Source: pinecone¶

Verification: MOCK ONLY, but driven through the real pinecone client library (Index(host=...)) against a mock of the data-plane API. Pinecone Local speaks the same protocol (--host http://localhost:5081). The destination is a real hms-server.

export PINECONE_API_KEY=...
hms-import pinecone --host https://my-index-xxxx.svc.region.pinecone.io --hms-url http://127.0.0.1:8080

Calls (https://docs.pinecone.io/reference/api/2025-10/data-plane/): POST /describe_index_stats (namespaces and counts), GET /vectors/list?namespace&limit&paginationToken&prefix (ids in sorted order, pagination.next), GET /vectors/fetch?ids=..&ids=..&namespace (values, metadata, sparseValues). Header Api-Key. Serverless and Pinecone Local only: pod-based indexes have no list endpoint (export through your own id list, or re-ingest from source data). --index-name resolves the host through the control plane instead of --host.

Mapping: every namespace (or the ones given with --namespace) becomes hms_namespace; the default namespace ("" or __default__) keeps bare ids, others get namespace::id, so the same id in two namespaces stays distinct. The text is a metadata key (--text-field, else text, content, chunk_text, ...) and is moved out of the metadata; --timestamp-field/--updated-field name metadata keys holding times. Sparse values are not dropped: they are kept as pinecone_sparse: {indices, values} in metadata (HMS has no sparse vector slot). An id that is listed but gone by fetch time is rejected with a reason.

Resume is seekable: the cursor is [namespace, page token, offset in page].