Bookerty: Search Chatbot

Given the current trend of AI agents covering a larger part of user search queries, I implemented a chatbot alternative search functionality in Bookerty’s web app and backend. Bookerty is a hospitality startup similar to AirBnB. The chatbot uses Vercel’s AI SDK to abstract the AI provider layer and Bookerty’s pre-existing internal search API’s to look for accommodation listings given a set of parameters deduced from the user chat.

The AI agent is given a set of claude SKILLs and some tool calls to carry out search, geocoding, Place of Interest search, directions, weather look-ups, so contextual querries like “I want to rent a 2-bedroom house next to a beach in Chalkidiki, ideally within a mile off a basketball court” can be answered. This helps a rounder user experience in booking accommondation emerge.

The chat endpoint uses SSE and streams responses to the client for speed. The endpoint also uses a custom protocol for accommodations that helps us render listing cards using frontend components rather than just relying on text, meaning we can be infinitely flexible with chat visualizations. Several architecture decisions, including rate limits and tool use limits, restrict excessive AI use from the public website. Hallucinations are also kept at bay by imposing strict rules on the model about information guessing.