IAssistant
Half a name, a rough date, Russian and Kazakh in one sentence. The assistant finds the record and answers in the language you asked in.
Assistant
The contract from Drive, the invoice from 1C, the payment from the bank, each line with its source.
Mixed languages, one term spelled the Kazakh way and the Russian way, one counterparty named three ways in four systems. All resolve to the same records.
Half a name, a rough date, a figure someone remembers. The question is matched by meaning as well as wording, so it finds the record and the two beside it.
Who signed, what was invoiced and what was paid live in three systems. One question reads all three, and nobody switches between tabs.
The answer comes back as one paragraph, and each sentence names the system it came from.
The index points back at the record in the system that holds it. An answer links to the source, and the source is what the reader opens.
Under each sentence sits the passage it came from and a link that opens the document at that page. Your team checks before it acts.
A sentence in the answer maps to a passage in a document. Both are shown, and the passage is quoted word for word.
If no passage supports a sentence, the sentence is not written. The answer shrinks to what the documents say, and it says so.
When the documents hold no answer, the assistant says so in one line. A plain “not in the documents” is worth more than a confident guess about a contract.
The model is not asked to fill a gap. If retrieval returns nothing that supports the question, the answer is that nothing does, with the systems that were searched.
Real questions with human answer keys, held fixed, and every change is scored against them before it ships. Exact identifiers must come back every time, and a wrong answer counts more than a missing one.
Permissions are checked inside the search, before anything is read. A payroll sheet closed to a new hire never reaches the model, the answer or the count.
A summary, a count and a citation are all drawn from the permitted set. There is no path by which a closed document shapes an open answer.
Each gets the answer their own access supports, and neither is told that a fuller one exists.
Your team asks where it already works. One index and one permission check sit under every surface.
The full console: a question, the answer, the passages under it and the source to open. Where a longer thread lives.
The same assistant in the messenger a team already has open. Ask from a phone, get the answer and its source, under the same permission check.
Ask about the page you are on, inside 1C or Bitrix24 in the browser, and get the answer beside the record, in the same tab.
Who asked, what was read and what was answered, tied to a person and a time. Your security team reads the log without asking us.
Each question is recorded against the person who asked it, with the time, the passages read and the answer given. Nothing is answered anonymously.
Your administrator opens the log, filters it by person or by document and exports it. It answers who has seen a document, and it does not need us to read it.
Run it on your servers or in a Kazakhstan datacentre, with the model and the key you choose. Retention is zero and training is excluded on every call.
The assistant runs on your own servers, on an air-gapped network or in a Kazakhstan datacentre. The index and the permission check run beside your data.
The model answers under your account with the provider, or runs on your own hardware. Swapping it touches neither the index, the permissions nor the log.
Every model call leaves through one point that carries the zero-retention terms, refuses a provider that may not see this data, and accounts for spend per tenant.
We connect Assistant to your systems and permissions, and the engineers who built it do the work. Tell us which question takes your team the most time.
