Agentic AI for Conversational Flight Search: Design and Evaluation of a Hybrid React and GPT-4o Framework

An affordable flight search remains a fragmented and cognitively demanding task. Travelers navigate several airline and online travel agency sites, compare fares that change by the minute, and re-enter the same personal details repeatedly. Rule-based chatbots have addressed this only partially, because they demand structured input and handle ambiguity poorly. This paper presents an agentic flight-search system that accepts natural-language queries, extracts and validates travel parameters, retrieves live fares, and returns ranked itineraries with explicit decision cues. The system pairs a React front end with a Node.js and Express back end. GPT-4o is invoked only for intent classification, entity extraction, clarification phrasing, and result summarization; all control flow, validation and ranking remain deterministic application code. The control design is model-agnostic; GPT-4o was chosen because it supports constrained structured outputs. Live fares are obtained through the Google Flights engine of SerpAPI. The system was evaluated on 30 scripted conversations spanning six input categories, with five trials per category, and on a usability study with ten participants. Twenty-eight of the 30 trials succeeded (93.3%). Vague queries were resolved in a mean of 2.0 turns through a single consolidated clarification, and both correction pathways preserved all unamended fields. Both observed failures and one flagged misclassification arose in language-model stages; the deterministic core contributed no errors. Participants rated the system 4.24 of 5 on average, with ease of use highest (4.43) and explanation of recommendations lowest (3.70). The results indicate that a hybrid design, in which a language model handles interpretation and ordinary code handles control, can deliver reliable, low-effort flight search while keeping failures localizable. Recognition of unrecognized parameter values is identified as the principal reliability gap.