Hey Product Hunt, I'm Vu from Mostly Right! 😊 I built Mojirama to see how far search can go when it sits on a good dataset. It covers 1,260,645 movies and TV shows going back to 1874, and you search it the way you would describe a film to a friend. Things to try: Things to try: - a plot: boy befriends alien (40 ms) - an actor: tom hanks (40 ms) - a director: tarantino (2 ms) - a character: darth vader (4 ms) - a studio: a24 (2 ms) - something similar: movies like interstellar (42 ms) - filters said out loud: korean thrillers (5 ms) - an original title: la casa de papel (10 ms) - a typo: game of thornes (28 ms) - emoji: 👽🚀 (1 ms) How it works. There is no model in the loop while you type. Every keystroke runs a search in memory: a keyword index that corrects typos from the catalogue's own titles and names, and an embedding index for anything that reads like a description. That is why the first results arrive in a few milliseconds. The query goes to Clef (https://www.cloudflare.com/resou...) half a second after you stop typing. Clef answers a few dozen multiple-choice questions about the query: what job each word does (topic, person, studio, filter, sort order, mood or filler), a yes or no for each genre, the decade, the language, movie or TV, and the order. Answers below 60% confidence are ignored. The rest become a search plan, which is a set of filters, an order and the topic words left to search. The planned results replace the first ones when they differ. The dataset is 16 sources joined into one table: TMDB for the catalogue, cast, crew and ratings, with Wikidata and Wikipedia on top. Every title is one row with no duplicates and the dataset refreshes every week. We also used Jev (https://typesafe.ai/blog/introdu...) to give each title tags and up to three emojis (Clef wasn't available at the time), which is what emoji search and the guessing game behind the "don't click here" button run on. It runs on Google Cloud Run. Each instance has 8 vCPUs and 32 GB and holds the whole catalogue and both indexes in memory. The dataset is public if you want to build on it: https://mostlyright.md/datasets/... Post any search it gets wrong, with what you expected to see.

Mojirama

Fast movie search that understands what you mean

Fast movie search that understands what you mean