Cartum / Horoshop — Knowledge Assistant (RAG)
Aug 2024 - Dec 2024
Maintained and iterated in production through 2025–2026.
Building a production Retrieval-Augmented Generation (RAG) system to automate internal support and onboarding for an e-commerce platform.
The Challenge
Cartum (formerly Horoshop) is a major e-commerce platform with extensive internal documentation, specifications, and support guidelines. New employees faced a steep learning curve, and the support team spent significant time searching for answers across fragmented knowledge bases.
The goal was to create an internal AI assistant capable of answering complex questions accurately, grounded in the company's proprietary data with strong anti-hallucination guardrails.
Solution & Architecture
I designed and implemented a full RAG pipeline. The system ingests heterogeneous documents (Markdown, Confluence, PDFs), chunks them semantically, and stores the embeddings in a localized Vector Database.
When a user asks a question, the system retrieves the most relevant chunks, injects them into a carefully crafted prompt, and uses a large language model to synthesize a precise answer.
- Data ingestion pipeline for automated syncing with internal wikis.
- Semantic chunking strategy optimized for technical documentation.
- Vector search using advanced indexing for low-latency retrieval.
- Prompt design + guardrails to keep answers grounded in retrieved context.
User Interface
The assistant is accessed via a clean, chat-based UI integrated directly into the internal tools suite, designed for speed and clarity.
Results
The Knowledge Assistant significantly reduced onboarding time for new hires. The support team reported a 40% decrease in time spent searching for internal processes and technical specifications, allowing them to focus on complex customer issues.