[FEATURE] Configurable search type (enforceable hybrid retrieval) and reranker model ARN for the retrieve and memory tools

[MrCsabaToth]

Problem Statement

Currently the retrieve and memory tools search type depends on the underlying data store:

• OpenSearch Serverless with a filterable text field configured → Bedrock can pick HYBRID automatically (or you can override).
• Any other vector store (e.g. plain OpenSearch Serverless without that config, Aurora PostgreSQL, Pinecone, Redis) → SEMANTIC only — hybrid isn't even available regardless of override.
  I'd want to enforce HYBRID search for best retrieval results.

Similarly, I'd also want to leverage a reranking: AWS KB already has the native capability to support that, we just need to specify the reranker model ARN we want to use.

Proposed Solution

Besides the knowledge base ID provide two more configurations:

1. search type: if empty then it can default to the current "auto" mode. If hybrid does not apply to a specific KB store then it can be ignored, or error can be thrown.
2. reranking model ARN: if specified (Bedrock has at least Cohere and Titan embedding, and a multi model embedder too, I expect the latest Nemotron embedder to show up as well) then reranking would be configured for the KB with the respective settings.

Use Case

Any RAG or memory which goes beyond hello world demos could benefit from these low hanging fruits.

Alternatives Solutions

Creating a custom tool with an AWS KB where I control these, but this would be more plumbing code, it'd be great if it's out of the box.

Additional Context

No response

[MrCsabaToth]

Downside would be the two extra knobs, but the upside could be superior retrieval accuracy

[krinn-matthew-bah]

This would definitely be a good feature

[yonib05]

Thanks for the request. As of strands-agents-tools v0.8.6, both the retrieve and memory tools are deprecated (#566): they log a warning on invocation, escalating to an error log in v0.9.0, and will not gain new capabilities.

Knowledge base access now lives in the SDK as MemoryManager with BedrockKnowledgeBaseStore (https://strandsagents.com/docs/user-guide/concepts/memory/bedrock-knowledge-base/). Retrieval configuration is exactly the kind of thing that belongs to the store now: its config already exposes retrieval options such as filter and max_search_results, and enforceable search type and reranker model configuration would slot in alongside them. If the store does not cover what you need, please refile this against strands-agents/harness-sdk and link this issue; the requirements you wrote up carry over directly.

Closing here as part of the deprecation.