BACKGROUND Timely, accurate, and understandable health information is critical for individuals with spinal cord injury (SCI) during rehabilitation, yet shortened stays, staff shortages, and information complexity leave many patients and caregivers underprepared at discharge. Large Language Model (LLM) chatbots show promise for delivering accessible patient education in SCI rehabilitation and typically augmented with domain-specific knowledge bases (KBs) to enhance accuracy and response quality, yet how KB source selection influences these systems remains understudied. OBJECTIVE This study investigates how different knowledge sources—curated website articles, journal articles, and their combination, as well as a no-KB baseline—affect the response quality of an LLM-driven SCI patient education chatbot. By identifying which configurations produce responses that are both clinically accurate and genuinely actionable, this study aims to guide KB development that support patient self-management and reducing preventable complications. METHODS Using a Retrieval-Augmented Generation (RAG) framework, we developed chatbots with GPT-4o, Llama 3.3-70B, and Claude 3.5 Sonnet across four KB configurations and evaluated their responses against a curated clinical question set. Response quality was assessed across five dimensions: semantic similarity to clinician-provided answers, understandability, actionability, factuality, and comprehensiveness. RESULTS The curated website-based KB achieved the highest performance overall, provided the highest semantic similarity, understandability and actionability (score = 73.29%, κ = 0.459, AC1 = 0.652), and zero non-factual responses. The vanilla model performed comparably in semantic similarity but produced one non-factual response and lower inter-rater agreement (score = 72.56%, κ = 0.325, AC1 = 0.553), indicating that relying solely on the internal model knowledge yields less reliable outputs. The combined KB showed no improvement over the curated websites KB or vanilla model from merging both sources (score = 72.46%, κ = 0.496, AC1 = 0.665). Finally, the journal article–based KB generated the most comprehensive responses but scored lowest overall (score = 65.48%, κ = 0.449, AC1 = 0.546), likely due to technical language and insufficient step-by-step guidance, and also produced one non-factual response. CONCLUSIONS Knowledge source selection substantially shapes the response quality of LLM-driven chatbot for SCI patient education. Among the configuration tested, a KB constructed from curated websites delivered the most balanced and reliable performance across five dimensions.