vai-workflow-asymmetric-search

Voyage AI's shared embedding space enables ~83% reduction in query-time embedding costs with minimal quality loss. But developers need to see this pattern in action — a clear, minimal workflow that demonstrates the technique.

Install

vai workflow install vai-workflow-asymmetric-search

How It Works

1. Embed — Embed the query with voyage-4-lite (cheapest model)
2. Search — Search the collection where documents were embedded with voyage-4-large
3. Rerank — Refine results for final relevance ordering

Execution Plan

Layer 1 (parallel):  embed_query | vector_search
Layer 2:             rerank_results

Example Usage

vai workflow run vai-workflow-asymmetric-search \
  --input query="How do I configure rate limiting for my API?" \
  --input collection="api_docs" \
  --input limit=5

What This Teaches

• The simplest demonstration of the shared embedding space
• embed_query and vector_search run in parallel
• The output template includes a note field that explains the asymmetric pattern
• The inputType: "query" parameter distinguishes query embeddings from document embeddings

License

MIT © 2026 Michael Lynn