Internal knowledge assistant for a logistics company

Problem

Operations staff spent hours each week searching SOPs, contracts and policy documents spread across shared drives and a legacy wiki. Answers were inconsistent, and the internal support desk was buried in repetitive questions.

The company needed a single place to ask questions in Russian and Uzbek and get reliable, sourced answers — without exposing documents outside the company network.

Approach & architecture

• Designed an ingestion pipeline that normalises DOCX/PDF/HTML sources, chunks by document structure, and stores embeddings in PostgreSQL with pgvector — no extra infrastructure to operate.
• Built hybrid retrieval (vector + keyword) with reranking, and a strict answer prompt that cites the exact passages used and refuses when confidence is low.
• Shipped a small React chat UI integrated with the company's SSO, plus an admin view for re-indexing and reviewing low-rated answers.

Result

• Support tickets on policy/procedure questions dropped ~35% within two months.
• Median answer time under 3 seconds; 92% of sampled answers rated correct, with accurate citations.
• Runs on a single VM inside the corporate network; documents never leave it.

What made it robust

• Evaluation set of 180 question/answer pairs run on every change to the prompt or retriever.
• Role-based document access enforced at retrieval time, not just in the UI.
• Structured logs and per-question cost tracking; handover with a runbook and architecture doc.