Caliby targets embedded vector search

The strongest idea here is not the benchmark flex; it’s the packaging. If Caliby holds up in real workloads, it gives agent builders a DuckDB-style local persistence layer for retrieval.

• –In-process storage removes the ops overhead of pgvector, Qdrant, or Milvus for apps that only need local or embedded memory.
• –The pgvector and FAISS comparisons are useful signals, but they still need independent, apples-to-apples benchmarking before anyone should treat them as settled.
• –Text + vector + metadata in one API fits RAG and agent-memory workflows better than stitching together a vector store and a separate document DB.
• –The market is crowded, so the project will likely win or lose on docs, ergonomics, and persistence reliability more than on raw ANN throughput.