rag-multi-query

Multi-Query RAG: Expand One Question into Three Searches in Parallel

A single query gives you one perspective. This pipeline expands the question into 3 query variants, searches for each in parallel via FORK_JOIN, deduplicates the results (documents like "d1" and "d4" appear in multiple result sets), and generates from the combined unique context.

Workflow

question
    │
    ▼
┌──────────────────────┐
│ mq_expand_queries    │  Generate 3 query variants
└───────────┬──────────┘
            ▼
┌─── FORK_JOIN ──────────────────────────────┐
│ ┌──────────────┐ ┌────────────┐ ┌─────────┐│
│ │mq_search_q1  │ │mq_search_q2│ │mq_search││
│ │              │ │            │ │  _q3    ││
│ └──────────────┘ └────────────┘ └─────────┘│
└──────────────────────┬─────────────────────┘
                       ▼
            ┌──────────────────────┐
            │ mq_dedup_results     │  Remove duplicate doc IDs
            └──────────┬───────────┘
                       ▼
            ┌──────────────────────┐
            │ mq_generate_answer   │  Generate from unique docs
            └──────────────────────┘

Workers

ExpandQueriesWorker (mq_expand_queries) -- Generates 3 query variants. When API key is set, uses LLM for expansion.

SearchQ1Worker (mq_search_q1) -- Returns docs d1 (centralized control) and d4 (retry/timeout logic).

SearchQ2Worker (mq_search_q2) -- Returns d1 (same as Q1), d7 (decoupled execution).

SearchQ3Worker (mq_search_q3) -- Returns d4 (same as Q1), d11 (choreography hidden dependencies).

DedupResultsWorker (mq_dedup_results) -- Merges all three result lists (handling null lists gracefully) and removes duplicates, keeping unique documents.

GenerateAnswerWorker (mq_generate_answer) -- Generates from deduplicated context. Estimates tokens via answer.split("\\s+").length.

Tests

27 tests cover query expansion, all three searches, deduplication logic, and answer generation.

Further Reading

• RUNNING.md -- how to build and run this example