Fast and Efficient Approximate Nearest Neighbor Search for High-Dimensional LLM Embeddings

Researchers presented their 2026 SISAP Indexing Challenge entries for two ANN tasks on LLM embeddings.

They built a kNN graph for 1024-dimensional BGE-M3 vectors and tackled MIPS on unnormalized Llama-3.2-8B features.

The method combines EVP-based quantization, reranking, dimensionality augmentation, and FLAS to boost speed, recall, and cache locality.

The work points to more efficient similarity search for large-scale retrieval and indexing systems using LLM embeddings.