A vector database gives you nearest neighbours and leaves the rest — filtering by stock, faceting, access control, analytics — to you. In AACSearch a vector field is just another field of a collection: it is queried through the same search call, narrowed by the same filter_by, and protected by the same scoped keys.
Add a float[] field. Either let the platform embed chosen text fields at index time, or send vectors you computed yourself.
vector_query takes an inline vector, the id of a document to find similar ones, or a base64 image for an image-embedded field.
Combine with filter_by and cap weak matches with distance_threshold, so «similar» never means out of stock or off-category.
An inline vector is the generic case — you embedded the query yourself and want the k nearest documents. Querying by document id turns the same index into a «similar items» block without any embedding call at all. A base64 image against an image-embedded field gives search by photo. All three are one parameter on the ordinary search endpoint.
Pure vector retrieval has no notion of an exact match. Part numbers, SKUs, model codes and brand names should go through keyword or hybrid retrieval, where an identifier either matches or does not. Use vector-only for similarity blocks, image search and queries where no wording overlap can be expected.
Embeddings from your own pipeline index and query like any other field — no lock-in to one model.
Reference a document instead of a vector and get its neighbours — a related-products block in one request.
Upload a photo or pass its URL to the image search endpoint and retrieve the documents closest to it.
k, distance_threshold, ef and flat_search_cutoff are exposed per query, so recall and latency are tuned per use, not per cluster.
New to the term? The glossary explains Vector embeddings from first principles.
{
"q": "*",
"vector_query": "embedding:([], id: sku-1042, k: 12, distance_threshold: 0.35)",
"filter_by": "category:=boots && in_stock:true",
"exclude_fields": "embedding"
}
// -> the 12 nearest in-stock boots to sku-1042Explore more
The best way to evaluate a search platform isn't reading one more comparison table.
Only then make the decision.