I originally built Agora as a side project to find a pair of red shoes for my wife. The MVP search took 5 - 10 seconds to return results and the app looked like it was designed in MS Paint (by yours truly). https://web.archive.org/web/20231214000400/https://www.searc...

Here are a few technical challenges that we've been working on recently:

1. Figured out how to return results in under 300ms. We are using Mongo as our primary database and a self-hosted version of Meilisearch for search on a high memory Digital Ocean server. The app is built with Remix, specifically to optimize for pre-fetching of data. To power the AI search, we convert our products into both text and image embeddings. Text embeddings with Cohere and image embeddings with Clip. We are now working on further server optimizations and data sharding to get the search speed under 200ms.

2. Expanded the crawler to index stores built with Shopify, WooCommerce, and major retailers like Walmart and Target. Because e-commerce data is constantly updated by store owners, we also built a "recrawler" that updates all stores on Agora every 24 hours. Then when a user clicks on a product, we do a live check of the variants and stock. This ensures that product data is always up to date.

3. Added deeper e-commerce query detection. We noticed that some searches are objective keywords like "camping tent" and others are subjective queries like "what is the best camping tent under $500". When a user searches on Agora, we pass the query through a text classifier to determine if it's subjective or objective, then change the AI tooling and UI accordingly. We served 2.1 million search queries in 2024 and are now using this data to get even more precise about query detection.

I'd love your feedback. And let me know if you have ideas on how to make the search faster.