Multi-Vector Retrieval (C++ Implementation)

This page documents the C++ implementation of multi-vector retrieval and evaluation in the VectorSearch system.

Overview

We implemented multi-vector aggregation in C++ to improve retrieval robustness when a query is represented by multiple semantically similar vectors.

How It Works

🔹 Query Aggregation

A set of query vectors (e.g., ID 0–4) is treated as one logical query. Each base vector is scored based on the average L2 distance to these multiple queries.

for (int i = 0; i < prep.data.N; ++i) {
    float total_dist = 0.0f;
    for (const auto& qvec : multi_query) {
        total_dist += cal_distSqrt(qvec.data(), prep.data.val[i], prep.data.dim);
    }
    score_all[i] = { total_dist / multi_query.size(), i };
}