Vector models by industry
Four places where matching on meaning beats matching on words. Each one pairs the vector search with the filter or the keyword query that makes it usable.
Retail: search that understands the ask
A shopper types something no keyword index will match, and the catalog still has the right answer.
Nearest products, within the shopper's constraints
The price filter is part of the search, so the ten results you get back are ten results you can show.
POST /v1/tenants/:t/vector/shop.products/topk
{ "query": [...], "k": 10, "dim": 768, "metric": "cosine" }Pair it with keywords
Vectors are strong on meaning and weak on exact terms - a model number, a brand. Run both and merge; the full-text side is the search API.
{ "query": { "match": { "name": "carbon marathon" } } }Support: deflect the ticket before it is filed
The customer describes a problem in their words. The answer already exists, written in someone else's.
Nearest resolved tickets
Embed the draft as it is typed and search resolved history; a good match becomes a suggestion instead of a ticket.
POST /v1/tenants/:t/vector/support.tickets/topkNearest-neighbour finds the ones that mean the same thing; the suggester fixes a misspelled product name before either search runs.
Legal and research: retrieval for grounded answers
The retrieval half of RAG. The quality ceiling of the answer is set here, not in the model.
The passages worth sending
Chunk the documents, embed the chunks, and retrieve the few that actually bear on the question.
POST /v1/tenants/:t/vector/kb.chunks/topkStore the source id and offset on the chunk row. Retrieval then returns something a reader can verify, which is the difference between an answer and a claim. RAG guide.
Media: finding the near-duplicates
Not identical files, which hashing finds, but the same thing re-encoded, cropped or re-uploaded.
Neighbours above a similarity threshold
Search each new asset against the library and treat anything very close as a candidate duplicate for review.
POST /v1/tenants/:t/vector/media.assets/topkWhat these have in common
- The vector is a column. It is written with the row, so there is no embedding store to sync and nothing to fall out of date.
- Filters belong inside the search. Filtering after a top-k is how you ask for ten results and show three.
- Meaning and words are different questions. The strongest results usually come from running both and merging.