Vector examples

← All examples

Start by writing an embedding, then choose the distance metric recommended by the model that produced it. The maintenance recipes cover index coverage, repair and compression.

Try a complete small exampleCopy Anchor Link

This uses three made-up numbers to check the write-and-search flow. It does not demonstrate the meaning of text. Use the tenant-scoped connection from the connection notes, and keep learning.sample separate from collections using a different dimension or metric.

First save one vector:

curl --fail-with-body -X POST "https://$OC_HOST/v1/tenants/$OC_TENANT/vector/learning.sample/put" \
  -H "Authorization: Bearer $OC_TOKEN" -H "Content-Type: application/json" \
  -d '{"id":"sample-1","embedding":[1.0,0.0,0.0],"dim":3,"metric":"cosine"}'

Then search with the same three values:

curl --fail-with-body -X POST "https://$OC_HOST/v1/tenants/$OC_TENANT/vector/learning.sample/topk" \
  -H "Authorization: Bearer $OC_TOKEN" -H "Content-Type: application/json" \
  -d '{"query":[1.0,0.0,0.0],"k":1,"dim":3,"metric":"cosine"}'

Look for sample-1 in the result. k: 1 asks for at most one match. The dimension and metric agree in both requests. Next, use a real embedding from your chosen model; the model supplies the numbers, while the database stores and compares them.

Search by similarityCopy Anchor Link

• Store one embedding — An embedding is a list of numbers that represents an item for similarity search.
• Attach filters to an embedding — Store small metadata fields beside an embedding so a search can be restricted to matching items.
• Find similar vector directions — Cosine similarity compares vector direction rather than magnitude.
• Find the nearest points — L2 is Euclidean distance: the straight-line distance between vectors.
• Search by dot product — Dot product compares both direction and magnitude.
• Compare speed and retrieval quality — Fast mode searches a narrower part of the index.
• Search within matching metadata — Combine similarity with exact metadata filters.
• Remove vectors from an index — Delete embeddings that are no longer needed.
• Group vectors for IVF search — IVF divides vectors into groups around representative points called centroids.
• Check IVF coverage and balance — Inspect whether stored vectors are covered by the index and whether one partition contains too many entries.
• Check and repair the search graph — HNSW links nearby vectors in a graph.
• Build a compressed vector index — IVF-PQ represents vector groups with compact codes.
• Combine dense and sparse matches — Dense embeddings capture similarity.

Before adapting a recipeCopy Anchor Link

Use a test database, load the named sample data, and check the expected result. Read each page’s limits before changing the query. See connection and SDK notes.