fix: Retry logic for when Aurora is in paused state

[axelpina]

Issue #848

Description of changes:

This PR addresses issue [#848] by adding a two-stage retry mechanism whenever the Aurora Serverless v2 cluster is auto-paused (min ACU=0). Both the high-level tool wrapper (knowledge.py) and the low-level retrieval function (vector_search.py) now detect the “resuming after auto-pause” error message, pause for 15s on first failure (to allow the cluster to wake up), then for 45s on second failure, before finally surfacing any remaining exception.

The retries only kick in when we hit the Aurora “resuming after auto-pause” error. The code looks for the exact substring (case-insensitive) "resuming after auto-pause" in the exception message. The other back-end mechanisms remain unchanged.

Background

When using Aurora Serverless as the vector store backing a Bedrock Knowledge Base, the first retrieval after a pause fails with:

The Aurora DB instance arn:… is resuming after being auto-paused. Please wait a few seconds and try again

Furthermore, once that initial error occurs, subsequent tool invocations in the same conversation skip the KB entirely. Adding retries ensures that a paused cluster has time to resume and that embeddings are actually retrieved on the first user turn.

Changes

• Import time

  • Added to both knowledge.py and vector_search.py to support time.sleep() delays.

• knowledge.py (search_knowledge)

  • Wrapped the call to search_related_docs() in a for-loop over two delays: [15, 45].
  • On catching any exception whose message contains "resuming after auto-pause", log a warning and sleep() before retrying.
  • All other exceptions (or after both delays) are logged and re-raised with full traceback.

• vector_search.py (_bedrock_knowledge_base_search)

  • Wrapped the agent_client.retrieve(...) call in the same two-stage retry loop.
  • Detects the same pause message (case-insensitive) in ClientError and re-issues the call after delays.
  • Consolidated error logging: non-retryable errors now bubble up immediately; retryable ones are distinguished by warning logs.

• Cleaned up unreachable code

  • Removed the standalone except ClientError at the end of _bedrock_knowledge_base_search, since all error handling now occurs within the retry loop.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[statefb]

Please fix the CI fails, thank you!

[statefb]

Could you replace it with reretry as merged on #860?