Avoid repeated context injection in RetrievalAugmentationAdvisor

Current Behavior

RetrievalAugmentationAdvisor is currently invoked on every LLM call, including intermediate calls triggered during tool execution.

In my setup, the advisor is configured with:

Query Transformers
RewriteQueryTransformer (uses LLM to rephrase queries for better retrieval)
Document Retriever
VectorStoreDocumentRetriever backed by Pinecone
Document Post-Processors
Custom NvidiaRerankingDocumentPostProcessor for reranking retrieved results

ChatClient Configuration

this.chatClient = chatClientBuilder
.defaultSystem(p -> p.text(systemPrompt)
.param(AgentEnvironment.ENVIRONMENT_INFO_KEY, AgentEnvironment.info())
.param(AgentEnvironment.AGENT_MODEL_KEY, agentModel)
.param(AgentEnvironment.AGENT_MODEL_KNOWLEDGE_CUTOFF_KEY, knowledgeCutoff)
)
.defaultTools(
FileSystemTools.builder().build(),
GrepTool.builder().build(),
GlobTool.builder().build(),
ShellTools.builder().build(),
AskUserQuestionTool.builder()
.questionHandler(new CommandLineQuestionHandler())
.answersValidation(false)
.build()
)
.defaultAdvisors(
retrievalAugmentationAdvisor,
ToolSearchToolCallAdvisor.builder()
.conversationHistoryEnabled(false)
.toolSearcher(toolSearcher)
.maxResults(3)
.build(),
MyLoggingAdvisor.builder()
.showUserText(true)
.showAssistantText(true)
.showAvailableTools(true)
.build()
)
.build();

Use Case

I am building a POC for an agentic JUnit test generation system.

The vector store contains domain-specific rules (e.g., rules related to Owner domain)
When generating tests for a specific class (e.g., OwnerController), the agent should:
Retrieve relevant domain rules
Use them to generate accurate test cases

Problem

While query rewriting, retrieval, and reranking work as expected, the issue is:

RetrievalAugmentationAdvisor is executed repeatedly for every LLM call, including tool calls.

This leads to:

Repeated query rewriting (LLM cost)
Repeated vector store retrieval (latency + cost)
Repeated reranking (extra processing)
Overall slower execution and increased cost

As a temporary workaround, I am considering implementing a custom advisor to control this behavior.
-->

Expected Behavior

It would be helpful to have a mechanism to:

Apply retrieval augmentation only once per user request / session
Avoid repeated execution during internal LLM calls (e.g., tool invocations)
Cache or reuse previously retrieved and processed context within the same interaction

Questions / Suggestions

Is there an existing workaround to prevent repeated retrieval during tool calls?
Would using a sub-agent be the recommended approach for this scenario?
Could a configuration option be introduced to:
Limit retrieval augmentation to the initial call
Or enable context reuse across subsequent calls

Start at RetrievalAugmentationAdvisor and trace how it is invoked for the initial request and for tool-triggered LLM calls. Compare the advisor lifecycle with the expected request-scoped reuse behavior, then verify that retrieval, query rewriting, and reranking occur only when intended during an agent interaction.