Why Keyvector
Keyvector is for teams whose data lives in Amazon Keyspaces and who now want to search it by meaning. Typical uses:
- “Find tickets like this one”
- Product or catalog search that understands the question
- Retrieval for an AI assistant (RAG)
- “More like this” recommendations
- Near-duplicate detection
Keyspaces can’t do any of that on its own: it has no vector type and no vector index.
The alternatives, and what they cost
Section titled “The alternatives, and what they cost”| Your other option | What it costs you |
|---|---|
| Migrate to a Cassandra with vector search (DataStax Astra DB, ScyllaDB, self-managed Cassandra 5) | Redoing a database migration you probably just finished, and leaving a managed AWS service |
| Move the data to DynamoDB or OpenSearch | Rewriting your data layer and running two databases during the move |
| Build the sync pipeline yourself | Weeks of work on the parts that are easy to get wrong (deletes, retries, restarts, backfill, composite keys), then owning it on call |
| Wait for AWS to add vectors to Keyspaces | An unknown wait |
What you get instead
Section titled “What you get instead”- No migration. Keyspaces stays your source of truth, and your application keeps writing to it unchanged. Turn on the table’s change stream and the index follows.
- The index stays right. Every insert, update, delete and expiry reaches the index, a failed batch is retried rather than skipped, and a restart picks up where it left off. See Status and testing.
- Results are never stale. A query reads the matching rows back from Keyspaces, so you get what the table holds now.
- Your data stays in your account. It runs on your ECS, under an IAM role that can read one table and write one index. With a self-hosted model your text never leaves your VPC.
- Pick your model. Open Sentence-Transformers models on Hugging Face TEI (fine-tune them, no per-token charges, no request quota), Amazon Bedrock, or any OpenAI-compatible API.
- Cheap to run. One small container per table. Rows are embedded only when their text changes, and vectors live in S3 Vectors, priced for storage rather than always-on servers.
- Open source, Apache-2.0.
When it isn’t the right fit
Section titled “When it isn’t the right fit”- If you’re free to change databases, a database with built-in vector search is simpler than running two systems.
- If you need search results to be transactionally consistent with writes: the index trails the table by seconds.
- If your data isn’t in Keyspaces, this version won’t help yet (see the roadmap).