vai-workflow-question-decomposition

Complex questions often span multiple concepts. A single vector search may retrieve documents about one topic but miss others. Decomposing the question into focused sub-queries and searching for each independently produces better coverage.

Install

vai workflow install vai-workflow-question-decomposition

How It Works

1. Decompose — An LLM call breaks the complex question into 2–3 focused sub-questions
2. Parallel search — Each sub-question is searched independently using query (with reranking)
3. Merge — Results from all sub-queries are merged and deduplicated
4. Rerank — The merged results are reranked against the original question
5. Synthesize — An LLM generates a comprehensive answer using the reranked results as context

Execution Plan

Layer 1:             decompose
Layer 2 (parallel):  search_sub1 | search_sub2 | search_sub3
Layer 3:             merged → final_rerank
Layer 4:             synthesize

Example Usage

vai workflow run vai-workflow-question-decomposition \
  --input query="How do microservices handle authentication and what are the performance implications compared to monolithic architectures?" \
  --input collection="engineering_docs"

What This Teaches

• LLM output (generate) can drive downstream step inputs — the decomposed sub-questions become search queries
• Steps with identical dependencies run in parallel automatically
• Reranking against the original question ensures final relevance reflects the user's actual intent
• The output template can return a structured object with multiple fields

License

MIT © 2026 Michael Lynn