Transparent workflow

How Janus.Guru builds a verifiable answer

Janus does not look for a ready-made answer on the open internet. It works inside a selected collection: retrieve evidence, assemble a response, and show the basis. When the basis is insufficient, a bounded answer is better than a confident guess.

The path

Five controls between a question and a response

01

Ingest and describe

Every source needs an owner, date, language, URL, status, and an explicit use boundary.

02

Split into evidence units

Fragments preserve enough context to be useful without turning every page into one opaque block.

03

Retrieve and rerank

Semantic and lexical signals are combined before the strongest candidates reach generation.

04

Generate with guardrails

The assistant is told which language, format, source policy, and refusal boundary apply.

05

Review the result

Sources, date, edition, and responsibility stay visible for the person who will use the answer.

Scientific basis

What the evidence status means

Janus uses established retrieval-augmented generation, information-retrieval evaluation, and data-provenance practices. QazKnowledge is still reviewing the product's claim-to-evidence passport, so this page does not call the system scientifically proven.

Already testable

Technical evidence

Source contracts, importer tests, retrieval benchmarks, query logs, and public collection limits.

Metrics and coverage
QazKnowledge review

Research passport

Claims, bibliography, applicability limits, and periodic review must be linked before publication.

Open QazKnowledge
Not proven by a benchmark

Boundary of inference

A retrieval benchmark does not prove every collection, legal conclusion, or deployment decision.

Responsibility boundaries