Rental Image Search in 2026: Caption Text, Pixel Matching, and Cache Boundaries

The story outlines a two-tier approach to image search for property listings. The first tier involves a text search on captions and metadata, which is fast and cost-effective. The second tier, pixel matching, is used for more complex visual queries, such as finding balconies or duplicate images. This tier is asynchronous, asynchronously indexing the images, and comes at a higher cost due to storage and cache requirements.

The decision to implement this split is based on practical considerations, such as the need for immediate search capabilities and the management of storage and cache costs. The system is designed to be simple and testable, with clear separation of responsibilities and retention rules for different data layers. Each component is treated as a separate column with its own retention rules, allowing for easy updates and revisions without affecting the original data.

The article also emphasizes the importance of maintaining security and privacy in the system. For instance, private photos should remain private, and the system should prevent misuse of the visual search feature. The final implementation includes a visibility filter that ensures only publicly accessible images are returned in search results. Tests are recommended to cover various scenarios, such as duplicate captions, replaced checksums, and private to public transitions.