rag-fusion

RAG Fusion: Multi-Query Search with Reciprocal Rank Fusion (k=60)

One query misses relevant documents that a rephrased version would find. This pipeline rewrites the question into variants, searches three indices in parallel via FORK_JOIN, fuses results using RRF with k=60 (score = sum(1/(k+rank))), and generates from the top fused documents.

Workflow

question
    │
    ▼
┌──────────────────────┐
│ rf_rewrite_queries   │  Generate query variants
└───────────┬──────────┘
            ▼
┌─── FORK_JOIN ──────────────────────────────────┐
│ ┌──────────────┐ ┌──────────────┐ ┌───────────┐│
│ │rf_search_v1  │ │rf_search_v2  │ │rf_search_v3││
│ └──────────────┘ └──────────────┘ └───────────┘│
└──────────────────────┬─────────────────────────┘
                       ▼
            ┌─────────────────────┐
            │ rf_fuse_results     │  RRF: score = sum(1/(60+rank))
            └──────────┬──────────┘
                       ▼
            ┌─────────────────────┐
            │ rf_generate_answer  │  Generate from fused context
            └─────────────────────┘

Workers

RewriteQueriesWorker (rf_rewrite_queries) -- Generates query variants. When API key is set, uses LLM to rewrite.

SearchV1Worker (rf_search_v1) -- Returns docs about versioning and retry policies with rank positions.

SearchV2Worker (rf_search_v2) -- Returns docs about JSON data flow and versioning (overlaps with V1 on versioning doc).

SearchV3Worker (rf_search_v3) -- Returns docs about polyglot workers and visual editors.

FuseResultsWorker (rf_fuse_results) -- Implements RRF with k=60. For each document ID across all result lists, sums 1/(k + rank). Handles null result lists gracefully. Returns fusedDocs sorted descending, fusedCount, and totalCandidates.

GenerateAnswerWorker (rf_generate_answer) -- Generates from the top fused documents.

Tests

41 tests extensively cover query rewriting, all three searches, RRF fusion math, and generation.

Further Reading

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