Frequently repeated inquiries about personnel,
rooms, welfare services and procedures are handled manually
in many university laboratories, which results in slow and
inconsistent information access. In this paper, a text based
question answering system for the 5G Wireless Innovation Center
of King Mongkut’s University of Technology North Bangkok is
presented. A domain specific dataset of 621 validated question
and answer pairs was constructed through a documented pipeline
of collection, cleaning, deduplication, privacy filtering and expert
validation. A DeepSeek R1 Distill Qwen 7B style component is
adapted through Low Rank Adaptation under 4 bit quantization
and is evaluated with a reproducible simulation harness
mirroring the local deployment pipeline. A confidence gated
hybrid inference method, denoted CGH-QA, is proposed in which
the fine tuned component, a BM25 retriever and a typo robust
character n gram dense retriever each propose an answer with
calibrated confidence, agreeing answers are fused with learned
weights and out of domain questions are refused by a calibrated
gate instead of being hallucinated. The system is evaluated on
held out, paraphrased, trained fact, misspelled, vague and out
of domain probes against five baselines using lexical, sequence
based and semantic proxy metrics. A held out sequence similarity
of 0.813 and a paraphrased sequence similarity of 0.752 were
achieved by the hybrid while all out of domain probes were
refused. The results indicate that local fine tuning and retrieval
can be combined through validation calibrated gating so that
institutional knowledge access is improved and unsupported
answers are reduced.

models,question answering,chatbot,LoRA fine tuning,DeepSeek,hybrid retrieval,confidence calibration,local deployment,dataset construction