To diagnose and resolve configuration, quota, payload size, and permission
issues when you generate automatic vector embeddings in AlloyDB for PostgreSQL, use
the following troubleshooting resolutions.

| Issue | Cause | Resolution |
|---|---|---|
| Modifying existing auto embedding configuration | Changing the model, source column, or refresh mode requires clearing the old configuration. | Call `ai.drop_embedding_config` to remove existing settings, then call `ai.initialize_embeddings` again with the new configuration. |
| Disabling transactional mode | Transactional mode is enabled using `ai.initialize_embeddings`. | Call `ai.drop_embedding_config` to remove the metadata and the trigger. |
| Slow inserts and updates in transactional mode | Embeddings are generated serially within the same transaction, so model latency impacts transaction time. | Check model latency and consider if transactional mode is necessary for your use case. |
| Slow auto embedding generation | Might be due to model request quota limits being exceeded. | Examine the model request quota for your embedding model. Adjust the `batch_size`. Contact [support](https://docs.cloud.google.com/support) if the issue persists. |
| `Error: 'AutoEmbeddingGeneration: Request size is greater than 4MB'` | The request size for a batch exceeds the maximum supported size of 4 MB. | Reduce the `batch_size` value in your function call and retry. |
| Permission error calling `initialize_embeddings` in transactional mode | Transactional mode creates triggers, which requires table owner privileges. | Ensure that the user calling the function is the table owner or has the owner role. |