Small-scale recommendation
memista powers 'more like this' recommendations: embed items, and search for an item's neighbours to surface similar ones. In-process, best for catalogues under ~100k items.
“More like this” is a nearest-neighbour query over item embeddings. For a small catalogue, memista handles it without a recommendation platform.
The approach
- Embed each item (from its text, tags, or a learned representation) and insert.
- To recommend from an item,
searchwith that item’s vector and return the neighbours (skip the item itself). - To recommend for a user or session, query with a user/session embedding if you can produce one.
Fit
memista is in-process and tested under ~100k items — the right size for a focused catalogue, a docs site’s related-articles widget, or an internal tool. It doesn’t do ranking models, business rules, or online learning; it does the similarity lookup underneath them. For a large, always-on recommender, pair a distributed vector store with a serving layer instead.
Related: run a nearest-neighbour search, use cases.
Frequently Asked Questions
Can I recommend by user vector too?
Yes — if you can produce a user or session embedding, query with it and return the nearest items. memista just does nearest-neighbour over whatever vectors you insert.