Aragats4K sits between your people and your AI models. For each question it picks the cheapest model that meets your quality and speed targets, answers from your own documents with citations it checks itself, and keeps every request within budgets, access rights and an audit trail.
Runs on your own infrastructure · Local open models, optional cloud model · Every answer audited
Send every question to the strongest model and answers are slow and expensive. Send them all to a small one and the hard questions come back wrong.
Models answer confidently from memory, and a citation can point to a passage that doesn't say what the answer claims.
Documents leak across teams through the assistant, and spending is known only when the invoice arrives.
Each step shows live in the console, so you can see why a question went where it did.
A small classifier labels the question: task, complexity and whether documents are needed, in about 0.4 s.
A hybrid search, by meaning and by keyword, runs over the documents this person may read.
It picks the cheapest model that meets the workflow's quality and speed targets, based on measured results, not price lists.
The model answers from the passages it was given and cites each one, like [S1].
The server checks every citation and every cited claim against its source. A failed answer is retried once on a stronger model, then logged.
Measured in October 2026 on a product-support knowledge base, with local open models. Costs for local models are notional prices set in the gateway.
Single runs vary by about one question. We re-measure on your own documents during a pilot, and the router re-checks itself every night.
PDF (scanned pages read by OCR), Word, Excel, CSV, Markdown, text and images. Answers cite the page, and every document keeps its versions.
Groups, repositories with read, write and manage rights, and public, group or private documents. Private means private, administrators included.
Per workflow: documents only, documents first, or automatic. Answers from model memory are labelled unverified, and Wikipedia can be added with citations.
Repeated and reworded questions are answered from the cache only when a model confirms both have the same answer. Only verified answers are kept.
Each workflow has a spending limit in the gateway, and each person and API key a monthly one. Over-budget requests are refused before they reach a model.
Cost and the cost avoided by routing, cache hit rate, the share answered from documents, where the time goes, and the questions your documents can't answer yet.
Single sign-on with roles, hashed API keys, credentials stripped from logs, personal data masked, and an audit log that can't be edited quietly.
Local open models run on your own GPUs through one gateway; a cloud model can be added as an extra tier. Apps connect through an API and a command-line tool.
Answers from manuals, policies and tickets, with the source on every answer and a list of what the documents don't cover yet.
One gateway for every model and app, with budgets per workflow and per person, and evidence for which model to use where.
Self-hosted, with access control per document and an audit trail that shows who asked what, which sources were used and what it cost.