The technology stack is outlined, including Kotlin, Spring AI, and pgvector, along with the process for ingesting hotel data and creating embeddings.
A simple vector search fails to respect hard constraints like price and amenities, leading to irrelevant results and a low evaluation score.
Manually separating a query into a residual semantic part and hard SQL filters significantly improves the relevance and accuracy of search results.
An LLM is prompted to parse a natural language query, successfully extracting structured filters like price and amenities from the unstructured text.
Instead of a single LLM, specialized tools like DJL for amenity matching and Duckling for price extraction offer more robust and efficient solutions.
The application architecture uses a composer service with focused, independent extractors for price, amenities, and residual intent, leveraging Kotlin's concurrency features.
The completed application successfully processes complex queries by decomposing them into hard filters and a residual vector search, delivering highly relevant results.
The talk concludes with key takeaways, emphasizing using the right tool for each job, filtering before vector search, and decomposing queries into modular components.
Jobs that call for the skills explored in this talk.