After an embedding model upgrade, some Cosmos DB ordinance items still have null vector fields. What job should run?
Select an answer to reveal the explanation.
Short Explanation
Null vectors are empty parking spots after the model upgrade. You can run a backfill job to find those Cosmos DB rows and write their missing embeddings, so RAG can retrieve those ordinance chunks again.
Full Explanation
After an embedding model upgrade, some Cosmos DB ordinance items may still have null vector fields and need deliberate repair. A backfill job that queries for null vector fields and writes the missing embeddings restores a complete municipal corpus, allowing RAG to retrieve upgraded ordinance chunks. Removing ordinance records with null vectors deletes source content instead of repairing the missing embeddings, reducing the material available for grounded answers. Ignoring null vectors and allowing RAG to fail silently leaves gaps in ordinance coverage without making the problem visible to operators. Copying random vectors from unrelated departments creates false similarity relationships and corrupts retrieval integrity. Exam caveat: model upgrades can change vector dimensions or semantics, so the backfill must generate embeddings with the upgraded model rather than reuse vectors from unrelated content. As a practical check, query Cosmos DB for null vector fields, run the backfill, and sample RAG results for ordinance chunks that previously had missing embeddings before declaring the upgrade complete.