How to merge results from two vectorStore queries into one prompt context?

[marcelodl]

Hello everyone, I need some guidance.

I’m using a vectorStore and want to build a prompt that includes two distinct contexts, both retrieved from the vector store, for example:

Given the context below:
-----------------
<context1>   // result set #1 from vectorStore (e.g., filter A / topK1)
-----------------

Write a document using the model below:
-----------------
<context2>   // result set #2 from vectorStore (e.g., filter B / topK2)
-----------------

I’m currently using QuestionAnswerAdvisory like this:

chatClient.prompt()
    .advisors(advisorSpec -> advisorSpec
        .param(CONVERSATION_ID, id))
    .param(TOP_K, 100)
    .param(QuestionAnswerAdvisor.FILTER_EXPRESSION, "key == '" + key + "'")
;

What’s the recommended Spring AI pattern to run two separate vectorStore queries (with different filters/topK) and inject both result sets into a single prompt?

• Is it possible to attach two QuestionAnswerAdvisors (one per filter) and have them both contribute to the final context?
• Or should I run two queries manually (pre-prompt), then pass them as separate variables/sections in the prompt template (e.g., <context1> and <context2>) via prompt().param(...)?

Any code snippet or best practice for merging multiple retrievals into one prompt would be greatly appreciated. Thanks!

[gengzi]

The official documentation of Spring AI states that VectorStoreDocumentRetriever allows for direct retrieval from the vector repository. You can create different vector retrievers and merge the results with ConcatenationDocumentJoiner

https://docs.spring.io/spring-ai/reference/api/retrieval-augmented-generation.html

[sachin-gupta99]

Hey @marcelodl , I ran into the exact same problem while building a multi-domain RAG service where I needed to query activity data, nutrition data, and
user profile data separately from PgVector and merge them into a single prompt. Here's what worked for me:

Approach 1: Manual retrieval + prompt template variables (Recommended)

This gives you full control over each query's filters and topK, and lets you place each context exactly where you want in the prompt.

// 1. Run separate vector store queries with different filters
SearchRequest activitySearch = SearchRequest.builder()
.query(userQuery)
.topK(10)
.filterExpression("domain == 'activity'")
.build();

SearchRequest nutritionSearch = SearchRequest.builder()
.query(userQuery)
.topK(5)
.filterExpression("domain == 'nutrition'")
.build();

List activityDocs = vectorStore.similaritySearch(activitySearch);
List nutritionDocs = vectorStore.similaritySearch(nutritionSearch);

// 2. Format each context
String activityContext = activityDocs.stream()
.map(Document::getText)
.collect(Collectors.joining("\n"));

String nutritionContext = nutritionDocs.stream()
.map(Document::getText)
.collect(Collectors.joining("\n"));

// 3. Inject both into the prompt as separate variables
String response = chatClient.prompt()
.system(s -> s.text("""
Given the activity context below:
-----------------
{activityContext}
-----------------

      And the nutrition context below:
      -----------------
      {nutritionContext}
      -----------------

      Use both contexts to answer the user's question.
      """)
      .param("activityContext", activityContext)
      .param("nutritionContext", nutritionContext))
  .user(userQuery)
  .call()
  .content();

This approach is clean and gives you full control — you decide the filter, topK, and exact placement of each context block in the prompt.

Approach 2: Using RetrievalAugmentationAdvisor with multiple retrievers

As @gengzi mentioned, Spring AI's newer RAG API provides VectorStoreDocumentRetriever and ConcatenationDocumentJoiner for this exact use case:

DocumentRetriever activityRetriever = VectorStoreDocumentRetriever.builder()
.vectorStore(vectorStore)
.similarityThreshold(0.7)
.topK(10)
.filterExpression("domain == 'activity'")
.build();

DocumentRetriever nutritionRetriever = VectorStoreDocumentRetriever.builder()
.vectorStore(vectorStore)
.similarityThreshold(0.7)
.topK(5)
.filterExpression("domain == 'nutrition'")
.build();

RetrievalAugmentationAdvisor advisor = RetrievalAugmentationAdvisor.builder()
.documentRetriever(activityRetriever)
.documentRetriever(nutritionRetriever)
.documentJoiner(new ConcatenationDocumentJoiner())
.build();

String response = chatClient.prompt()
.advisors(advisor)
.user(userQuery)
.call()
.content();

This merges all retrieved documents into a single context block automatically. It's cleaner if you don't need separate placement of each context in the
prompt.

For your use case where you explicitly want and as separate sections, its better to go with Approach 1. It's more code but gives you precise control over how each context appears in the prompt.

Hope this helps!