Worked on the meta-llama/llama-stack repository to enhance data ingestion reliability across multiple vector database providers. Focused on backend improvements using Python, the work unified upsert semantics for Milvus, ChromaDB, and SQLite-vec, addressing issues of data duplication and silent re-ingestion. The approach involved updating the VectorIO protocol and refining EmbeddingIndex documentation to clearly define the upsert contract, ensuring consistent behavior across providers such as PGVector, Qdrant, Elasticsearch, OCI 26AI, and Infinispan. This effort prioritized data integrity and developer experience, aligning ingestion logic and documentation to support robust, cross-provider vector data workflows in API-driven environments.