Approximate nearest neighbor (ANN) graph indices such as HNSW and Vamana construct graph topology in full-precision or high-fidelity quantized metric spaces, using binary quantization (BQ) only as a post-hoc search-time distance estimator. We ask whether BQ can build the graph itself. We present QuIVer (Quantized Index for Vector Retrieval), a training-free ANN graph index that performs edge selection, pruning, and graph navigation entirely in a 2-bit Sign-Magnitude BQ metric space. QuIVer combines: (i) a 2-bit encoding that preserves sign and magnitude strength at 1/12 the memory of float32 vectors; (ii) Vamana alpha-diversity pruning directly on BQ distances; and (iii) symmetric BQ beam search using XOR/AND/Popcount, followed by float32 reranking over a small candidate set. On six embedding datasets spanning 384--3072 dimensions, QuIVer achieves at least 88% Recall@10 at 13--41K multi-threaded QPS with 58--262-second construction and less than 1.3 GB hot memory. At matched recall on Cohere-1M, it outperforms the official DiskANN Rust implementation by 2.5--3.3x, hnswlib by 3.6--4.7x, and FAISS HNSW by 3.8--4.9x in multi-threaded throughput. Controlled experiments on additional datasets, including word vectors, CV features, uniform random vectors, multimodal CLIP embeddings, and low-rank synthetic data, delineate QuIVer's applicability boundary: high recall requires cosine-native distributions with low effective dimensionality, while monotonic recall gains with increasing ef suggest that BQ noise mainly affects navigation efficiency in the tested regimes.

翻译：近似最近邻（ANN）图索引（如HNSW和Vamana）在全精度或高保真量化度量空间中构建图拓扑，仅将二元量化（BQ）作为事后搜索时的距离估计器。我们探究BQ是否能够自行构建图结构。本文提出QuIVer（向量检索量化索引），这是一种免训练的ANN图索引，其边选择、剪枝和图导航完全在2比特符号-幅值BQ度量空间中完成。QuIVer结合了：（i）一种2比特编码，在float32向量1/12内存占用下保留符号与幅值强度；（ii）直接在BQ距离上执行Vamana alpha多样性剪枝；（iii）使用XOR/AND/Popcount的对称BQ波束搜索，随后对小型候选集进行float32重排序。在六个维度为384至3072的嵌入数据集上，QuIVer在多线程每秒查询数（QPS）为13K至41K、构建耗时58至262秒、热内存低于1.3 GB的条件下，实现了至少88%的Recall@10。在Cohere-1M上匹配召回率时，其多线程吞吐量优于官方DiskANN Rust实现2.5至3.3倍、hnswlib 3.6至4.7倍、FAISS HNSW 3.8至4.9倍。在包含词向量、计算机视觉特征、均匀随机向量、多模态CLIP嵌入以及低秩合成数据的额外数据集上进行的控制实验，界定了QuIVer的适用边界：高召回率要求具有低有效维度的余弦原生分布，而随ef增大呈现单调增长的召回率表明，在测试条件下BQ噪声主要影响导航效率。