A multi-model database is one database that can store several kinds of data, such as documents, graphs, time series and vectors, and answer questions across them together. Without one, a company runs a separate database for each kind of data.
Each kind of data has its own specialist database. A document database stores records such as a customer profile. A graph database stores the links between records. Time series data is readings stamped with the time they were taken, such as a sensor reading every minute. And vector embeddings are lists of numbers that an AI model uses to capture the meaning of text or images, so that similar meanings sit close together.
Running four databases means four sets of servers, four ways of writing queries and four sets of skills in the team. It also means copying data between them and keeping the copies in step. And a question that needs all four kinds must be split into four queries, whose answers the app then has to combine.
A multi-model database keeps all four kinds in one store, with one way of querying them. A single query can then filter by a field in a document, follow a link, read the last week of readings, and find text by meaning, all at once.
Consider a farm advice start-up in Nashik that serves 2,000 grape growers. A soil moisture sensor on each farm sends a reading every 10 minutes. That is 288,000 readings a day, stored as time series. Each farm's profile, with its soil type, grape variety and owner, is a document. Links connect each farm to its buyers and its cold store, as a graph. And 5,000 advice notes written by the start-up's experts, in Marathi and English, are stored as vectors. A grower asks the start-up's chatbot why his vine leaves are turning yellow after the rain. One query finds the advice notes closest in meaning to the question. It keeps only the notes for farms on the same soil type whose moisture readings stayed high for the past three days. The chatbot then answers from those notes. This way of answering from a company's own documents is called retrieval-augmented generation (RAG).
There are trade-offs. A specialist database is often faster at its one job than a database that does several. And because all the data sits in one product, moving away from it later is harder, which is called vendor lock-in.
How is it different from a database that has added one feature, such as vector search? Many databases now store vectors beside their main kind of data. A multi-model database is designed from the start to treat each kind of data as equal.