University deployment
DGX Chatbot: Hybrid RAG Assistant on NVIDIA DGX A100
A locally-hosted hybrid Retrieval-Augmented Generation chatbot serving 2,000+ postgraduate students at Universitas Gunadarma, built on a custom Indonesian retrieval pipeline with Qdrant, BM25, Sastrawi, and Reciprocal Rank Fusion.
- students served
- 2,000+postgraduate cohort
- hybrid retrieval
- 4-wayQdrant + BM25 + Sastrawi + RRF
- vLLM serving
- LocalQwen3-8B + Qwen2.5-VL multimodal
01Problem
Postgraduate students lacked a unified institutional assistant that could answer curriculum, research, and admin questions in Bahasa Indonesia while keeping sensitive data on-premise.
02Approach
- Built a hybrid Indonesian retrieval pipeline combining Qdrant vector search, BM25 lexical search, Sastrawi stemming, and Reciprocal Rank Fusion.
- Served fine-tuned Qwen3-8B and Qwen2.5-VL models locally via vLLM on NVIDIA DGX A100.
- Designed the orchestration to support multimodal ingestion (text + image) for research queries.
03Contribution
- Designed and implemented the hybrid retrieval pipeline.
- Deployed and tuned the local vLLM serving stack on the DGX A100 cluster.
- Coordinated rollout with the Universitas Gunadarma HPC programme.
04Outcome
- Serving 2,000+ postgraduate students with a fully local Indonesian-language assistant.
- Hybrid retrieval combines dense semantic, sparse lexical, and Indonesian linguistic signals for high-precision answers.
- Documented a reproducible template for local LLM serving on DGX infrastructure.
05Limitations
- The chat interface is deployed on the internal Gunadarma network and is not publicly accessible.
- Latency is delivered at batch cost, not real-time streaming, due to on-prem GPU budget.