Source: vespa¶

Verification: MOCK ONLY. The mock implements the visit API of https://docs.vespa.ai/en/reference/document-v1-api-reference.html (continuation tokens, wantedDocumentCount, format.tensors=short, bearer auth). The destination is a real hms-server.

export VESPA_TOKEN=...
hms-import vespa --url https://app.example.vespa-app.cloud --vespa-namespace shop --doctype item --vector-field embedding --hms-url http://127.0.0.1:8080

Endpoint: GET /document/v1/{namespace}/{doctype}/docid?wantedDocumentCount=N&continuation=...&format.tensors=short[&cluster=..][&selection=..]. The response is {"pathId", "documents": [{"id", "fields"}], "continuation"?, "documentCount"}; the walk ends when continuation is absent.

Mapping: the user-specified part of id:ns:type:key:user is the id; the text field and timestamp fields are --text-field, --timestamp-field, --updated-field; all other fields are metadata, plus vespa_namespace, vespa_doctype and the g=/n= locality (vespa_group/vespa_number). --namespace-from none|namespace|doctype|group picks the HMS namespace.

Tensors: one-dimensional dense tensors in the values list, cells list (indexed addresses 0..n-1), a bare list, or the int8 hex string form. Mapped (sparse), multi-dimensional and mixed (blocks) tensors are not vectors: such records are rejected with the reason, never imported without their vector. Several tensor fields require --vector-field.

Auth: bearer token ($VESPA_TOKEN) or mTLS (--cert, --key, --ca-cert). Resume re-reads and hash-verifies the committed prefix (continuation tokens are per page and are not stable across runs).