peterdm

Are you seeing a blueprint start to emerge for a standardized way to build and deploy search indices in the context of applications that need vector-space features? (E.g. if you start with ANNOY, you get kNN but then how do you add in the ability to refine, filter, rescore, sort with text, etc?)

Upvoted. This is how Postgres supports "fuzzy searching" which helps with misspellings. https://www.rdegges.com/2013/easy-fuzzy-text-searching-with-...

Completion response-time will be slower than Solr, Elasticsearch, Algolia, etc... but if you're already running Postgres, this may be the fastest to deliver for you.

For e-commerce search, personalization with Elasticsearch takes a similar level of effort as with Solr. Don't re-platform under the impression it will make personalization easier. It still takes data collection and experimentation but can be accomplished on Elasticsearch. Feel free to contact me if you have questions.

Feature vectors do tend to get incorporated in relevance tuning (regardless of the engine), but from what I've heard of Vespa, features (and ML in general) are first-class citizens, whereas with Elasticsearch and Solr, text statistics are your first-class citizens, and you're adding in additional features and integrating ML at the periphery.

Until about 2014-2015, many large companies wouldn't have looked twice at Elasticsearch. The companies you list using Solr have been invested in search for 10 years or longer (predating Elasticsearch), and may have high switching costs.