A private, secure, grounded AI you administer yourself — offline, on‑prem, in your cloud, or hybrid. Here's how it answers, the modes you can allow, a look inside the admin portal, where it fits, and how a machine talks to it.
SYNTHEKOS runs an AI model over your documents. It answers from your material, names the source of every answer, and says plainly when it doesn't know — instead of inventing something expensive to be wrong about. Run it as a sealed offline box, on a local server, on a private cloud, or hybrid; give people and machines exactly the access they should have, and nothing more.
The model and documents live where you put them. Nothing leaves unless you decide it should.
Answers name their source, clickable to the page. No match, no bluff — it tells you.
Admins curate the corpus and grant each person, group or machine its own capabilities.
Pull the cable and it still answers. Being right matters most where the network isn't.
You ask a question. SYNTHEKOS searches your own documents for the passage that answers it. If it finds it, it answers from that text and names the file. If it doesn't, it declines — rather than filling the gap from a model's memory. The honesty is structural, not a setting.
A person, a portal visitor, or a machine sends a question.
It retrieves the exact passage from the corpus in scope.
Reads from the source and names the file — the right 42 Nm out of four adjacent numbers.
Empty sources, an honest abstention — never a guess dressed as fact.
TRUTHGATE. Before an answer goes out, a small deterministic check compares every number in it — with its unit — against the passages it came from. For machines, portals, kiosks and the face, an answer that fails is withheld and replaced with the honest abstention; in the text chat the check currently runs in watch mode and is logged. It checks numbers, not the meaning of sentences, and says so.
One assistant, four answer policies. An administrator decides which are allowed for each person, group or machine — exactly like every other permission. A locked‑down deployment can allow Grounded only; an open research portal can enable them all.
Answers only from your documents, cites the source, and abstains when it isn't covered. No outside knowledge — ever. Optionally layer +Web on top (below) to widen the evidence while keeping cite‑or‑abstain.
Live todayAnswers from the model's general knowledge — for brainstorming and everyday help where grounding isn't required.
Live todaySearches your corpus first and answers from it, cited. Facts — names, dates, numbers, quotations — come only from your documents; where they are silent it may add general knowledge and says so. With nothing in the corpus, the answer is labelled “from general knowledge”, with no source. TRUTHGATE never hides a hybrid answer: a number that is not in your documents is flagged ⚠.
Live todayA layer over any mode: when your corpus can't answer, it consults live web sources (via a self‑hosted SearXNG) and cites each link. Off by default; admin‑granted, per user; the user can toggle it and pick when it fires (fallback, augment, always).
Live todayEvery mode is an administrator‑granted capability, and the policy travels with the identity, not the page.
When the answer isn't in the corpus yet, teach it — in a sentence, with one click on any reply, or out loud to the face. What you teach joins your own knowledge base: grounded, cited back to you, and yours to edit or delete. It's saved as corpus you control, never fine-tuning — so every taught fact stays transparent and reversible.
Type what it should know — or tell the face "remember that…"; it reads it back and asks before saving. Indexed into your knowledge, cited in future answers, marked with who taught it and when.
One click on any message captures it straight into your knowledge base.
A "My knowledge" panel lists everything you've taught — change it, or remove it entirely.
Learning is a per-user capability, off by default — like every other dial.
Everything is managed from the admin portal: who exists, what they can do, what they can see, and how they reach it. These are annotated views of the real screens.
| Login ID | Role 2 | Access 3 | Answer mode 4 | Actions |
|---|---|---|---|---|
| admin | admin | self · upload | Hybrid | Edit |
| reception | user | facilitated | Grounded | Edit |
| data-bot | euser | machine · key | Grounded | Key 5 |
| family | portal | open link | Open | Edit |
Screens are annotated representations of the live admin UI; exact labels evolve as the product does.
Other devices — sensors, controllers, apps, dashboards — can query the box with an authenticated key and get the same grounded, cited answers a person would. And when nothing matches, they get an honest empty‑sources signal they can act on: proceed, or hold safe and escalate.
Issued once from the machine's identity, stored server‑side — never in browser JavaScript.
POST /api/ask with the key and a prompt — or a set of readings a controller wants judged against the loaded doctrine.
It retrieves from that identity's corpus (its own documents ∪ any shared corpus it's allowed).
A grounded, cited answer when the corpus covers it — or an honest abstention with empty sources when it doesn't. Every reply also carries the number check (verified, abstained, or unverified = the draft was withheld) and the model that wrote it. The check covers numbers and their units, not the meaning of sentences.
# a device asks — with an authenticated key { "prompt": "What torque for the coolant flange bolts?" } → { "answer": "42 newton-metres.", "sources": ["KESTREL-7-manual.pdf"], "grounding": "grounded", "model": "gemma4:latest", "truthgate": { "verdict": "verified", "checked": 1 } } # when the loaded documents don't cover it, it says so { "prompt": "What's the capital of France?" } → { "answer": "I don't have that in the loaded materials.", "sources": [], "grounding": "grounded", "model": "gemma4:latest", "truthgate": { "verdict": "abstained", "checked": 0 } } # and if the model's draft carried a number that is not in the documents, # the caller never receives it → { "answer": "I don't have that in the loaded materials. (The draft answer was withheld: \"55 Nm\" — not in the evidence.)", "sources": [], "grounding": "grounded", "model": "gemma4:latest", "truthgate": { "verdict": "unverified", "checked": 1 } }
Anywhere being wrong is expensive, the connection is unreliable, or the data can't leave — and anywhere a group wants a grounded assistant over its own knowledge.
HR or IT policy on one box; staff ask and get cited answers, and an audit of what the handbook doesn't yet cover.
Verified procedures read aloud under a machine — the right number, gated at the steps that can hurt you.
A community or association runs a self‑administered portal — members just ask; admins curate.
Company documents, per‑team access, its own admins — grounded and audited, on infrastructure you control.
A public URL portal over one product's manuals — cited, honest when the manual is silent, gated by a link or password.
Sovereign by construction — ships, mines, care settings and secure sites, beyond the network.
A face that greets visitors, answers from its own documents only — a listing, a product range, a building — and hands the rest to a person.
The same material as an illustrated PDF, plus a briefing on request.
SYNTHEKOS can answer through an animated face that listens and speaks — for a signed-in person (the avatar, chosen per person by the administrator) or as a walk-up kiosk for visitors. Same documents, grounding, truth gate and audit as the chat; always clearly a synthetic presence.
Lip-synced answers, with a plain, honest look when it must abstain. Tap to talk, or a continuous conversation you can pause.
A kiosk can greet whoever walks up, then go quiet and forget the conversation when they leave. The camera stays in the browser — nothing recorded or sent.
Hello and thanks are answered politely; real questions come from the documents; the rest goes to a person you name.
For a signed-in person it keeps the conversation and can be taught by voice — confirmed, and attributed.
See it: kiosk greeting demo → (scripted, runs in your browser). Still on the horizon: letting on-device expression cues colour how the face speaks — never what it knows.
Also on the horizon — the TruthFirewall. The cite‑or‑abstain gate, generalised so it can sit in front of a far more powerful external model: the big model may draft the answer, but nothing reaches the user until it's checked against your own documents — supported claims pass (cited), unsupported ones are withheld. Frontier capability with sovereign trust; the honesty check stays local and yours.