RAG & Local LLMs

EVA Cosmetics Offline HR Assistant

Context: Enterprise Project· Role: AI Engineer / Architect· Status: On-Premise Internal System

A fully offline bilingual assistant that lets employees query internal HR policies in Arabic and English, with no cloud dependency.

Business Problem

Employees needed a simple way to find HR information inside internal documents, in either Arabic or English, without exposing sensitive policy content to any external service.

Users

Employees across the company asking HR policy questions in whichever language they're comfortable in.

Constraints

Arabic & English support No cloud dependency No external data transfer Local model execution

Solution

I designed a fully offline hybrid RAG pipeline combining ChromaDB vector retrieval with BM25 keyword search, answered by a locally-hosted Aya-Expanse 8B model — semantic search catches paraphrased questions, while keyword search catches exact policy terms, numbers and department names.

Architecture

HR Assistant architecture: Arabic and English HR documents retrieved via ChromaDB and BM25 in parallel, merged, and answered locally by Aya-Expanse 8B
Arabic/English HR Documents
→
Hybrid Retrieval (ChromaDB + BM25)
→
Local LLM (Aya-Expanse 8B)
→
Employee Answer

Your Contribution

Designed
the offline hybrid RAG architecture.
Built
Arabic and English document workflows.
Integrated
ChromaDB and BM25 into one hybrid retrieval layer.
Deployed
Aya-Expanse 8B locally via Ollama, with zero cloud dependency.

Technology Stack

Python ChromaDB BM25 Aya-Expanse 8B Ollama Arabic NLP

Results

Bilingual
HR-policy access
Zero-cloud
fully local execution
Zero-exposure
of internal HR documents

Challenges & Trade-Offs

Semantic vs. keyword retrieval, privacy vs. capability

Combining vector and keyword search trades some system simplicity for retrieval coverage. Running a local 8B model instead of a larger hosted one trades some raw capability for a strict no-cloud, no-external-exposure guarantee on HR data.