- By default, every field is considered searchable. Ex: URLs, prices, dates.
- By default, no field is considered filterable. Ex: users will not be able to sort by price.
- Users raised issues related to indexing performance, and best practices with regards to how the index settings may impact the experience
- Limited knowledge of the different index settings lead to user confusion, anger, and sometimes violence
- Rules to detect common field values, such as URLs, UUIDs
- Field name lookups to understand the usage of each field, such as
id, datetimes, and so on - Entropy used to measure the amount of randomness of certain values over the whole dataset
- if we observe a small amount of different values in the entire dataset, the field is probably a "category" (color, genres, ...), and thus should be filterable
- if we observe a large amount of different values, the field is likely to represent unique content, thus should be searchable
- Base library in Rust
- can be shipped on the web (fast client-side validation with WASM)
- can be embedded inside of Meilisearch in the future
- directly embedded in the Cloud 🦭
https://meilisearch-cloud-staging-pvc3tikmm-meili.vercel.app
Users do not need to be aware of new settings and features enable via settings: we will offer them the best experience by default.
| dataset | count | default | generated | Faster by |
|---|---|---|---|---|
| All cities | 1M | 40.02s | 37.38s | 6.6% |
| Hacker News | 1M | 184.51s | 170.99s | 7.33% |
| Movies | 32k | 4.12s | 3.72s | 9.71% |
| Magic The Gathering card game | 74k | 35.71s | 19.91s | 44.25% |
| Twitch messages | 5M | 248.58s | 111.80s | 55.03% |
| Music Brainz | 100k | 83.80s | 35.86s | 57.21% |
- Language detection, in order to download synonyms and stop words accordingly
- Continuous analysis of documents to keep settings as up-to-date as possible
- Provide an explanation for all suggested settings to the user
MrBeast 🫶