeline, so every document gets identical treatment.

Browser upload one file at a time Watched folders automatic · subfolders too Shared ingest fingerprint · skip duplicates save copy · queue job Job queue a row in the database worker claims the job 1 · Read PDFs & Word parsed, scans OCRed, images described by AI 2 · Split into overlapping passages 3 · Index each passage becomes a meaning-vector (multilingual) 4 · Ready passages + vectors stored, document answerable If any step fails, the document is marked "failed" with the reason shown in the interface — nothing hangs.
Every document is fingerprinted by its content, so re-uploading or re-scanning the same file is a no-op. Watched folders are additive-only: new files get imported, but nothing in Smarty is ever deleted because a source folder changed — your knowledge base can't be emptied by an accidental folder wipe.
Getting answers out

Ask in any language, get a cited answer — or an honest no

Smarty doesn't hand your whole library to the AI. It first finds the handful of passages most related to the question by meaning, not keywords — a question asked in Slovenian finds the relevant passage even if the document is in English. Then the local model is instructed to answer only from those passages, citing each claim.

Question any language Understand question → meaning-vector Search the index closest passages by meaning, ready documents only relevant passages found Local AI writes the answer only from the numbered passages, in the language of the question Streamed, cited answer click [1] to see the source passage, its file, page, and download link word by word Honest refusal "I don't have documents covering this." the AI is never asked nothing relevant
The refusal path is a feature, not a failure: if the index holds nothing close enough to the question, Smarty answers with a fixed sentence and skips the AI entirely — it cannot hallucinate an answer from documents it doesn't have. For a nurse, a clerk, or a teacher, that reliability matters more than eloquence.
What it takes to run

One machine, modest requirements, no bills that grow

The whole system — web interface, database, AI models — runs on a single server or a well-equipped PC with a consumer GPU. The AI models are installed once during setup; from that moment on, the machine never needs to see the internet again. There are no per-seat licenses, no API metering, no usage bills — the marginal cost of the thousandth question is zero. And because everything is standard, self-hosted technology (PostgreSQL, containers, open AI models), your IT team can back it up, monitor it, and audit it with the tools they already use.