Retrieve and Memory tools to support rerankingConfiguration

[JimmyBcn]

Hello!

I'm in the process of migrating few bedrock agents into strands agents.

I've been successfully using bedrock agents querying knowledge bases. At invocation time, I've been able to add a rerankingConfiguration into the vectorSearchConfiguration of the retrieval configuration. That has been working well so far for those bedrock agents.

Now that I try to migrate them to strands, I've seen that neither the retrieve tool nor the memory tool allow injecting a rerankingConfiguration . Both tools just have a very limited default configuration in their implementation:

retrieval_config = {"vectorSearchConfiguration": {"numberOfResults": number_of_results}}

I could modify it to specify a filter property in that vectorSearchConfiguration, and it's working as expected.
However, at the time of adding a rerankingConfiguration the query to the knowledge base always returns no results. The same configuration when used while invoking a bedrock agent is working fine. Even using the AmazonKnowledgeBasesRetriever connector in llamaIndex does its job.

So my questions are:

• are those tools ready to perform a query using rerankingConfiguration?
• why does not allow to pass a full vectorSearchConfiguration as other frameworks do?

This is how I modified the code at the retrieve tool (indeed, I have access to the reranking model in Oregon)

        # Default retrieval configuration
        # retrieval_config = {"vectorSearchConfiguration": {"numberOfResults": number_of_results}}
        # https://docs.aws.amazon.com/bedrock/latest/APIReference/API_agent-runtime_KnowledgeBaseVectorSearchConfiguration.html
        retrieval_config = {
            "vectorSearchConfiguration": {
              "numberOfResults": 10,
              "overrideSearchType": "HYBRID",
              "rerankingConfiguration": {
                "type": "BEDROCK_RERANKING_MODEL",
                "bedrockRerankingConfiguration": {
                  "metadataConfiguration": {
                    "selectionMode": "ALL"
                  },
                  "modelConfiguration": {
                    "modelArn": "arn:aws:bedrock:us-west-2::foundation-model/amazon.rerank-v1:0"
                  },
                  "numberOfRerankedResults": 5
                }
              },
            }
        }

Thanks. I'm very excited to finally see my agents running with strands!

[JimmyBcn]

All right. Look like using a reranking model decreases the score of the results by something near to 0. I could fix it by specifying a min_score=0 in my request.

So now only my second question is still relevant: why does not allow passing a full vectorSearchConfiguration as other frameworks do?

[mehtarac]

Our tools repository is currently a community driven package and we welcome community contributions. This sounds like it would be a great contribution to the enhance the the tools capabilities and would love to have a PR for it if you have an implementation in mind. If interested, please open a feature request and a PR and we an assign you the issue such that you are credited as the author of the contribution. Thanks!

[JimmyBcn]

I finally follow the same approach as in this example and implemented my own tool, where I can completely configure the retrieval configuration.

But I will consider making a contribution in order to have it available in the already existing retrieval tool. Thanks for the suggestion.

Thanks