Corporate AI knowledge base with role-aware answers

Context

A company's real knowledge lives in regulations no one rereads, documents no one can find, and the heads of a few experienced people. New hires re-ask the same questions; expertise walks out the door when someone leaves; and the honest answer to "where is that written down?" is often "nowhere findable".

The knowledge base makes that tacit material retrievable — and answers in a way that respects who is asking.

Approach

Four source types — regulations, working documents, correspondence and captured expert experience — are ingested into one retrieval contour. Scanned and photographed material passes through OCR first, so paper-era documents are not second-class citizens in the index.

Answering is retrieval-augmented by design: the LLM is constrained to respond from retrieved source passages, which makes answers verifiable against the originals rather than plausible-sounding inventions. Where retrieval finds nothing, the system says so instead of improvising.

Architecture

Employees ask questions in Telegram — the tool already open on their phone — while the heavier machinery stays invisible. File storage, CRM and 1C feed the contour: documents, client history and operational records flow in from the systems where they already live, so the base tracks the company rather than becoming another stale wiki.

The pipeline is ingest → OCR where needed → index → retrieve → answer, with role filtering applied at the retrieval step, not bolted on at the end.

Key engineering details

Role-aware scoping is the detail that makes the system deployable in a real company. Retrieval is filtered by the asking employee's role before generation, so a line employee and a finance lead can put the same question to the same base and each receives only what they are entitled to see. Access control lives in the retrieval layer — the model never holds material it must pretend not to know.

Grounding doubles as an audit surface: because every answer traces to source passages, a wrong answer is a fixable data problem, not an unexplainable model behavior.

Outcome & what shipped

What shipped is a product line ready to pilot on a client's own document corpus: the ingestion pipeline with OCR, the retrieval contour, role-scoped answering and the Telegram front end, with file storage, CRM and 1C as feeding integrations.

Deployed, it works as an internal AI helper, an onboarding accelerator and a document search — and as insurance against expertise evaporating with staff turnover. Planned next: answer-with-citation for auditability and onboarding checklists wired to the base.

What we built

• Document ingestion

  Regulations, documents, correspondence and expert notes pulled through OCR into one retrieval contour.

• Grounded answers

  RAG constrains every answer to retrieved source material — verifiable, not improvised.

• Role-aware scoping

  Retrieval filtered by the asking employee's role before generation, not after.

• Ask where you work

  Employees query from Telegram; file storage, CRM and 1C feed the base behind the scenes.