little big brain
Product / Search

Semantic search on your graph.
Managed embeddings. Checked hits.

Declare which classes to embed. We run the model and keep the vectors current as records change.

Sign upRead the guide
[01] In the console

One declaration per class.

Pick the class and the fields to embed. The class is searchable once its vectors are built.

A search for login problems after SSO: issues, organizations and deals ranked by Jev, each with its scoreA search for login problems after SSO: issues, organizations and deals ranked by Jev, each with its score
The Status page: written, indexed and searchable at commit 7, with three classes current for searchThe Status page: written, indexed and searchable at commit 7, with three classes current for search
No vector store to run.

We run the embedding model and embed changed records again.

Checked hits.

Hits are checked against one published commit. Deleted records do not come back.

Reranking.

On BEIR SciFact, the Jev reranker lifted nDCG@10 from 0.713 to 0.797.

A SPARQL query with search:similarTo, joined with each issue's customer, and five scored rowsA SPARQL query with search:similarTo, joined with each issue's customer, and five scored rows
The nearest issues, each joined with its customer.
[02] SPARQL

Vector search inside a query.
Meaning and links together.

Write a search:similarTo pattern in a query. Join the hits with any graph pattern. The engine selects the plan.

  • Plans: nearest, filter_first or search_first.
  • search:score, search:top and search:rerank in the same pattern.
  • The response names the plan it used.
Search by meaning
[03] What is included

What comes with search.

One managed model per graph. You choose what to embed.

POST /v1/search

Ranked records for a text. Filter by class, with its subclasses, and by link.

search:similarTo

Vector search as a triple pattern, joined with the rest of the query.

rerank

Turn reranking on for a graph or for one search.

lag_commits

Responses say how far the embeddings trail the graph.

model

Switch the embedding model of a graph in one call.

preview

See the text and the cost of embedding a class before you declare it.

[04] How it works

Declare it once. We keep it current.

Step 01

Declare.

Name a class and the fields to embed.

Step 02

Embed.

We build the text and embed it. Changed records are embedded again.

Step 03

Search.

Ask in words and get the nearest records.

[05] FAQ

Questions about search.

Which embedding model do you run?

A managed OpenAI embedding model. A graph has one model, and you can switch it.

Do I pay for embeddings?

Search by meaning is included on Free and on paid plans.

How large are the vectors?

Two measured catalogs took 7.3 KB and 7.8 KB per record at 1,536 dimensions.

What does reranking cost in time?

It adds about 0.3 s to a search. After 2.5 s, the search falls back to its own order.

Next step

Try it on your data.

Sign upRead the docs