Supply your own embeddings.
Generate vectors with your chosen model and store them under your record IDs. Send a query vector to retrieve the nearest matches.
Your model·Vectors·Matches
Use the vector API Vector search
Find the meaning.
Keep the context.
Find relevant content even when the wording differs. Search embeddings, retrieve similar matches and connect every result to the records your application needs.
Meaning has neighbours
Your queryNearby matches
MatchesRecord IDsSimilarity scores
More ways to find
what matters.
Give search and discovery a sense of similarity, alongside the structured data you already use.
01
Search by meaning
Help people find relevant documents and passages even when they use different words.
02
Discover similar records
Use the similarity captured by your embedding model to surface related content or items.
03
Retrieve useful context
Return record IDs with your matches, then bring their source data into an application or AI workflow.
Prepare once. Search by similarity.
Bring your vectors.
Or start with your text.
Choose the ingestion route that fits your application and embedding provider.
Use a saved embedder.
Choose the text columns to embed and run a job on a dedicated configuration. Search the prepared table with a plain-text question using the same embedder.
Table text·Embedder·Matches
Prepare an embedding run How do embeddings stay current?
Schedule embedding runs for new rows and, optionally, changed text. Without a schedule, start a run when you want vectors refreshed. When supplying vectors yourself, your application manages their updates.
Use a consistent model, dimension and distance metric for stored and query vectors. Saved embedders require your provider endpoint and credentials.
Search that fits your workload
Balance memory,
search effort and recall.
Choose an index with measurements from your own dataset and the results your users need.
Neighbour graph
HNSW
A starting point for approximate similarity search. Tune search effort and measure recall on your own data.
Partitioned search
IVF
Organize vectors into cells and choose how many to inspect for each query.
Compressed vectors
IVF-PQ
Evaluate a compressed representation when memory matters. Measure how compression affects your results.
Filtering behaviour depends on the index. IVF-PQ does not support metadata filtering; review the index and filter reference before choosing.
Dense + sparse vectors
Two rankings. One candidate list.
Store both representations under matching IDs and combine their rankings with reciprocal rank fusion. Retrieve the original records to add context.
Explore hybrid vector retrievalDenseACB
SparseBAD
CombinedABCD
Illustrative ordering Hybrid vector retrieval combines dense and sparse vectors. Full-text BM25 is a separate search path.
Put similarity to work.
Choose an entry point, then connect the returned IDs to the rest of your data.
Build with OriginChainDB