--- id: vector-databases title: "Vector Databases: Approximate Nearest Neighbor Search, Embedding Storage, and Retrieval at Scale" schema_type: article category: ai language: en confidence: high last_verified: "2026-05-24" created_date: "2026-05-24" generation_method: ai_assisted ai_models: - claude-4.5-sonnet derived_from_human_seed: true conflict_of_interest: none_declared is_live_document: false data_period: static completeness: 0.85 atomic_facts: - id: af-vector-databases-1 statement: >- Microsoft Research DiskANN (2026) achieved the state-of-the-art in web-scale vector search — indexing and querying billion-scale embedding datasets with >95% recall at <5ms latency per query using graph-based ANN algorithms with SSD-aware index layouts, enabling hybrid vector+keyword search at Bing and Microsoft 365 Copilot for trillions of documents. source_title: "Microsoft Research (2026) — DiskANN: Vector Search at Web Scale / MSR Project Akupara / DataCamp 2026 vector database comparison" source_url: https://www.microsoft.com/en-us/research/project/project-akupara-approximate-nearest-neighbor-search-for-large-scale-semantic-search/ confidence: high - id: af-vector-databases-2 statement: >- arxiv (October 2025) introduced DSANN — a distributed storage ANN system for billion-scale vector databases — combining concurrent indexing, distributed storage with load balancing, and fault-tolerant query serving — demonstrating linear scalability to 10 billion 128-dimensional vectors while maintaining sub-10ms query latency, addressing the infrastructure challenge of deploying vector search at internet scale. source_title: "arxiv 2510.17326 (2025) — DSANN: Distributed Approximate Nearest Neighbor Search on Large-Scale Vectors" source_url: https://arxiv.org/abs/2510.17326 confidence: high primary_sources: - id: ps-vector-databases-1 title: "DiskANN: Graph-based Approximate Nearest Neighbor Search on Disk" type: academic_paper year: 2026 institution: Microsoft Research / NeurIPS url: https://www.microsoft.com/en-us/research/project/project-akupara-approximate-nearest-neighbor-search-for-large-scale-semantic-search/ - id: ps-vector-databases-2 title: "DSANN: Distributed Approximate Nearest Neighbor Search of Large Scale Vectors" type: academic_paper year: 2025 institution: arXiv url: https://arxiv.org/abs/2510.17326 known_gaps: - Multi-modal hybrid search combining dense vectors, sparse vectors, and structured filters - Dynamic index maintenance — updating vector indices without full rebuild disputed_statements: [] secondary_sources: - title: "A Comprehensive Survey of Vector Databases: Architectures, Algorithms, and Applications for AI" type: survey_paper year: 2024 authors: - multiple institution: ACM Computing Surveys url: https://doi.org/10.1145/3635100 - title: Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs (HNSW) type: journal_article year: 2018 authors: - Malkov, Yu. A. - Yashunin, D. A. institution: IEEE TPAMI url: https://doi.org/10.1109/TPAMI.2018.2889473 - title: "Pinecone, Weaviate, Chroma, Milvus: A Comparative Analysis of Modern Vector Database Systems for LLM Applications" type: report year: 2024 authors: - DB-Engines institution: DB-Engines / GitHub url: https://db-engines.com/en/article/Vector+DBMS - title: "FAISS: A Library for Efficient Similarity Search and Clustering of Dense Vectors (Meta AI)" type: technical_report year: 2019 authors: - Johnson, Jeff - Douze, Matthijs - Jégou, Hervé institution: Meta AI Research url: https://arxiv.org/abs/2401.08281 updated: "2026-05-24" --- ## TL;DR Vector databases are the storage engine powering modern AI — from RAG (Retrieval-Augmented Generation) to semantic search to recommendation. They store embeddings (numerical representations of text, images, audio) and perform approximate nearest neighbor (ANN) search to find the most similar items in milliseconds across billions of vectors. ## Core Explanation Why vector databases: LLMs and embedding models convert unstructured data (text, images, audio) into fixed-length dense vectors (embeddings) — typically 768-4096 dimensions. Semantic similarity = cosine similarity or Euclidean distance between vectors. Problem: finding the k-nearest neighbors among N vectors requires O(Nd) comparisons — prohibitively slow for N > 1M. ANN algorithms trade some recall for dramatic speedup: (1) Graph-based — HNSW (Hierarchical Navigable Small World) builds a multi-layer proximity graph; search traverses graph edges, arriving at nearest neighbor in O(log N) steps. Widely used (FAISS, hnswlib); (2) Quantization-based — product quantization (PQ) compresses vectors by splitting into subvectors, clustering each subspace — achieves 8-32x compression; (3) Tree-based — Annoy, KD-trees, randomized projection trees; (4) Hash-based — locality-sensitive hashing (LSH) maps similar vectors to same bucket. ## Detailed Analysis Leading vector databases (DataCamp 2026): Pinecone (managed cloud, proprietary index), Weaviate (open-source, hybrid vector+keyword), Milvus (cloud-native, distributed), Qdrant (Rust-based, high-performance), Chroma (lightweight, developer-friendly), pgvector (PostgreSQL extension). FAISS (Meta, open-source) remains the gold standard for research and custom deployments — GPU-accelerated IVF+PQ indexes handle 1B+ vectors with <5ms latency. Microsoft DiskANN (2026): breakthrough in SSD-based indexes — traditional ANN requires all vectors in RAM (hundreds of GB for billion-scale). DiskANN places graph index on SSD with carefully designed data layout minimizing random reads, achieving RAM-level latency at 10x lower cost. DSANN (2025): distributes billion-scale indexes across hundreds of machines with linear scalability, ensuring high availability. Applications: (1) RAG — retrieve relevant documents for LLM context; (2) Semantic search — users query by meaning, not keywords; (3) Multimodal search — text-to-image, image-to-image retrieval; (4) Recommendation — find similar items to user interests; (5) Deduplication — find near-duplicate content across massive corpora. Key challenges: (1) Filtered search — combining vector similarity with structured metadata filters (price range, date, category) without destroying ANN performance; (2) Freshness — inserting new vectors requires index rebuilding (batch) or online insertion with gradual quality degradation. ## Further Reading - FAISS: Facebook AI Similarity Search (Meta) - Pinecone, Weaviate, Milvus, Qdrant Vector Databases - ANN-Benchmarks: Benchmarking ANN algorithms