传统出行选择模型依赖大规模结构化数据且在行为解释与预测性能之间往往难以兼顾，大语言模型为实现类人的认知决策提供了新路径。然而，通用大语言模型因缺乏特定领域的行为先验，直接推理常面临行为错位挑战。为此，本文提出一种融合检索增强与行为偏好校准的预测框架（ReCAP），该框架遵循检索－归纳－演绎的范式。首先，设计基于语义向量空间的平衡检索策略，从异质性样本中精准匹配同质参考群体，缓解传统检索面临的样本类别不平衡与长尾分布偏差；其次，利用大语言模型的少样本归纳推理能力，从参考群体中提炼出包含时间、成本及舒适度敏感性的结构化行为偏好画像，并将其作为显式先验约束注入决策推理过程，实现通用模型向领域专家的动态校准；最后，基于该画像在具体情境中进行决策推理完成预测。瑞士地铁数据集上的结果表明：所提框架在微观个体分类与宏观分担率拟合两方面均表现出较优性能，相较于多项Logit及随机森林等基准模型，预测准确率与加权F1值平均提升10%以上，预测分布与真实分布之间的詹森-香农散度有效降低，缓解了对非主导类别的识别偏差。同时，通过自然语言生成的推理链清晰展示了决策背后的权衡逻辑，显著增强了行为预测的可解释性与透明度。

Traditional travel choice models rely on large-scale structured data and often struggle to balance behavioral interpretability with predictive accuracy. Large Language Models (LLMs) offer a promising alternative for human-like cognitive decision-making. However, general LLMs often suffer from behavioral misalignment during direct inference due to a lack of domain-specific priors. To address this, this paper proposes ReCAP, a prediction framework integrating retrieval augmentation with behavioral preference calibration. The framework follows a Retrieve-Induce-Deduce paradigm involving three stages. First, a balanced retrieval strategy based on semantic vector space matches homogeneous reference groups. This approach effectively mitigates sample class imbalance and long-tail distribution biases found in traditional retrieval. Second, the framework utilizes the few-shot capabilities of LLMs to distill structured behavioral preference profiles covering sensitivities to time, cost, and comfort. These profiles serve as explicit prior constraints to dynamically calibrate the general model into a domain expert.  Finally, the model makes context-specific predictions based on these profiles. Experiments on the Swissmetro dataset demonstrate that ReCAP excels in both micro-level individual choice classification and macro-level mode share fitting. Compared to baselines like multinomial logit (MNL) and random forest (RF), it improves accuracy and the weighted F1 score by over 10% on average. Furthermore, it significantly reduces the Jensen-Shannon divergence (JSD) between predicted and true distributions, alleviating identification bias against non-dominant categories. Additionally, the generated natural language reasoning chains elucidate the trade-offs behind decisions, enhancing the interpretability and transparency of behavioral predictions.