Definition
A search paradigm in vector databases that uses indexing structures to find data points in high-dimensional space close to a query vector, sacrificing 100% precision (recall) for massive gains in retrieval speed and reduced computational overhead. In RAG pipelines, it enables sub-second retrieval from millions of document embeddings by avoiding exhaustive 'brute-force' distance calculations.
Unlike exact K-Nearest Neighbor (k-NN), ANN prioritizes low-latency 'good enough' results over absolute mathematical accuracy.
"Finding a book in a library by going to a specific genre aisle and shelf rather than scanning every single spine in the building."
- Vector Embedding(Prerequisite)
- HNSW (Hierarchical Navigable Small World)(Common Algorithm)
- Recall(Performance Metric)
- Cosine Similarity(Distance Metric)
Conceptual Overview
A search paradigm in vector databases that uses indexing structures to find data points in high-dimensional space close to a query vector, sacrificing 100% precision (recall) for massive gains in retrieval speed and reduced computational overhead. In RAG pipelines, it enables sub-second retrieval from millions of document embeddings by avoiding exhaustive 'brute-force' distance calculations.
Disambiguation
Unlike exact K-Nearest Neighbor (k-NN), ANN prioritizes low-latency 'good enough' results over absolute mathematical accuracy.
Visual Analog
Finding a book in a library by going to a specific genre aisle and shelf rather than scanning every single spine in the building.