Real-time patient guidance is vital in healthcare. This paper presents Chikitsak, an AI-enabled medical chatbot combining BioBERT for semantic retrieval and GPT-2 for fluent, context-aware response generation. The system uses fine-tuned BioBERT embeddings with FAISS indexing and a tag-aware negative sampling strategy, achieving 92% retrieval accuracy. Trained on 25,000 annotated doctor–patient interactions (70% from North American/European sources), it achieves BLEU-4 = 0.42, ROUGE-L = 0.51, and BERTScore-F1 = 0.73. While dataset bias is acknowledged, potential impacts on minority populations are analyzed, and mitigation strategies such as multilingual expansion, bias audits, and rare-condition oversampling are proposed. Medical experts rated 85% of responses as clinically appropriate and 72% as sufficiently detailed, surpassing baseline GPT-2 by 15%. The hallucination rate is reduced to 5%, outperforming ClinicalBERT (10%) and baseline GPT-2 (12%). Low-confidence outputs revert to ranked Q&A references for safety. Ethical considerations, including patient data privacy, explainability, and regulatory compliance (GDPR/HIPAA), are addressed. Limitations include reliance on simulated evaluation and the absence of real-world usability testing, planned for future work. The open-source design will be shared for reproducibility, with future improvements targeting advanced LLM integration, knowledge graphs, multilingual capability, and clinician-in-the-loop learning.