Search a Vector Index

A fixed search lets your application page through neighbors without a new search or a complete embedding download. The result retains exact index and source versions, the query, and its candidate budget.

Open one fixed searchCopy Anchor Link

Use a query from the same embedding model and with the same dimensions as the indexed column. Select the source Table and Vector Index from the same snapshot.

snapshot = client.project("demo").branch("main")
frames = snapshot.table("frames")
index = snapshot.vector_index("frame-search")
results = index.search(frames, query, limit=200, probes=32, rerank=2_000)

limit selects the result count. probes and rerank control recall for this search. The search contract defines request limits and filter behavior.

Both the index and the source Table require read access, even when the runtime reuses cached data. Expired credentials and grants for other roots do not produce a search result.

An unbuilt index requires a refresh. A source version mismatch produces an error before search. The client does not switch source versions or download all embeddings.

Read visible columnsCopy Anchor Link

Keep the result handle for the gallery session. Request only the columns for each visible page.

start = min(offset, len(results))
stop = min(start + 25, len(results))
page = results.page(start, stop, columns=["frame_id", "camera", "image"])

Every page retains neighbor order and the three search result columns.

Requested source columns follow these columns. Media payloads remain separate from the page metadata. A page accepts at most 64 MiB of Arrow buffers. The client bounds retained batches before it combines them. The TypeScript client returns null for an empty page range.

Interpret a short resultCopy Anchor Link

The result exposes the effective partition and candidate limits through info. The exhaustion flag distinguishes a complete population from a fixed budget. A new search with different recall controls produces a separate result handle.

The ScanDAL HTTP reference describes the underlying search operation. The Rust SDK guide describes client setup.