Most systems advertising artificial intelligence today solve the problem by sending your documents to someone else’s server. For a public body holding restricted information, that is not a feature: it is a leak.
Orpyca does the opposite. Search, precedents and drafted answers with citations can run inside your own infrastructure, without a single byte leaving the institution. And wherever a feature uses an external provider — the conversational assistant today, or drafting if you so choose — the system enforces a boundary that cannot be switched off.
You ask “what did we answer about the village street lighting” and the system finds the records dealing with the matter, even if none of them uses those words. The index feeds itself: every time a document is registered it enters the search without anyone having to reindex anything.
Before answering a right-of-petition request, the system shows what the institution has answered before on the same matter, with the registration number and date of each precedent. It is not a summary you must take on trust: it is a list of sources you can open.
The conversational assistant has no privileged account. It works with the token of the user talking to it, and every operation it attempts goes through role and clearance checks again. If you cannot see a case file, neither can the assistant, and no amount of persuasion changes that. Nor does it write on its own: each registration or transfer is proposed on a card and carried out only when the person confirms. Today it runs on Anthropic’s Claude API and strips restricted material before the conversation is sent.
Orpyca exposes more than two hundred operations as tools for artificial intelligence agents: register a document, look up a record, search, open a case file, view its index, list the retention schedule, review the processing inbox and consult the inventory. Each user only sees the tools their permissions allow.
You can dictate to the assistant instead of typing. The audio is transcribed by a service of its own, inside your infrastructure and without a graphics card; if that service is not running, dictation does not work rather than sending your voice elsewhere. It is off by default and meant for sentences, not for recording a meeting.
This is the part worth reading carefully, because it is the difference between a system you can use in a public body and one you cannot.
Material classified as restricted or above is never sent to an external provider. It is not a configuration checkbox a distracted administrator could untick: the checkbox does not exist. If the context of a question includes restricted material, that material is excluded and the answer degrades to plain retrieval, saying so.
Clearance filtering happens server-side, before anything reaches the model. The model never receives documents the user could not open themselves, so it cannot leak them by accident.
Every drafted answer records the provider and the model that produced it, in separate columns and with its entry in the audit trail. Six months later you can state precisely which provider received a given query.
If the artificial intelligence provider is not configured or does not respond, the system returns an explicit error. It never fabricates a plausible answer to disguise an unavailable feature.
This applies to search and to drafted answers with citations, and the default configuration is the sovereign one: everything inside your infrastructure. The conversational assistant is separate: today it only works with Anthropic’s Claude.
The semantic representation model runs inside the container itself and the generative model on a model server you host. No data leaves the institution. This is the default.
Semantic search and cited precedents only, without automatic drafting. It is the lightest option and the one demanding the least hardware.
For public material, if the institution so decides. The classification boundary still applies and admits no exceptions.
Nothing. It is included in the software, which is free and open source.
All you need is the hardware to run it on, and the figure is worth stating plainly: search by meaning and cited precedents work on the minimum server, with no graphics card. Drafted answers from a local model do need one — from 8 GB of video memory for the default model, and 24 GB if you want a model that drafts regulatory language well.
The conversational assistant does not use your hardware: it runs on Anthropic’s API, and that usage is billed by the provider directly to the institution.
It is a one-off investment and the equipment is yours. Set against a per-user monthly subscription, a card pays for itself in months and is still there the following year.
