Move your vectors
between databases.
Without moving them through us.
Vector Transfer copies dense embeddings, IDs, and record metadata from one vector database to another. You submit and watch the job in a dashboard. A worker you run on your own machine does the reading and writing, so database credentials, vectors, and checkpoints stay in your infrastructure. If an interrupted transfer stops, it resumes from the last acknowledged page instead of starting over.
Your worker, your data path
The worker connects outbound to Polign over HTTPS and reports progress. Vector data does not pass through the control plane. The dashboard sees worker identity, connection names, job settings, status, counts, and audit information, nothing else.
Checkpoints and resume
A checkpoint advances after the destination acknowledges a whole page. Kill the worker, restart it with the same identity, and it continues from local state. Cancel a job and resume it later past the same checkpoint.
Honest about writes
Execution is at least once: a crash between a write and its checkpoint can replay one page, so upserts must be idempotent by ID. Existing destination IDs are replaced. Cancellation does not roll back what was already written.
Private endpoints reachable
Because the worker runs on your network, it can reach databases with no public endpoint. S3 Vectors uses the machine's AWS credentials or workload identity; other providers take credentials at the worker's prompt and never send them to Polign.
Or let Polign run it
For databases with public HTTPS endpoints, choose Polign hosted in the dashboard. Polign stores the connection credentials encrypted and processes the vectors. The trade-off is stated in the guide: the server can decrypt what it stores.
Try it with nothing installed
One binary includes the dashboard, a 1,000-record sample source, and a local JSON destination. It runs on loopback and makes no cloud calls. Download it from the guide and run a transfer in a minute.
Supported databases
All fifteen work as both source and destination. Each has documented settings and limits; the provider catalog lists them.
- Pinecone serverless indexes
- Weaviate collections, with tenants
- Milvus 2.6 and later over REST
- Elasticsearch 8.19 and later
- OpenSearch with basic auth
- pgvector over native PostgreSQL TLS
- Amazon S3 Vectors with workload identity
- Chroma API v2
- Redis HASH records
- MongoDB with native TLS
- FAISS through the included file adapter
- Solr collections and cores
- Qdrant collections
- Turbopuffer namespaces
- polign_db collections
What a transfer does not do
Read this before moving your own data. Transfers leave the source in place and move exactly what is there.
- No re-embedding: vectors are copied as they are, at their source dimension
- No schema creation: prepare the destination for the dimensions, ID type, and metadata first
- No index definitions: index settings are not copied between products
- No metadata conversion: incompatible types fail explicitly instead of being silently changed
- Keep the source still during a transfer and any resume
- Sign in required for the dashboard; the local demo needs no account
Start a transfer.
Sign in, set up a worker, and watch it run.