Introduction to Vector database and Chroma DB

Quick memory notes

Vector database types

• In-memory

  • Store vectors in RAM for very fast reads and writes

  • Suited for real-time analytics and recommendations

  • Examples: RedisAI, TorchServe

• Disk-based

  • Persist vectors on disk for data sets larger than memory

  • Use indexing and compression to speed retrieval

  • Examples: Annoy, Milvus, ScaNN

• Distributed

  • Spread vectors across multiple nodes for horizontal scale and fault tolerance

  • Good for massive data sets and high throughput

  • Examples: FAISS, Elasticsearch with vector plugin, Dask-ML

• Graph-based

  • Represent vectors as properties in graph nodes and edges

  • Capture complex relationships for graph analytics

  • Examples: Neo4j, Amazon Neptune, TigerGraph

• Time-series

  • Handle time-stamped vectors for temporal patterning and anomaly detection

  • Examples: InfluxDB, TimescaleDB, Prometheus

Dedicated vector databases vs databases that support vector search

• Dedicated vector databases

  • Built for vector storage, indexing, querying and analysis

  • Use specialised structures and techniques like inverted indexes, product quantisation and locality-sensitive hashing

  • Support nearest neighbour, similarity and distance operations

  • Optimised for speed and high dimensional vectors, scale across clusters

  • Tunable indexing and search parameters for use case fit

  • Examples: FAISS, Annoy, Milvus

• Databases that support vector search

  • General databases or processing frameworks with vector features or plug-ins

  • Store vectors as blobs, arrays or user-defined types

  • May offer standard or custom indexes and integrations for vector operations

  • Can execute similarity search via add-ons or external libraries

  • Typically less optimised than dedicated systems, evaluate performance and scale for needs

  • Examples: SingleStore with IBM watsonx.ai support, Elasticsearch vector add-on, PostgreSQL (PostGIS noted), MySQL vector indexes, RedisAI in memory, MongoDB, Apache Cassandra

Where they are used

• Recommendation systems, information retrieval, social network analysis, knowledge graphs, monitoring, forecasting and anomaly detection

Application of Vector DB

How companies and individuals use vector databases – quick memory notes

• Image and video analysis

  • Store high-dimensional feature vectors for colour histograms, texture descriptors and deep embeddings

  • Run similarity search to find, summarise and recommend images or videos

  • Real-time processing with horizontal scalability for surveillance, object recognition and live event analysis

  • Example: photo app keeps embeddings of user photos to suggest similar images for tagging or album organisation

• Recommendations

  • Embed items and use nearest neighbour search for personalised suggestions

  • Scale to high concurrency with fast indexing and query processing

  • Cross-domain suggestions via shared embedding spaces

  • Example: streaming service uses movie embeddings to recommend related titles after a viewing

• Geospatial and location-based services

  • Use R-tree or quadtree indexing for addresses, polygons and GPS points

  • Support spatial queries like proximity, range and spatial joins

  • Real-time geospatial analytics: spatial clustering, pattern recognition for traffic, fleet management, dynamic routing and hotspot detection

  • Example: navigation app queries a vector store of restaurant GPS locations to list nearby options

• Social and marketing insights

  • Distributed storage and parallel processing for big data and simultaneous queries (SEO, user profiles)

  • Optimised caching and query execution plans to reduce latency and speed trend analytics for influencers and advertisers

  • Autoscaling and dynamic resource allocation for performance and cost control in cloud and on-premise environments

  • Example: social platform scales databases as users grow, tracks interests and clicks without slowing response times

• Key takeaways

  • Vector databases enable feature extraction, similarity search and real-time processing for media

  • Embeddings power fast, scalable, cross-domain recommendations

  • Geospatial processing supports GPS services, fleet management and real-time traffic routing

  • Social and marketing platforms benefit from horizontal scalability, low latency and adaptive resource usage