Semantic search and RAG with Cloudflare Vectorize V2, supporting 5M vectors, 31ms latency, and async mutations.

• Vectorize V2 (GA September 2024) brings 25× capacity increase (200K → 5M vectors), 18× faster queries (549ms → 31ms), and async mutations with breaking API changes including returnMetadata boolean → string enum
• Metadata filtering with 10 indexes per collection, range operators ($gte, $lt, etc.), and cardinality-aware query optimization for RAG and semantic search workflows
• Batch insert optimization: 5000-vector batches deliver 18× throughput improvement; individual inserts impractical at scale
• Native Workers AI integration for embeddings (@cf/baai/bge-base-en-v1.5, 768 dims) plus OpenAI support; dimension limit of 1536 requires reduction for larger models
• 14 common errors prevented: dimension mismatches, metadata index timing, V2 async mutation handling, TypeScript type gaps, Windows dev registry issues, and topK conditional limits