DynamoDB Native Vector Search
Hands-on evaluation of native vector search in DynamoDB on a product catalog, testing semantic recall, filters, cross-language queries and score calibration.
Architecture
CDK stack with an on-demand table and a custom-resource Lambda that creates the 1024-dimension cosine vector index, since CloudFormation doesn't support it yet. Embeddings come from Titan v2 on Bedrock. A dataset built with lexical and semantic distractors, near duplicates and Spanish queries shows where the index helps and where it fails.
Highlights
- Custom resource for a feature CloudFormation lacks
- Dataset designed to expose edge cases
- Deploy and teardown for a few cents