ANN
Approximate nearest neighbor search is a retrieval strategy that trades a small loss in recall accuracy for order-of-magnitude gains in search speed. It is the core method by which vector databases achieve millisecond-level retrieval.
Core Facts
向量数据库的核心操作是近似最近邻搜索（ANN），通过HNSW、IVF等索引算法将相似度计算加速到毫秒级响应
In the offline phase of the traditional RAG process, documents are chunked and converted into vectors via an embedding model and stored in a vector database; in the online phase, similar document chunks are retrieved using an Approximate Nearest Neighbor (ANN) algorithm.
Timeline (last 90 days)
Approximate nearest neighbor (ANN) index structures (such as HNSW and IVF) need to be rebuilt or tuned after the data distribution changes, otherwise recall will decline.
An ANN index is bound to a distance metric; HNSW organizes the nearest-neighbor graph according to a certain metric at build time, and switching to another metric usually requires rebuilding the index.
HNSW is a hierarchical navigable small world index structure that belongs to the approximate nearest neighbor (ANN) strategy, trading a small amount of recall accuracy for order-of-magnitude speed improvements.
The cost of multi-vector approaches is that each document requires storing multiple vectors, making index sizes typically several to tens of times larger than single-vector approaches and placing higher demands on vector database storage and ANN search performance.
Commonly used distance metrics for vector retrieval include cosine similarity and dot product; practical systems typically employ approximate nearest neighbor (ANN) index structures such as HNSW or IVF to trade a small amount of precision for speed.
ANN solutions trade a slight loss in accuracy for orders-of-magnitude improvements in retrieval speed, returning the most similar results in milliseconds even at scales of millions to billions of vectors.
Vector index efficiency issues are typically addressed by approximate nearest neighbor (ANN) algorithm libraries, with common solutions including FAISS, Annoy, and the graph-based HNSW algorithm.
Hybrid retrieval, which combines the precision of keyword retrieval with the semantic recall of vector retrieval, has become the de facto standard for modern search engines.
The paper proposes combining ANN-based AI with probability-wave-based 'entangled brain simulation' to advance the development of 'Human-like Processing Units' (HPUs).
The two-tower model uses a user tower and an item tower to generate embeddings separately, pre-computes item vectors offline and builds an index, and only needs to compute the query vector online before performing ANN retrieval.
2 more timeline events
All Facts (14)
向量数据库的核心操作是近似最近邻搜索（ANN），通过HNSW、IVF等索引算法将相似度计算加速到毫秒级响应
80%VerifiedIn the offline phase of the traditional RAG process, documents are chunked and converted into vectors via an embedding model and stored in a vector database; in the online phase, similar document chunks are retrieved using an Approximate Nearest Neighbor (ANN) algorithm.
65%UnverifiedApproximate nearest neighbor (ANN) index structures (such as HNSW and IVF) need to be rebuilt or tuned after the data distribution changes, otherwise recall will decline.
50%UnverifiedAn ANN index is bound to a distance metric; HNSW organizes the nearest-neighbor graph according to a certain metric at build time, and switching to another metric usually requires rebuilding the index.
50%UnverifiedHNSW is a hierarchical navigable small world index structure that belongs to the approximate nearest neighbor (ANN) strategy, trading a small amount of recall accuracy for order-of-magnitude speed improvements.
50%UnverifiedThe cost of multi-vector approaches is that each document requires storing multiple vectors, making index sizes typically several to tens of times larger than single-vector approaches and placing higher demands on vector database storage and ANN search performance.
50%UnverifiedCommonly used distance metrics for vector retrieval include cosine similarity and dot product; practical systems typically employ approximate nearest neighbor (ANN) index structures such as HNSW or IVF to trade a small amount of precision for speed.
50%UnverifiedANN solutions trade a slight loss in accuracy for orders-of-magnitude improvements in retrieval speed, returning the most similar results in milliseconds even at scales of millions to billions of vectors.
50%UnverifiedVector index efficiency issues are typically addressed by approximate nearest neighbor (ANN) algorithm libraries, with common solutions including FAISS, Annoy, and the graph-based HNSW algorithm.
50%UnverifiedHybrid retrieval, which combines the precision of keyword retrieval with the semantic recall of vector retrieval, has become the de facto standard for modern search engines.
50%UnverifiedThe paper proposes combining ANN-based AI with probability-wave-based 'entangled brain simulation' to advance the development of 'Human-like Processing Units' (HPUs).
50%UnverifiedThe two-tower model uses a user tower and an item tower to generate embeddings separately, pre-computes item vectors offline and builds an index, and only needs to compute the query vector online before performing ANN retrieval.
50%UnverifiedDense retrieval combined with ANN index technologies (such as HNSW and FAISS) can achieve millisecond-level recall across millions to billions of vectors.
50%UnverifiedHNSW算法的时间复杂度约为O(log N)，近似最近邻算法在召回率95-99%下将检索时间从线性降至对数级别
50%