Hybrid Search

Why Hybrid Search?

Pure vector search is excellent at semantic similarity but can miss exact term matches. Consider:

• Query: "What is policy HR-2024-LEV-003?" — the policy ID is a unique string; vector search may not find the right document if the ID wasn't prominently embedded
• Query: "How does VLOOKUP work?" — keyword "VLOOKUP" is specific; vector search on "how to look up data" may retrieve unrelated spreadsheet docs

Hybrid search solves this by running both vector search and keyword search in parallel, then merging results using Reciprocal Rank Fusion (RRF).

Hybrid Search Implementation

public class HybridSearchService
{
    public async Task<IReadOnlyList<MemoryRecord>> SearchAsync(
        string collection,
        string queryText,
        float[] queryEmbedding,
        int topK,
        CancellationToken ct = default)
    {
        // Run both searches in parallel
        var vectorTask = _vectorStore.SearchAsync(
            collection, queryEmbedding, topK * 2, minScore: 0.5f, ct: ct);

        var keywordTask = _keywordSearch.SearchAsync(
            collection, queryText, topK * 2, ct: ct);

        await Task.WhenAll(vectorTask, keywordTask);

        // Merge using Reciprocal Rank Fusion
        return ReciprocallRankFusion(
            vectorTask.Result,
            keywordTask.Result,
            topK: topK,
            vectorWeight: 0.7f,    // configurable
            keywordWeight: 0.3f);
    }

    private IReadOnlyList<MemoryRecord> ReciprocallRankFusion(
        IReadOnlyList<MemoryRecord> vectorResults,
        IReadOnlyList<MemoryRecord> keywordResults,
        int topK, float vectorWeight, float keywordWeight)
    {
        var scores = new Dictionary<string, float>();
        const int k = 60;   // RRF constant

        for (int i = 0; i < vectorResults.Count; i++)
            scores[vectorResults[i].Id] =
                scores.GetValueOrDefault(vectorResults[i].Id) +
                vectorWeight / (k + i + 1);

        for (int i = 0; i < keywordResults.Count; i++)
            scores[keywordResults[i].Id] =
                scores.GetValueOrDefault(keywordResults[i].Id) +
                keywordWeight / (k + i + 1);

        return scores
            .OrderByDescending(s => s.Value)
            .Take(topK)
            .Select(s => vectorResults.Concat(keywordResults)
                .First(r => r.Id == s.Key))
            .ToList();
    }
}

Keyword Search Backend

For PGVector, the keyword search uses PostgreSQL full-text search (tsvector). For Qdrant, it uses Qdrant's sparse vector support (SPLADE model) for dense-sparse hybrid search. The configuration selects which approach to use:

// Hybrid search config
public class HybridSearchConfig
{
    public bool Enabled { get; set; } = false;      // off by default
    public float VectorWeight { get; set; } = 0.7f;
    public float KeywordWeight { get; set; } = 0.3f;
    public HybridSearchBackend Backend { get; set; } // PgFullText, QdrantSparse
}