Olzhas
A large company always has more knowledge than it can serve. Magnum has internal regulations, onboarding and day-to-day work instructions, materials for every role — and a constant stream of people who ask their manager, a colleague, or whoever is nearby instead of opening the document. Olzhas is the layer that sits between the employee and the existing knowledge base. The employee asks a question. The assistant finds the right piece of content and answers from it.
- 01 A question in plain words → an answer from approved documents, with sources.
- 02 Russian and Kazakh, including mixed phrasing.
- 03 Every update is checked against 100 real employee questions.
Employees always asked.
The knowledge base existed and was kept up to date. It was just hard to use in the moment: too many documents, no quick way to land on the right paragraph, no time to read everything when you have a shift starting.
So employees did the natural thing — they asked a manager, asked a colleague, asked the person next to them. The team ended up answering the same questions week after week, and onboarding always took longer than it should.
- A large library nobody had time to search through.
- Managers and colleagues spent hours a week on the same repeat questions.
- New employees waited on people to explain what was already in the docs.
And the questions are the same, every single day.
A single chat over everything the company has already written down.
Olzhas is a RAG assistant on top of Magnum's internal knowledge. The team uploads materials in the admin panel. The system indexes them. Employees ask questions and get answers grounded in the company's own content.
The goal was not to be clever. The goal was to take the most-repeated questions off managers and colleagues and give new employees an answer in the moment, not in a meeting two days later.
- 01
Ask in plain language
Employees write the question the way they would ask a colleague. The assistant handles the search.
- 02
Answer from the company's own content
Answers are grounded in indexed internal materials, so the assistant stays inside what the team has actually approved.
- 03
Stay current
The team updates the knowledge base in one place. The assistant picks up new and changed material as it is added.
- 04
Run from one panel
An admin panel for the editors to upload, replace and remove materials without an engineering ticket.
Hybrid RAG built around the way the company actually keeps content.
The model is the boring part. The interesting part is retrieval: an employee asks in conversational phrasing, and the answer lives in one paragraph of one PDF out of hundreds.
So the search is hybrid: vector (BGE-M3 embeddings, a Qdrant index) plus full-text, with re-ranking of the results. The answer is built only from the retrieved documents and arrives with sources — the employee sees which files it was composed from. OCR pulls out tables and diagrams embedded in PDFs as images, and answers link to training videos — Russian voice-over with Kazakh subtitles. The base lives in the admin panel: the team uploads and replaces PDFs without engineers, the system re-indexes itself.
Hybrid search: vector (BGE-M3 on Qdrant) + full-text, with re-ranking.
Answers strictly from approved documents, with sources named.
OCR reads tables and diagrams embedded in PDFs as images.
Links to training videos: Russian voice-over, Kazakh subtitles.
Russian and Kazakh, including mixed phrasing.
Re-indexing from the admin panel — upload a document, the system absorbs it.
Retrieval gets its own evals. Separate from the answer.
In RAG, quality breaks in two different places: the system can fail to find the right document — or find it and still answer past it. We measure those separately. The base instrument is a benchmark of 100 real employee questions in Russian and Kazakh, with reference answers pinned to the source documents.
- 01 Read the questions We collect real employee questions and label the failures: not found, wrong document, made up.
- 02 100-question benchmark Reference answers in Russian and Kazakh, each pinned to a specific document in the base.
- 03 Retrieval eval We check retrieval on its own: did the right document land in the results, and how high. If search misses, no answer can save it.
- 04 Answer eval The answer is checked against the reference, and the citations — is the right document named as the source.
- 05 Run on every change A new prompt, model or document markup passes the benchmark first — then ships.
Hallucinations get their own line. The assistant answers only from approved documents: when the base has no answer, the correct behaviour is to say so, not to improvise. And every answer carries its sources, so the employee can check for themselves.
What changed
Managers stop being a search engine
The recurring questions move from people to the assistant. The team keeps the harder work where it belongs.
New employees ramp faster
The answer to "how do we do X" arrives in the moment of the question, not in a follow-up message the next day.
The knowledge base finally gets used
Material the company already produced now has a usable surface. Updating a document immediately reaches everyone.
Quality is measured, not asserted
The 100-question bilingual benchmark runs on every change — before it reaches employees.
Sitting on a knowledge base that nobody opens?
We help companies put a chat layer over the content they already have — a RAG pipeline tuned to your documents and an admin panel your operations team can actually run.