Vector search and embeddings

A vector collection holds items (text plus your own metadata) as embeddings. Searching it with a question returns the items closest in meaning, not just matching words — the core of semantic search, recommendations and RAG.

#Index and search

await sovaria.vector.upsert('help-articles', [
  { id: 'a1', text: 'How to reset your password', metadata: { lang: 'en' } },
  { id: 'a2', text: 'Exporting invoices as PDF', metadata: { lang: 'en' } },
])

const hits = await sovaria.vector.search('help-articles', 'I forgot my login', {
  topK: 3,
  metadataFilter: { lang: 'en' },
})

#Retrieval-augmented generation

1. 1

   Index your content

   Upsert each document (or each paragraph of long ones) with an id and metadata such as lang, product or tenant.

2. 2

   Search with the user's question

   Take the top 3–5 hits, filtered by metadata the user may see.

3. 3

   Answer with context

   Send the hits to your model as context. Doing this in an edge function keeps your model key on the server — see Functions.

#Manage collections

vector.listCollections(), listRecords(collection), deleteRecord(collection, id) and deleteCollection(collection) (Python: vector_list_collections, vector_list_records, vector_delete_record, vector_delete_collection). Full signatures: JavaScript: Vectors and embeddings and Python: Vectors and embeddings.