vai-workflow-hybrid-precision-search

A single vector search strategy returns biased results. Broad models prioritize recall, while precise models prioritize precision. Metadata filters find structurally matching documents that may miss semantic intent. No single approach is optimal for all queries.

Install

vai workflow install vai-workflow-hybrid-precision-search

How It Works

The workflow runs three retrieval strategies in parallel against the same collection:

1. Broad retrieval — search with voyage-4-lite for high recall at low cost
2. Precise retrieval — query with voyage-4-large for high precision with reranking
3. Filtered retrieval — search with a metadata filter for structural matching

Results from all three strategies are merged with deduplication, then reranked against the original query to produce a final, optimized result set.

Execution Plan

Layer 1 (parallel):  broad_search | precise_search | filtered_search
Layer 2:             merged
Layer 3:             final_rerank

Example Usage

vai workflow run vai-workflow-hybrid-precision-search \
  --input query="How does rate limiting work in distributed systems?" \
  --input collection="engineering_docs" \
  --input filter='{"category":"architecture"}' \
  --input limit=5

What This Teaches

• Three steps with no dependencies on each other run in parallel automatically (Layer 1)
• The merge tool with dedup: true eliminates duplicate documents found by multiple strategies
• The shared embedding space allows mixing voyage-4-lite, voyage-4, and voyage-4-large results
• Final reranking produces a unified relevance ordering regardless of which strategy originally found each document

License

MIT © 2026 Michael Lynn